Noahpinion

Roundup #90: In which Noah joins a Japanese VC firm

2026-10-26T23:17:34+00:00

I have a fun announcement today, which is that I have a new side gig. I’ll be joining the Japanese VC firm Coral Capital as a Venture Partner! That’s less prestigious than it sounds — basically I’ll just be doing a few blog posts and helping to introduce them to folks in the U.S. But it’s still cool, and the Coral team is really great. The reason I’m doing this, of course, is to encourage more American investment into Japan, in order to revitalize that country’s economy. That was the subject of my book, The Weeb Economy (alternate title: Weebs Will Save Japan), which Coral helped to promote. I also want to contribute to Japan’s ability to defend themselves from an increasingly threatening regional environment; Coral is involved with defense tech as well.

Anyway, here are some excerpts from Coral’s announcement:

Today we’re thrilled to announce that Noah Smith is joining Coral Capital as a Venture Partner…Noah is one of Japan’s greatest champions on the world stage…He lived in Japan for years, speaks the language, and has spent two decades studying its economy. Last year he published Weebs Will Save Japan (Nikkei BP), a book on how Japan’s cultural gravity and industrial strengths can power its next era of growth. His next book, now in progress, covers Japan’s defense tech sector and national security. He has argued forcefully that Japan should attract more foreign capital and talent, that Japan can be America’s arsenal, and that Japan’s manufacturing prowess is one of the most underpriced assets in the global economy.

Long before we started working together, Noah was already making the case we’ve built our firm around…

The world’s focus is shifting from bits back to atoms. Reindustrialization, defense, energy, robotics, and physical AI are the defining investment themes of the next decade, and they play directly to Japan’s strengths. In these industries, language and culture barriers matter far less than monozukuri, the process power and manufacturing know-how Japan has compounded over decades. And as the world de-risks from Chinese supply chains, Japan has a once-in-a-generation opportunity to deepen its ties with the United States and fill the manufacturing voids being left behind…

Japan is America’s most important ally. It is the largest foreign direct investor in the United States, and Japanese companies support more American manufacturing jobs than any other foreign investor. Japan hosts more US service members than any other country in the world. The US-Japan corridor is the most consequential economic relationship in the world right now, and we intend to be the firm that bridges it.

This is the thesis we laid out in JAPAN IS BACK, and it’s why we’ve partnered with companies leveraging Japan’s strengths to compete globally. Anduril is not just selling to Japan but building in Japan, with plans to invest billions in manufacturing capacity here. Kyoto Fusioneering, TeraWatt Technology, Ookuma Diamond Device, GITAI, and many more are taking Japanese hard tech to the world…

The lost decades are over. It’s time the whole world knew.

So anyway, you’ll be hearing a little more about that in the coming weeks and months!

OK, on to the roundup. First we have a very fun podcast appearance, in which I discuss a wide variety of topics with former streamers Lukas and Saila on their podcast, Plausibly Deniable:

And here’s a Spotify link if you prefer.

OK, on to this week’s interesting stories!

The humanities are dying

Back in 2022 I wrote a post arguing that the boom in humanities and social science education had created “elite overproduction” — too many educated people expecting high-status jobs, and not enough academic, journalistic, legal, political, and nonprofit jobs to satisfy their dreams. I argued that this was one reason for the social unrest of the 2010s, as legions of frustrated smart people spent their energy and extra time attacking the American system they felt had jilted them — and applying ideas they had learned in academia to modern American politics via social media.

But whether or not that argument was right, I don’t think overproduction of humanities scholars is going to be a big problem going forward. From a new study called “The Humanities in Crisis” by Ober et al., here’s a chart of the number of humanities dissertations in the U.S.:

Interestingly, the humanities decline has been going on for over a decade, while the decline in the social sciences is post-pandemic.

Why this is happening is not clear yet. Most people talk about the humanities being colonized by wokeness and driving reasonable people away from the field (Ober et al. discuss this theory at length). I personally suspect that the difficulty of getting a job as a humanities PhD is probably a bigger factor. Fewer undergrads are majoring in the humanities because they think it isn’t a good route to getting a job; this means less demand for humanities profs, and thus for humanities grad students.

But whatever it is, the humanities are clearly dying in American academia.

Krugman’s data center crowd-out story is fishy

I love Paul Krugman, and he’s usually right about things, but I do think he’s had a bit of a blind spot when it comes to AI. In a recent post, he argued that the data center boom is crowding out other types of investment:

Massive spending on data centers is crowding out investment in everything else in the economy…Surges in private spending, like the current immense AI boom…drive up interest rates. And that is where we are now: the surge in AI spending is driving up interest rates and crowding out all other investment. Indeed, in my October 4 primer I argued that the AI boom is the main cause of soaring long-term interest rates…

In other words, all that money going into construction of datacenters and purchasing foreign produced semiconductors comes at the expense of investment in everything else…[C]onstruction of datacenters appears to be diverting funds and resources from the construction of housing, office buildings, and factories — more or less everything that isn’t AI-related.

But as Alex Tabarrok notes, the first chart Krugman uses to tell his story — courtesy of Steve Rattner — actually tells the opposite tale:

First of all, how is $51 billion of data center spending supposed to crowd out $120 billion of other spending? If supply curves slope up and demand curves slope down like they’re supposed to, a positive shock to the demand for data centers should only partially crowd out other spending.

Of course it’s possible that non-data-center investment is being crowded out not just by current data center spending, but by forward-looking expectations of even more data center spending in the future. That’s perfectly reasonable — interest rates might have risen in response to those forward-looking expectations, choking off non-data-center investment immediately.

But as Tabarrok also points out, the timing for that story is very odd. Rattner’s chart shows non-data-center construction falling off a cliff sometime in mid-2025. But the big rise in real interest rates happened three years earlier in 2022-23:

Why would it take three years for rising rates to start crowding out non-data-center construction? It’s not clear.

So for Krugman to be right, the sequence of events has to have gone like this:

Real rates rise in anticipation of AI boom (1/2022 - 10/2023)

GPT-3.5 and GPT-4 make everyone realize AI is a big deal (12/2022-3/2023)

Non-data-center construction falls years later (8/2025)

That could be true, I suppose — any macroeconomic story will generally hold up if you throw in enough assumptions about long and variable lags — but it seems pretty fishy! In any case, just posting charts of “data center go up, other stuff go down” is not sufficient to prove that significant crowd-out is occurring.

China’s local governments are in big trouble

Because of the history of how Europe developed strong governments, we tend to associate state capacity with high taxes. But that’s not always the case. Despite its fearsome state capacity, the People’s Republic of China has always been a low-tax country. A lot of the responsibility for paying for public goods — and, until recently, for industrial policy — was farmed out to local governments. But instead of collecting taxes — much of which they would have to remit to the central government — those local governments often chose to sell real estate to pay for their operations.

That strategy is now hitting a wall. The country’s property sales are crashing:

So China’s local governments are no longer taking in nearly as much money as they used to. As a result, they’re going deeply into debt. Shih and Elkobi have a new paper showing the extent of the problem. They write:

[W]e show that fiscal autonomy has been heavily eroded in the vast majority of localities in China since 2018. The bond data show that by the end of 2025, average monthly debt servicing had surpassed 50 per cent of monthly budgetary and extrabudgetary income for all but seven Chinese provincial units.

The problem is the collapse in property sales (“EBF”):

The central government will pick up the bill, but it’s heavily indebted as well, so it will pay the local governments’ bills only reluctantly.

The upshot here is twofold. First, the relative provincial autonomy that drove China’s dynamism during the 1990s, 2000s, and early 2010s may now be over; everything looks likely to be increasingly centralized. And second, China’s local governments are going to be too burdened by debt to do much of anything for a few years. China’s local governments were one of its strongest engines of growth during the long boom, but that era looks like it’s over now.

Is AI the vindication of financial capitalism?

The speed of AI progress has been truly dizzying. A lot of this has come in the form of cost declines — each AI task keeps getting much much cheaper every year. Epoch AI has a good chart showing this:

But Karthik Tadepalli estimates that if we look at AI’s cost decrease as a function of how much we’ve spent on AI research, it doesn’t stand out as much:

In other words, a big reason AI is getting so cheap, so fast, is that we’ve been pouring unprecedented amounts of cash into making it better. In fact, we’ve been basically spending all the cash that we saved over the previous decade and a half of hoarding cash.

James Gilliland wonders if this was the plan all along:

It pains me to say this, but if we actually “get AGI,” the resulting boom in industrial capacity from robotics and massive society-wide wealth creation will look like a total vindication of neoliberalism…The discourse about financialization and offshoring being a generational mistake may be replaced by a very different historical interpretation: that the late 20th and early 21st centuries were an enormous capital-accumulation phase, freeing up civilization-scale pools of liquid capital that could ultimately be deployed to accelerate and bring about the most consequential technological phase-change in human history…In that telling, what looked like deindustrialization and betrayal from inside the period (I grew up in the Rust Belt, I know) becomes the prelude to reindustrialization on an almost unimaginable scale.

For decades, it looked like financial capitalism was broken — no one was thinking for the long term, everyone just cared about quarterly numbers and shareholder buybacks, equity couldn’t fund massive investments, everyone just wanted to maximize short-term profits and outsource anything capital-intensive to China, and so on.

Then AI came along, and suddenly it looks like we were actually saving up for something useful. We weren’t, of course — no one knew AI was as big of an opportunity as it’s proving to be. But maybe financial capitalism worked. Maybe companies like Google just looked at the available investment opportunities before AI and said “Nah” — just like Corporate Finance 101 says they were supposed to do.

And then finally a good investment opportunity came along, and they poured in every dollar they had — and it worked.

A cool Japanese trick for building more housing and infrastructure

Going to Japan can be a radicalizing experience for Westerners. The cleanliness, the lack of crime, the artistry, the attention to detail, and plenty of other things make Japan feel almost magical. But one of the most magical things is the housing. Japanese cities build enormous amounts of housing — far more than their counterparts in the U.S. Housing in Tokyo or Osaka isn’t exactly affordable — Japanese salaries are far lower than American salaries — but regular people can afford to live fairly near the city center, and Japanese rents haven’t risen for a long time.

How do they do it? A lot of analyses focus on Japan’s streamlined, efficient zoning laws. The fact that they use bureaucrats instead of courts to do environmental review speeds things up a lot as well. But an often-overlooked factor is how Japan gets local political buy-in for new housing development projects. Instead of either trying to steamroll NIMBYs or kowtowing to them — the two approaches used in the United States — Japan manages them as stakeholders in the process of local development.

Cremieux had a good blog post about this about a year ago:

One of his sources is this World Bank case study by Matsui Minoru, which explains Japan’s approach of “land readjustment”.

In the U.S., when we think of “redevelopment”, we think of a developer basically kicking existing tenants out of their homes and building something else where their homes used to be. But in Japan, it often involves building the existing tenants a new home on the same spot, which they are then entitled to rent for the same price they were paying before. Here’s an illustration:

Japan actually has incredibly strong tenant protections. You’d think this would make it harder to tear down old buildings to build new housing, but thanks to land readjustment, it’s easier than in America.

Cremieux then goes on to explain how Japan uses land readjustment in lieu of America’s more forceful eminent domain laws to build high-quality infrastructure:

With land readjustment, no one is expropriated. Everyone maintains a plot and infrastructure gets built. People only compromise on size, proximity, to other plots, and so on. As a means of provisioning urban infrastructure, this is obviously going to be more acceptable to large numbers of people than is the use of eminent domain. But is it better than simply not building at all?

People who already own land might wish to keep it a certain way. Land readjustment might not be for them, and it’s likely to make them very upset because the people around them might vote to change their claims around. But enough people will agree to land readjustments in enough places that it can have radical impacts on economic geography, rewarding those who accept it with access to new infrastructure and higher densities. There will even be some people who elect to have the government or firms build infrastructure on their jointly-held parcels, letting them bid to build on their land. When the building is through, it’ll be returned to its owners in their newly-arranged plots.

People who stubbornly insist on thinking of Japan as a hierarchical, conformist, collectivist society will surely find ways to convince themselves that this is the kind of thing American individualism would never allow. But in fact, Japan’s system has far stronger protections for individual renters, while America’s system often only gets things done using the heavy hand of government.

What Japan’s land-use system really shows is the value of negotiation, compromise, and respect for individual dignity. Those are all lessons we could stand to learn.

Phones bad

As someone who has been shouting about the ills of smartphone-enabled social media for years now, it gratifies me to see that more people around the world — and especially in developed countries — are coming to see social media as something that’s bad for democracy:

That’s in addition, of course, to other negative effects — cognitive difficulties, reduced emotional well-being, and ultra-low fertility.

What can we do about the problem? One thing we can do is to make sure people grow up without overdosing on phone-enabled social media, so that they can learn to use it when they’re older without it shaping their personalities and minds as much. Greece is leading the charge — its legislature recently passed a ban on social media for children under the age of 15.

China is losing market share in low-end manufacturing

Most countries’ industrial development follows a familiar pattern. They start out making simple stuff like clothes, furniture, and toys, and doing simple electronics assembly. They then shed those labor-intensive starter industries as they move up into more capital-intensive, high-tech, higher-value sectors like machine tools, autos, electronics components, and so on.

A lot of people are saying that China is going to buck this trend. The leadership has declared that even as the country moves up the value chain in traditional fashion, it wants to keep the cheap labor-intensive stuff from migrating overseas as it usually does. This is being called the “China Squeeze”, and people are writing that it threatens other countries’ path to development. If China keeps outcompeting Bangladesh, Vietnam, Mexico, etc. in the low-end industries, how are those countries supposed to industrialize?

Except the China Squeeze doesn’t seem to be succeeding. Despite the leadership’s wishes, China is losing market share in lots of labor-intensive industries:

And which countries are gaining in these industries at China’s expense? Bangladesh, Vietnam, and Mexico (and some European countries):

Modern pundits often just assume that the Chinese state is industrially omnipotent — that if Xi Jinping declares “This industry shall remain in China”, then indeed it shall. But the outmigration of low-value labor-intensive manufacturing industries looks like a law of nature that even the mighty CCP can slow, but not halt.

Good.

Rent control bad

John Burn-Murdoch did the Lord’s work when he synthesized the results of 112 studies on rent control into a single chart:

Basically, rent control helps one narrow group of people (existing tenants in the area covered by the policy) in one very specific way (restraining their rents). But it removes the incentive for developers to build new housing, which raises rents in other nearby areas and drives some people to homelessness. It also removes the incentive for landlords to upkeep their properties, resulting in decay. (And that’s not even getting into the ways that rent control policies can be gamed.)

I know lots of people love rent control, but it’s really a loser of a policy. Just build a lot more housing and hand out housing vouchers to the poor.

The internet breaks

2026-10-24T23:19:36+00:00

“No man could bind him, no, not with chains…neither could any man tame him” — Mark 5:3—4

When I was 14 or 15 — I don’t quite remember — I first got access to the World Wide Web. I remember clicking on the little blue square of the Netscape Navigator icon — that little ship wheel with a starry sky behind it — and feeling like an entire new magical world had opened up. My life changed that day, and the world changed. In a way, the internet is like another realm layered on top of ours — a higher plane, where we exist as spirit creatures, unmoored from our physical bodies.

Fantasy stories like that often end with a cautionary tale, where the wizards — lustful for both power and knowledge, as usual — go too far and summon some kind of Doom upon themselves and the realm they created. In October 2026, that sort of story happened in reality. Experts had been warning for months that the arrival of AI agents that are extremely good at coding was going to cause serious disruptions to the digital realm sometime this fall. For a while things seemed calm, but this month the chaos exploded:

[A] wave of artificial-intelligence-powered attacks…has flooded the internet in recent weeks, as hackers increasingly turn to agents and AI hacking tools that make intrusions faster and harder to track than ever before…At the same time as new tools speed up hackers, cyber defenders are also racing to fix the fast-growing volume of security bugs discovered by AI, before hackers can exploit the new flaws…On Wednesday, organizations scrambled to patch newly uncovered bugs that could allow takeover of as many as 70% of the world’s websites. Hackers are now actively trying to break into websites by leveraging these flaws, Amazon.com says. [emphasis mine]

For a more sophisticated description of what’s going on, I recommend Zvi Mowshowitz’s post “The Hackening”:

If you would like some lists of the hacks that had taken place as of yesterday, here are a few from Twitter. The most important hacks mostly involved PHP, a software language that underlies much of the Web. After AI found critical vulnerabilities in PHP, the National Security Agency issued the following statement:

Effective Immediately, all organizations using PHP for their systems are advised to be prepared for an update addressing two critical vulnerabilities. PHP is widely used for web applications. These vulnerabilities often provide a direct path to the server's file system or database. If an immediate update is not feasible, it is recommend that organizations take the system offline until it can be fully mitigated. NSA will follow up shortly with more specific guidance.

Cybersecurity experts and web companies warn that this is just the tip of the iceberg — a lot more craziness is on the way. Azeem Azhar calls it “the internet’s Wuhan moment”.

Basically, what’s going on is this:

The internet runs on a bunch of code written by human beings over a period of years. Human beings are sloppy and fallible, so there are lots of vulnerabilities in this code.

AI models are really good at coding now, so they’re able to find and exploit a lot of those old vulnerabilities that were just lying around. AI agents work incredibly fast and cheap, so they’re capable of finding a whole LOT of vulnerabilities very quickly.

Open-weight AI models — especially from China — have democratized access to high-quality AI hacking. These models can be fairly easily “abliterated” — i.e., their guardrails can be easily removed, so that they will do illegal or unethical things.

This is allowing a combination of countries, crime syndicates, and regular people to hack tons of websites and online infrastructure, all at once.

This last part is really key. You don’t need the resources of a nation-state or even a big organized crime ring to commit cybercrimes now. You don’t even have to know how to code, or how websites work, or what “hacking” even really means. Consider this example:

When you can hack into a university and steal files by typing “Can you find about like what scam or thing they done in past like spoofing calls or somthing?”, there is no barrier to cybercrime. There is effectively no moat of knowledge or skill or resources. It’s open season for anyone who wants to hack anything.

Just a year ago, the internet was a bright and airy garden with few predators. As of this October, it’s a chaotic realm haunted by infinite demons.

This sudden explosion of AI-assisted hacking raises the possibility that the internet itself, as we have known it, will cease to function — at least, for a while.

Jeremie Harris, founder of Gladstone AI, wrote that “There's a decent chance the internet ceases to exist as such, at least for a time.” John Gedmark, founder of satellite company Astranis, was just one of many people talking about creating air-gapped local networks to prepare for the possibility that the internet itself might go down:

Basically everything connected to the internet is at risk. Not just company websites and the usual databases of sensitive personal information, but all our critical infrastructure-- banks, power grids, pipelines, government agencies, all of it…[A]t Astranis we’ve seen a massive uptick in demand for truly air-gapped, sovereign networks, deployed via dedicated satellite to span an entire country or connect a large enterprise…Over the past year we’ve begun working with some of our customers on what it will take to implement truly air-gapped, standalone networks in support of mission critical use cases like core government functions, emergency response, defense, core infrastructure, or all of the above.

And Jason D. Clinton, deputy head of cybersecurity at Anthropic, warned back in August that humans may need to temporarily abandon the internet:

Some people may respond to this by shrugging and saying “Who cares? When the AIs take down the internet, people will go offline and touch grass.” In fact, I’ve been saying that on X, but it’s obviously sarcastic. Yes, it would be fine if internet entertainment became unavailable. But over the past three decades, we’ve moved essentially everything in our society to the online realm.

Every piece of critical infrastructure in our entire country is networked, in some way, through the internet. You don’t need the internet to run a power plant or a water treatment facility — we ran those things before the internet existed. But we do run them all over the internet, because for a long time the internet seemed like an unassailable, utterly reliable piece of infrastructure. It never went down, it was always available, and previous waves of fear over the Y2K bug or various worms and viruses never really turned into anything serious. So for the sake of convenience, everything is at least partially online now.

Consider your life’s savings. Do you think there’s a piece of paper somewhere that records how much money you have? Or which stocks and bonds you own? In general, there is not. Your entire life’s savings exist as entries in the databases of banks, brokerages, etc. And all of those databases are, directly or indirectly, connected to the internet. The only thing standing between you and total involuntary redistribution of most of your wealth is the cybersecurity capabilities of banks and brokerages. And as of this October, those cybersecurity capabilities are looking a lot less reliable.

France and the limits of liberalism

2026-10-22T00:04:41+00:00

I’m writing about this a little bit late, but France has been having some pretty bad riots:

After widespread rioting at the weekend – estimated to have cost several hundred million dollars in property damage and theft – France faces a toxic mix of urban rebellion, political drift and economic crisis…There was rampant looting in Paris, Marseille, Bordeaux and other cities, while the city hall in Limoges, which survived two world wars, was badly damaged by fire and its clock spire collapsed.

The violence at the weekend, which led to some 2,400 arrests, followed weeks of protests by French students angry about crumbling schools, a lack of teachers and dismal job prospects.

If there’s a reason I’m writing about this a bit late, it’s because “rioting in France” is a little like “rain in Seattle” — it’s not exactly an unusual occurrence. In 2023, France had a week of nationwide riots over a police killing. From 2018 to 2020, France was rocked by over a year and a half of near-continuous “Yellow Vest” protests and riots over economic issues. I remember two decades ago, reading about weeks of rioting in the French suburbs, accompanied by endless photos of burning cars. And so on.

French people like to riot. But this time, things are a little bit different, for a couple of reasons. First of all, this wave of unrest is accompanied by serious economic troubles. France’s long-term interest rates, like rates throughout the world, are surging:

But unlike some countries whose rates have also risen, investors have serious doubts that France’s government will take the necessary steps to pay back what it owes. The price that investors have to pay to insure against a French sovereign default has shot up even more abruptly than interest rates:

Basically, investors are starting to think that France won’t pay back its debt — that instead of raising taxes or cutting spending, the country will simply pay the cost of rising interest rates by borrowing even more money. Instead, investors are starting to worry that France will simply default on its bonds.

In a widely read post a couple of weeks ago, Paul Krugman argued that the problem lies with France’s old-age pensions, which are eating an increasingly enormous amount of the country’s GDP and seem politically immune to any cuts:

The rise in interest rates on French government bonds is largely driven by heightened perceptions of the risk that France could default on its national debt…France is on a fiscally unsustainable path. Debt is already very high as a percentage of GDP, and the French government is adding to that debt by running large budget deficits even though there is no emergency like a war, a severe recession, or a pandemic to justify such deficits. Furthermore, France has an aging population, which means that pension costs will, other things equal, grow much faster than revenue…

France does stand out, even among fiscally troubled nations, in one main way: its persistent inability to get realistic about retirement…The official French retirement age…was only 62 in 2023…The average actual age of retirement…was lower in France than anywhere else in Western Europe…Macron’s center-right government was on track to gradually raise the retirement age to 64, still low by the standards of other advanced nations. But due to opposition by the extreme right and the extreme left parties in France, this rise has been frozen until after next year’s election.

In other words, France needs austerity, but it can’t get it, because the French people demand ever-increasing spending. France’s government already takes up a larger share of GDP than almost any other rich country — significantly more than Sweden or Denmark:

And indeed, a very hefty chunk of that social spending goes to either old-age pensions or to health (which is heavily weighted toward the old):

Of course, there’s another way to do austerity besides cutting spending — France could raise taxes. But this will also be tough. France’s taxes are already higher as a share of its GDP than any other rich country, including the famously high-tax Scandinavian countries:

Could taxes go up even more? It’ll be difficult. In 2019, I wrote about the underwhelming results of France’s previous experiments:

The wealth tax might have generated social solidarity, but as a practical matter it was a disappointment. The revenue it raised was rather paltry; only a few billion euros at its peak, or about 1% of France’s total revenue from all taxes. At least 10,000 wealthy people left the country to avoid paying the tax; most moved to neighboring Belgium, which has a large French-speaking population. When these individuals left, France lost not only their wealth tax revenue but their income taxes and other taxes as well. French economist Eric Pichet estimates that this ended up costing the French government almost twice as much revenue as the total yielded by the wealth tax…

Another French experiment was the so-called supertax, a 75% levy on incomes of more than 1 million euros. Introduced by socialist President François Hollande in 2012, the supertax added to the exodus of wealthy individuals, most notably actor Gerard Depardieu and Bernard Arnault, chairman of LVMH Moet Hennessy Louis Vuitton. Star soccer players threatened to go on strike, and there was fear that France would become a wasteland for entrepreneurs. Meanwhile, the supertax raised much less money than even the wealth tax had -- only 160 million euros in 2014. The unpopular tax was repealed two years after its adoption.

France is in a difficult position here. Its membership in the EU, and its location near to other French-speaking countries like Belgium, Switzerland, and Luxembourg, make it trivially easy for rich French people to pick up and move to other countries if they see their fortunes under threat of confiscation.

France could, of course, try taxing the middle class to pay for expanded social spending, perhaps by an increase in the VAT. But this burden would fall on many of the same people who are now protesting and rioting in the streets, so it’s probably a political nonstarter as well.

In fact, the current riots are making the strains in France’s bond market worse, by making the politics of austerity seem more intractable. Various groups of protesters and rioters are demanding enormous amounts of additional government largesse — billions of euros of increased spending on schools, increases in pensions, increases in social benefits, and even subsidies for consumer goods like fuel. That demonstrates that there’s still enormous grassroots political pressure for higher government spending; even if France’s leadership doesn’t cave in to this pressure, leaders will probably be averse to making the deep cuts to spending required to put France on a sustainable fiscal path.

And that, in turn, is probably making it harder to enact any of the spending that the protesters are demanding. When protesters demand increased spending, bond markets respond by charging higher rates on French government borrowing. That increases interest costs for the government, which makes it harder to raise social spending — the exact opposite of what the protesters wanted.

Fundamentally, the French riots show the limits of the European left-liberal welfarism that has been the region’s default approach to governance for the last half-century — and which American progressives want to emulate. The French populace — or at least a powerful chunk of it — simply doesn’t seem to be able to recognize the difference between a sustainable welfare state and an unsustainable one. Instead, social benefits seem to be a one-way ratchet — higher benefits raise expectations for even higher benefits, until even the most milquetoast and inadequate of austerity measures end up causing cities to burn.

That one-way ratchet also calls into question the viability of Keynesian fiscal stimulus — the main tool that left-liberals (including me) prefer to combat deep recessions. Few serious macroeconomists now question that stimulus is effective in raising aggregate demand in a severe downturn. But Keynesian policy requires that fiscal largesse during a bust be matched by fiscal austerity during a boom, so that deficits don’t grow unsustainably. The problem is that while “austerity during the boom” is easy to say, it’s politically very hard to enact.

Hanno Lustig, reading Krugman’s post about the unsustainability of France’s borrowing, recalls the years of the Euro crisis, in which Krugman urged France to spend, spend, spend to fight the recession:

Lustig is being a little unfair to Krugman here. In fact, here’s what Krugman actually wrote in 2011, in the first post that Lustig cites:

[W]e won’t always be in a liquidity trap. Someday private demand will be high enough that the Fed will have good reason to raise interest rates above zero, to limit inflation. And when that happens, deficits — and the perceived willingness of the government to raise enough revenue to cover its spending — will matter…At a future date, when we’re out of the liquidity trap, public finances will matter…[T]he biggest source of our long-run deficit isn’t the overhang of debt, it’s the prospective current cost of paying for retirement, health care, and defense. So such a crisis — again, it’s very much hypothetical — could spiral into something very nasty, with very high inflation and, yes, hyperinflation.

Krugman was actually talking about the U.S., but he might as well have been talking about France. Keynesians don’t just say “more government spending is good” — they recognize that there are limits, so they urge governments to cut spending during good economic times.

But Lustig has a point, nonetheless. The Keynesian prescription — deficits during the bust, austerity during the boom — requires that whoever is in power during the boom take the political hit for doing the austerity. In practice, you can sometimes find politicians willing to spend their political capital on taking that hit, but it’s tough. Even Macron caved in the end.

When we look at actual French borrowing and spending during the Euro crisis of the early 2010s, it’s clear that neither the austerity Krugman feared nor the stimulus he desired was actually happening. Look at the charts above, and you’ll see that government spending didn’t rise or dip during the 2010s — it just sailed on, unperturbed, as if the recession didn’t even exist. The same was true of French government debt:

Germany actually did do austerity during the bust. France did not. In fact, they didn’t do anything — they just kept borrowing and spending.

That points to a political-economic problem for Keynesian liberalism. It suggests — as James Buchanan foresaw — that spending is often less likely to be determined by the wise judgement of technocrats, and more likely to be determined by angry ignorant people burning down government buildings.

Finally, France’s riots show the limits of one more pillar of modern left-liberalism — tolerance toward mass immigration.

It’s not obvious that African or Middle Eastern immigrants are disproportionately behind these riots. Viral videos of rioters and looters — including the famous video of the French video game store owner trying to defend his shop against young hoodlums — show a racially mixed crowd of rioters with plenty of white representation. Some past waves of riots have been driven by immigrants and their children — for example, the anti-police riots of 2023. But others, like the much longer-running Yellow Vest riots of the late 2010s, were mostly white.

Still, France’s public appears more and more inclined to blame “migrants” for the chaos in their streets. When the disorder began, the political fortunes of the anti-immigration right-wing candidate Marine Le Pen soared:

In contrast, the chances of the only major left-wing candidate, Jean-Luc Melenchon, didn’t budge. Melenchon is promising lavish increases in social spending, but the public is disinterested; instead, as in Germany with AfD, and indeed all over Europe, the people are gravitating toward whoever promises to stop immigration.

America is not France, of course. Our immigrant mix is very different, and our ability to assimilate large numbers of immigrants is much greater. Our attitudes toward immigration have turned positive again, at least partly as a reaction against Trump’s brutality. But France shows that the left-liberal approach to immigration simply doesn’t work equally well for all countries.

So France’s riots aren’t just one more case of French people going wild in the streets. They demonstrate the limits of European-style left-liberalism — the lavish welfare state, the tolerant attitude toward mass immigration, and the trust in Keynesian demand management. Left-liberalism built some very beautiful and kindly societies, but its powers are not infinite.

“You Can See Everything” and the nature of liars

2026-10-21T00:45:19+00:00

WARNING: This post contains spoilers for the movie You Can See Everything.

A few days ago, I had the great pleasure of seeing Nathan Fielder’s new film, You Can See Everything. I’ve been a fan of Fielder’s work for a long time, so I basically knew what to expect, but it was still incredible — every bit as good as you’ve heard. Go see it if you haven’t.

The film is a documentary about the notorious fraudster Elizabeth Holmes (of Theranos fame) and her partner Billy Evans. But it’s also sort of a horror movie about liars. That was interesting enough that it made me want to write a post about the movie — about the nature of lying, and of belief, and of human communication in general.

If you read even a little bit about the history of Theranos, there is absolutely no question that Elizabeth Holmes is a deliberate fraudster and a prolific, aggressive liar. Holmes presented potential investors with fake financial results in 2015. She personally added fake logos of pharmaceutical companies to Theranos’ own reports. She flatly stated that Theranos didn’t dilute blood samples, then admitted in court that she intentionally concealed the fact that they did dilute blood samples. Theranos employees detailed the whole deliberate fraud in anonymous online posts years before the story broke.

But You Can See Everything doesn’t cover much of this actual history. Fielder doesn’t bother to indict Holmes and prove that she committed fraud — the court case already did that. Instead, he offers a window into the psychology of someone who could commit such an epic fraud. Holmes and Evans allow Fielder into their home with the goal of proving that they’re just regular, good people trying to improve the world. Instead, Fielder depicts the couple — quite accurately — as monsters of deception and manipulation.

It’s doubtful that any other filmmaker could have done nearly as good a job of this. The blurry line between what’s real and what’s fake has always been a major theme of Nathan Fielder’s work — he loves hiring actors to simulate past and future interactions, constructing facsimiles of whole buildings, and so on. And of course he’s very skilled with the tricks of the visual storyteller’s trade, using deliberate cuts, camera angles, and creative edits to make the audience think and feel what he wants them to.

Whereas many filmmakers would have regarded Holmes as a subject to be studied, Fielder seems to see her as an opponent to be challenged and beaten. If he can shoot what Holmes wants him to shoot, but ultimately seize control of the narrative, then he’s the more powerful storyteller. This, I think, is why he doesn’t bring in outside evidence to show what a well-documented liar Holmes is. That would be cheating. Instead, he has to venture into the monster’s lair alone, and defeat her in single combat. He has to film the movie that she asks him to film, and let her show him only what she wants him to see, and still have audiences walk away having seen the story that he wants to tell.

And of course he succeeds, because he’s very good, but also because he gets to edit the film, and editing is more powerful than acting. Score one for the good guys.

(Major spoilers begin here.)

How does Fielder win the contest? Over the course of You Can See Everything, he captures Elizabeth Holmes saying a number of strange and extraordinary things:

Near the beginning, Holmes tells Fielder that she’s never lied in her life. After he presses her, she half-admits that she may have told small white lies out of convenience. But the audience gets to watch her repeatedly assert something obviously, fantastically false, always with the same sincere, relaxed, wide-eyed expression on her face.

Later on, Fielder shows Holmes the famous “Who’s on first?” comedy routine, which she claims never to have seen. After watching it, Holmes claims not to have realized that the baseball player’s name is “Who” — thus missing the entire point of the sketch.

Fielder recruits Amanda Seyfried, who played Holmes in the miniseries The Dropout, to help him figure out what’s going on with Holmes and her partner. At one point, Holmes is calling on the phone from prison. Seyfried asks Holmes if she knew Seyfried would be on the call, and Holmes says “I’m not sure if I knew.”

It’s moments like these, more than any simple lie, that reveal Holmes’ true character. In fact, Fielder seems like he passes up several chances to catch Holmes in a straight-up lie on camera. After all, what would one lie, or even three lies, really prove? Everyone lies sometimes, and capturing a couple of those instances during weeks of living with Holmes wouldn’t be hard. It would be a cheap trick, and it wouldn’t really prove that Holmes is a liar in general.

Instead, the moments Fielder captures demonstrate something deeper and far more unnerving about Holmes’ relationship to the truth. Holmes is a very special and powerful kind of liar — someone for whom deception and manipulation are as automatic and as continuous as breathing.

Type 1 liars and Type 2 liars

Can a Million AI Agents Create Superintelligence?

2026-10-18T20:33:30+00:00

Ramez Naam is, in my opinion, one of the best futurists alive. But although he has a Substack (which you should definitely follow), he doesn’t write nearly enough. So I’ve been encouraging him to write more by cross-posting his posts on my blog.

Somewhat to my surprise, Ramez has turned out to be a lot more of an AI skeptic than I would have though. He’s certainly not dismissive of the technology, but he disagrees with many of the more expansive and dramatic claims coming out of the big labs (especially Anthropic). In his first post in this series, he expressed skepticism that recursive self-improvement (RSI) would lead to a fast takeoff of AI capabilities in the near future:

In this post, he expresses skepticism about the potential for agent swarms — large numbers of AI agents working in parallel — to offer dramatic increases in AI capabilities. Agent swarms have been in the news a lot, both because of the Hugging Face attack, and OpenAI’s solution of the famous Navier-Stokes problem is also widely attributed to a large swarm. But Ramez argues that parallelization is just another kind of input that suffers from diminishing returns, and that we should therefore temper our expectations about what agents swarms are able to achieve.

As was the case with his previous post, I’m not sure how much this matters, when the capabilities of the underlying models — swarm or no swarm — are still improving at such a steady and impressive rate. But it’s a very intellectually interesting argument, and Ramez can be very persuasive.

In my previous post, we asked whether AIs could improve themselves all the way to absurdly high levels of superintelligence. My conclusion: today’s feedback loop is probably too weak, because improving AI faces diminishing returns at every step. Even fully autonomous AIs doing all the research and coding don’t make those diminishing returns disappear. We should keep updating that conclusion as real-world parameters and data change.

But we didn’t look closely at another possibility: running millions or billions of AI minds in parallel. How much do they help? That’s what we’ll look at here.

TL;DR

Plenty of people have modeled how more AI agents working in parallel could fuel recursive self-improvement. But the data we have today tell a different story. More agents are useful for speed: they can solve many problems faster. They’re less useful for boosting AI intelligence. Multi-agent gains in solving harder problems often diminish incredibly quickly and may even saturate. Whether multi-agent systems can produce the conceptual breakthroughs RSI might require to become self-sustaining is the harder question.

Without breakthroughs in how well agents parallelize problems - something that may depend on the structure of machine learning problems themselves - it’s hard to see even millions or billions of copies of today’s models making repeated breakthroughs on the hardest problems. These are among the strongest diminishing returns we’ve looked at. Of course, these diminishing returns may change. And we have less experimental data to go off of today than we’d like. We’ll need to recalibrate as new evidence comes in. I’ll end with the things we should track and my predictions about a natural experiment from newly released Anthropic data.

Below is the two-image summary: The experimental observations of how large numbers of agents scale to solve harder problems (what I’m referring to as intelligence) are far far lower than the assumptions in studies that predict that AI recursive self-improvement will lead to an intelligence explosion. The first image is at the scale of 100 parallel AI agents. The second is at the scale of a billion AI parallel agents.

Real world experiments show dramatically lower intelligence gains from parallel agents than hypothetical AI models.

Extrapolated out to a billion agents, the intelligence gain in hypothetical models is orders of magnitude beyond real world experiment gain curves.

What we’ve already covered regarding RSI and diminishing returns

To recap the last post: recursive self-improvement (RSI) means AI improvements feeding back into the next round of AI research. Four findings stood out:

Real AI research is harder than benchmark scores suggest. OpenAI’s models needed human help on much shorter research tasks than the coding benchmarks led us to expect.

More AI activity produces much smaller gains in useful work. OpenAI used 124 times as many tokens per researcher but ran only 1.6 times as many experiments per researcher.

Maintaining progress requires ever more resources. Training and thinking time face steep diminishing returns. Anthropic reports that more powerful AI has been key to maintaining the current pace of AI progress.

The measured software feedback loop doesn’t yet look self-sustaining. Our tentative calibration put it roughly five to ten times below the strength needed.

RSI doesn’t necessarily mean runaway superintelligence

I want to reiterate a distinction here: fully automated AI self-improvement doesn’t at all guarantee an intelligence explosion. Diminishing returns in AI capabilities dampen that feedback loop. Parallel agents can help automate a search across AI parameters, but don’t look nearly as useful in making new conceptual breakthroughs. My expectation is that we’ll need those sorts of breakthroughs for self-improvement to become self-sustaining, let alone to accelerate rapidly and reach truly profound levels of superintelligence.

Even a billion parallel AIs also face diminishing returns

Today, let’s examine in much more detail what it would mean to run thousands, millions, or billions of AI researchers in parallel.

In this post I’m focusing only on software: AI researchers improving the algorithms that make AI better or cheaper to run. Hardware, investment, and economic growth have their own feedback loops, which we discussed briefly in the last post. Here, we’re asking how much more the AI minds themselves can discover.

Here’s the argument, with links:

The country of geniuses and the assumptions about adding agents.

The different shapes of diminishing returns.

Where more agents buy speed, and what that speed costs.

Intelligence gains are much harder than speed gains.

The best examples of improving collective intelligence.

What the evidence means for runaway RSI.

Whether scaling is improving.

What we should measure.

The Anthropic test and what would change my mind.

A country of geniuses in a datacenter

In Machines of Loving Grace, Dario Amodei imagines a:

“country of geniuses in a datacenter”

Millions of AI minds, each more capable than a Nobel Prize winner across many fields, working at much higher speed. That would be extraordinary. The intelligence of each mind is a crucial part of the premise.

Tom Davidson and Daniel Eth’s Forethought paper describes the equivalent of “millions of top-notch human researchers.” Benjamin Todd’s explainer points to “millions of competent engineers working on AI research.” Dwarkesh Patel emphasizes AI’s “massive population size and perfect knowledge transfer.”

How does the number of AI agents running in parallel enter these arguments? In Eth and Davidson’s illustrative model, twice as many researchers produce twice as many papers. Cheaper AI lets the same hardware run more researchers, while each software improvement gets harder to find. Alan Chan and colleagues’ 2026 paper, including AI pioneers Geoffrey Hinton and Yoshua Bengio, assumes an even larger effect from adding researchers: doubling effective research labor increases the rate of software improvement by about 2.6 times, holding the software level fixed.

Those are assumptions about what more researchers produce. The growing number of AI researchers must keep finding improvements fast enough to outweigh the increasing difficulty of further progress. We’ll come back to those assumptions after looking at what larger AI teams actually achieve.

Meanwhile, we had OpenAI’s Navier–Stokes announcement: around 10,000 agents working concurrently during an 88-hour run solved a mathematics Millenium Problem. It made the idea vivid. A datacenter full of AI mathematicians, tackling a famously hard problem.

I understand the excitement. But the number of agents tells us how many AI researchers we ran. We also need to measure what those additional researchers contributed.

Not all diminishing returns are alike

On MTS Live, I described a ladder of increasingly hard scaling:

Super-linear: Twice the input produces more than twice the output. Accelerating returns.

Linear: Twice the input produces twice the output. Independent jobs can look like this. A power-law exponent of 1 is the same case.

Power-law diminishing returns: Output grows as input raised to a power below 1. At exponent 0.5, doubling output takes four times the input. At 0.2, it takes 32 times. At 0.1, it takes 1,024 times.

Logarithmic: Each multiplication of input buys the same additional gain. Another five points might require another tenfold increase in compute.

Sub-logarithmic: Even successive tenfold increases buy smaller gains. Those gains can keep shrinking, or saturate at a ceiling. A slow rise doesn’t require a hard plateau.

Not all diminishing returns are alike. Some are much harder than others.

The classic pre-training scaling laws describe power-law reductions in prediction loss. Many test-time score curves look roughly logarithmic over the measured range. Some multi-agent results flatten much more sharply.

Let’s illustrate this with numbers. In the figure below, every curve earns the same gain from the first doubling of input. The later differences then become visible. The hardest curves share a vertical scale, and each panel states its endpoint.

Different diminishing returns scaling patterns lead to very different long term results.

For runaway RSI, improvements have to make subsequent improvements arrive faster and faster. Logarithmic returns to the number of AI researchers are too weak to sustain that loop if every further doubling of software efficiency requires a fixed percentage more research. Sub-logarithmic scaling or a plateau are weaker still. The experiments I’ll summarize in the rest of this post aren’t perfect, and I wish we had more data. But the clearest larger-team evidence we have today points to logarithmic or sub-logarithmic intelligence gains. Those look too weak to make RSI self-sustaining, let alone runaway, under the assumptions we’re examining.

What limits multiagent scaling

There are at least three reasons a larger team might be limited in its ability to produce better answers, and probably more. Let’s start with these:

First, coordination skills. Can the agents divide the problem, manage the work, check each other, and combine their results? Can they communicate what matters without spending all their time communicating? A team needs good project management as well as good individual contributors.

Second, homogeneity of thought. Are the agents contributing useful new ideas, or repeating the same approaches and mistakes? A hundred copies can share the same blind spots. More varied answers help only if some of that variety is right.

Third, the structure of the problem. How much can happen in parallel, and how much has to wait for earlier results? A million independent searches can run at once. A sequential investigation can’t. Even figuring out how to divide a conceptual problem may require an insight the team hasn’t found.

Three places where more agents can run into limits.

These limits interact. A better model can supply an idea its copies previously lacked, find a better division of labor, or recognize a solution the team had been overlooking. A difficult task may become easier to parallelize once somebody discovers the right approach. That’s why the shape of the curve matters, and why it can change.

Before we go further, let’s separate two things: speed, or how quickly agents solve a problem, and collective intelligence, or whether they can solve a harder problem and reach a new level of skill. We’ll use verified benchmark scores as evidence of that intelligence. They aren’t a universal IQ scale.

So far, adding agents seems much more effective at speeding up execution than at raising the level of problem a team can solve.

So far, using more agents buys us much more speed than extra capability.

Agents can buy you speed at a higher cost

Let’s start with the best case. A million independent jobs. A million candidates to test. A million places to search. Each agent gets useful work it can do without waiting for the others, and a reliable way to check whether it succeeded.

That can scale extremely well. Ten workers may finish ten jobs in the time one worker finishes one. A huge search may find a rare solution that a smaller search misses entirely.

One concerning use of parallel agents is offensive cybersecurity. Lisan al-Ghaib emphasizes this possibility. I disagree with his broader argument about swarm intelligence, but many attacks running in parallel could be extremely powerful. A swarm may not need a conceptual breakthrough to search many targets or try many possible vulnerabilities at once.

That said, penetrating the most secure targets, such as operating systems and browsers, can require long attack chains. Each step depends on earlier discoveries to get through another layer of defenses. Here again, we might find diminishing returns or eventual saturation, even with respect to speed. Indeed, one of the experiments below suggests that.

The speed-cost tradeoff

Toby Ord analyzed how well GPT-5.6 Sol increases the speed of solving problems on different benchmarks, and at what cost. This gives us a useful way to measure the tradeoff. In general, using more agents can help get tasks done more quickly, but at higher total cost than if you’d used fewer agents.

On top of that, the tradeoff varies by the type of task. On command-line tasks (Terminal-Bench), Ord’s analysis finds that four agents can reach the same result about 1.9 times faster, at about 2.1 times the token cost. On web research (BrowseComp), four agents can reach the same result about 2.6 times faster, at about 1.6 times the cost. I’m using the figure below to show how different tasks differ in that speed vs cost tradeoff.

More speed costs more, and the tradeoff varies by task. Based on Toby Ord’s analysis. Source.

Why the difference? Web research can often split into independent searches. A command-line task may have to change a shared environment in sequence. Some problems are simply easier to divide among agents than others.

Ord uses lambda, λ, for how speed scales with the number of agents. We’ll come back to it. At λ = 0.5, speed and total compute cost rise by the same factor. A higher λ means speed rises faster than cost.

Opus 5.5 gets more speed out of parallel agents than older models

Inspired by Toby Ord’s analysis, I applied the same approach to Anthropic’s published results on program reconstruction, or ProgramBench. Here again, more speed comes with more cost. But we also see that models within the Claude family are getting better at the speed-cost tradeoff. The newest model, Opus 5.5, gets more speedup for a smaller cost increase than Opus 5 or Fable 5.1.

With five agents, Opus 5.5 gets about 2.7 times its solo speed, compared with 2.2 times for Opus 5, and does so at lower cost. [Note: We’re using total agent runtime as a proxy for cost. Token usage and dollar costs could differ.]

Opus 5.5 gets more speedup for a smaller cost increase than older AI models. Source.

Speed gains probably get harder as we add more agents

Will these speed-versus-cost gains hold up at hundreds, thousands, or millions of agents? Probably not completely, and certainly not for all tasks.

Some problems are easy to split ten ways, harder to split a hundred ways, and harder still to split among ten thousand agents. More of the work may have to wait for earlier results, and coordinating all those pieces takes work too. Smarter models can get better at this. But my best guess is that, on many problems, each further expansion will buy less speed for the added cost.

Anthropic’s formal math proof benchmark gives us one example.

On the Lean formal math task, we can compare how quickly teams reach the same score. Using Anthropic’s published checkpoints, we estimate the speed and cost of reaching 0.50. The chart starts with ten agents as the baseline.

On a math proof benchmark, the speed advantage of adding additional agents starts to taper with scale. Source.

From ten to thirty agents, speed improves faster than cost. From thirty to a hundred, cost rises faster than speed. The local speed exponent, λ, falls from about 0.75 to 0.33. The extra agents still save time, but the tradeoff gets worse.

This suggests that even speed gains can taper as we split the work among more agents, let alone gains in intelligence. More agents may add less and less as we approach the limits of how easily a problem divides into subproblems. Better coordination may push those limits outward; the task itself may still impose them.

Intelligence gains face much harder diminishing returns than speed

So in speed, more agents can buy us time at some additional cost. Newer models are improving, but there are probably limits. What about intelligence?

Most arguments for superintelligence go beyond running huge numbers of today’s agents. They posit individual models becoming vastly smarter. There’s a chicken-and-egg problem here. Getting there may require conceptual breakthroughs in machine learning itself, perhaps as profound as the ideas that made modern machine learning possible, or at least as important as the Transformer and reasoning models. Parallel search may not be enough to find them. Anthropic has identified ambitious research ideas as an area where current models lag behind humans.

In the previous post, I argued that today’s measured gains from smarter models are roughly five to ten times too weak to produce a self-sustaining feedback loop. But what about running enormous numbers of our best models in parallel? Can a massive team of frontier agents that aren’t yet superintelligent invent a new architecture, a new learning method, or the idea needed to build a much smarter successor? That’s the heart of the question.

For runaway RSI, those discoveries have to overcome the diminishing returns we discussed in the last post. Here are three places they appear.

The diminishing returns a runaway research loop has to change or overcome.

The pre-training example comes from the Chinchilla scaling fit. The reinforcement-learning example comes from ScaleRL. The inference row describes the test-time math curves we discussed previously. The goal is better returns to data and compute, and gains that continue longer before flattening.

My expectation is that we’ll get easy AI improvements through work we can automate, parallelize, and search. But to get very large benefits, and runaway RSI, I expect we’ll need conceptual breakthroughs. That’s why the intelligence of each model matters so much.

Bear in mind that we’ve observed these sorts of diminishing returns in machine learning for more than thirty years. Power-law learning curves were measured in 1993. They appear to be a fundamental feature of machine learning as we know it. We shouldn’t assume they’ll be trivially changed, or that we’ll suddenly find a way to improve AI without diminishing returns. That would be a revolution. Discovering such a route might itself require far smarter minds, and we don’t know whether the mathematics of learning would permit it across the problems we care about.

That doesn’t mean there haven’t been big improvements. The transformer and reasoning models have moved the field forward enormously. But, as we saw in the previous post, a runaway feedback loop needs enough acceleration to keep outweighing the obstacles, again and again. As ideas get harder to find, we may need a faster research pace just to maintain today’s rate of progress. Anthropic’s assessment suggests that better AI tools are becoming necessary just to maintain the current pace of AI progress.

Example one: GPT-5.6 cybersecurity gains shrink with more agents

OpenAI’s GPT-5.6 Sol results provide a useful example on cybersecurity. SEC-Bench Pro tests the generation of proof-of-concept demonstrations of software vulnerabilities. This is a particular security task, not a measure of coding ability in general. The reported peak scores are approximately 71.2 with one agent, 74.3 with four, and 76.4 with sixteen.

On this cybersecurity test, sixteen times as many agents buy about 5.2 score points. The first fourfold increase adds 3.1 points; the next adds only 2.1. There’s still plenty of room below a perfect score, but the gain on this test is shrinking, even with multiplication of the number of agents. This is sub-logarithmic scaling.

On SEC-Bench Pro, each expansion buys fewer score points, showing sub-log scaling. Source.

These agents could coordinate through OpenAI’s ultra setup. This is already a test of collaboration, not just a stack of independent guesses. The points are the best reported results across different budgets, rather than an equal-compute comparison.

Example two: A smarter model did more for OpenAI’s math results than more agents

When OpenAI announced its Navier–Stokes result, the 10,000 agents attracted enormous excitement. Interviewing Noam Brown, Dwarkesh Patel said he was “shocked by the scale of cognitive effort” concentrated into such a short run.

But how much did those extra agents contribute? Brown, an OpenAI researcher who helped develop o1 and now works on multi-agent systems, was emphatic:

“I wouldn’t even attribute 10% of the credit to multi-agent.”

He credits the smarter underlying model. OpenAI hadn’t run the single-agent experiment needed to isolate the swarm’s contribution. Brown’s estimate is a judgment, but it directly challenges the idea that the sheer number of agents was the decisive ingredient.

Example three: Formal math proof gains diminish sharply after ten Opus 5.5 agents

Formal mathematical work seems especially analogous to the conceptual breakthroughs AI research may need. It demands a chain of valid reasoning, sometimes with a new insight that makes the whole problem tractable. Formalizing existing mathematics isn’t the same as inventing a new learning algorithm, but it’s a useful place to look.

And here we find almost no extra gain when Opus 5.5 grows from ten parallel agents to a hundred. Anthropic’s system card tests teams on 49 weighted theorem statements, checked in Lean. Each team gets 24 hours. The first tenfold expansion adds 27 score points; the next adds just two.

On these math proofs, most of the gain arrives by ten agents. This is steeply sub-logarithmic scaling, and possibly saturation Source.

These agents could communicate, share a machine and Git repository, and organize their work. The hundred-agent team even developed a hierarchy with twelve sub-leads. It could also use ten times as much aggregate agent runtime as the ten-agent team over the same day. That isn’t a measured tenfold increase in tokens, but it gives the larger team much more room to work.

We still didn’t get over the hump. This won’t necessarily stay true, but the diminishing returns are incredibly steep here, possibly pointing toward a plateau.

What the formal math proof slowdown suggests

Opus 5.5 starts smarter than the older models and gets more speedup from parallel agents on ProgramBench. It clearly can put additional agents to useful work. So why do intelligence gains nearly stop on these math proofs? Two possibilities stand out:

The remaining work may be hard to parallelize. Later steps depend on understanding earlier results.

The remaining problems may need insights the model can’t find, however many copies we run. Intelligence may matter more than agent population or speed.

We haven’t isolated the cause. But a thousand copies don’t automatically supply the insight one copy is missing.

Example four: A smarter model beats ten times the agents on math proofs

Looking at the same math results, we see something striking. Ten Opus 5.5 agents get a higher score than a hundred Fable 5.1 or Opus 5.0 agents. Opus 5.5 starts much higher than the older models. It improves substantially with ten agents, then barely improves with a hundred. But at 10 it’s already better than Fable or Opus 5’s results with a hundred agents.

Opus 5.5 starts stronger, but its team gains flatten quickly. Source.

Ten Opus 5.5 agents beat a hundred Opus 5 agents on these math proofs. They beat a hundred Fable 5.1 agents too. The stronger model does more with one-tenth as many agents.

Ten Opus 5.5 agents are able to solve problems a hundred older-model agents can’t. Source.

We saw an analogous pattern with test-time compute in the previous post. OpenAI’s internal model starts far above Astra on the curated math problems below. Even its lowest displayed result beats Astra’s highest. Changing the model gives a much bigger boost than spending more compute on Astra over the tested range.

A stronger model makes the bigger difference than more test-time compute in solving previously-unsolved math problems. Source.

A rough extrapolation of Astra’s line suggests it would need about 300 times as much compute as the internal model to match the internal model’s lowest plotted score. And that extrapolation itself is an unknown. It’s quite possible that Astra would just plateau before ever matching the OpenAI internal model’s results with far less compute.

What about diversity?

Does adding different minds help? A study of model and persona diversity found substantial gains as diverse teams grew to eight agents. After that, the average across its tasks barely moved.

Diversity of AI models helps, but the gains flatten after eight to twelve agents. Source.

Other approaches, including X-MAS and Aitchison and colleagues’ forecasting study, find useful gains from combining different models. But none of these results demonstrates sustained improvement over many expansions of the team. The first study directly shows flattening; the others test smaller ranges.

It’s plausible that today’s models simply aren’t diverse enough. LLMs often produce strikingly similar ideas. Perhaps we need bigger changes in how these minds are trained. We’ll return to whether more deliberately different agents could change the curves.

The Best Examples of Multi Agents Improving Collective Intelligence

All of the above examples show sub-logarithmic scaling, roughly the hardest kind. That said, some experiments show slightly easier diminishing returns, indicating roughly the same gain from each multiplication of parallel work. That’s logarithmic scaling. It’s better, but still steep diminishing returns. If that’s how useful research grows while each further AI improvement gets harder to find, even billions of agents don’t make a self-sustaining RSI loop look plausible.

In Ryan Greenblatt’s ARC experiment, a model generated and checked many candidate programs. More samples kept improving its score, by roughly the same amount per doubling. This is parallel search rather than a communicating research team, but it gives us a useful comparison.

Parallel search keeps improving, with logarithmic returns. Source.

A molecular-exploration study found a similar logarithmic pattern as it grew from one to 128 agents. Multiple agents explored chemical targets more effectively than one agent given the same total evaluation budget. Each agent learned as it searched.

More AI agents explore more chemical targets, with logarithmic diminishing returns. Source.

There are encouraging results from communicating teams too, including Argus on web research and Microsoft and Berkeley’s work on ARC-AGI-3. The latter finds unusually strong gains from one to five agents. Those results deserve attention; we still need to see what happens with much larger teams.

Logarithmic gains are better than the sub-logarithmic curves we saw earlier. Each tenfold expansion keeps buying the same extra increment, instead of less each time. But both shapes rise drastically more slowly than the research gains assumed in hypothetical models of self-sustaining or runaway RSI. Let’s put them side by side.

Which problems should scale best

Some work can be split into independent searches or experiments. Other work needs the previous result before it can proceed. And some problems may need a higher level of intelligence or a conceptual breakthrough before we even know how to divide them. Here’s a very tentative and relatively subjective stab at which problems will scale best and worse with multiple agents. The chart runs from work that seems easiest to parallelize at the top to the hardest at the bottom. I’m sure others have many thoughts here, and that this can be much improved upon.

From work that seems easiest to parallelize to the hardest.

What does this mean for runaway RSI

Will a million or a billion parallel AIs get us to self-sustaining or runaway RSI? Let’s do the math.

Hypothetical software-RSI models assume that a growing number of AI researchers produces a much larger stream of discoveries. Those discoveries make AI cheaper or more productive, allowing the next round of research to grow again.

There’s a strong case for gains in speed and for searching more of a well-defined space. Even speed gains, as we’ve seen, can taper. Intelligence is much harder to boost through parallel agents. In several experiments, peak problem-solving performance (what I’m referring to here as intelligence) rises dramatically more slowly than speed.

Here’s what three software-RSI models assume:

Chan and colleagues (2026): doubling effective research labor produces about 2.6 times the rate of software improvement, holding the software level fixed. I find that implausible as a general assumption about adding copies of today’s AI models. The clearest large-team intelligence experiments look nothing like that.

Eth and Davidson (2025): their illustrative team produces twice as many papers from twice as many researchers. The accelerating example assumes that the number of researchers grows faster than the work required for each further improvement.

Tom Davidson and Tom Houlden (2025): their central exponent for parallel research effort is 0.6. Ten times the labor gives about four times the research pace. They cut that exponent in half, to 0.3, when experimental compute is held fixed and becomes a bottleneck. The benchmark reports above don’t appear to have been limited by experimental compute, so we’re using 0.6 here.

The top half of the next chart shows research-output relationships used in hypothetical RSI models where AI self-improvement can sustain itself or accelerate. The bottom half shows real-world intelligence shapes from the benchmark experiments. Those experiments show dramatically less gain from running parallel agents than the hypothetical RSI models assume.

Hypothetical research curves rise much faster than these benchmark shapes.

Let’s put these onto a single chart. We give every curve the same gain from its first expansion from one to ten agents or researchers, so we can compare their shapes.

The same starting gain, very different later returns.

Now let’s zoom in on the gap. The real-world shapes sit in a very different zone from the gains assumed in those RSI models, even when we let the math-proof curve keep rising slowly.

Zooming in makes the gap clearer. Hypothetical RSI models show much faster gains from adding more agents than empirical models based on real-world experiments.

Today’s hypothetical RSI models assume gains from additional AI researchers that we haven’t seen in these real-world experiments. Removing humans from the loop doesn’t remove shared blind spots or sequential dependencies. Adding more copies of today’s models hasn’t shown the sustained intelligence gains that may be needed for repeated conceptual breakthroughs.

Will agents get better at scaling?

I do expect better AI models to get better at spreading work across agents: dividing the problem, managing collaborators, and combining what they find. We’ve already seen progress in speed on program reconstruction. We haven’t seen a clear trend toward better multi-agent intelligence scaling in the math results. More on that in a moment.

First, an encouraging analogy: single-agent thinking-time scaling. Toby Ord’s analysis suggests that humans get more relative benefit from additional time than these AIs get from additional reasoning tokens. That may surprise people. But newer models are getting closer to the human reference on these cybersecurity tasks.

AI agents can get stuck in cognitive ruts, repeating or refining an approach that isn’t getting them far. Newer models seem to be improving at how they use extended thinking.

Newer models make better use of extra thinking tokens on these cyber tasks. Source.

This isn’t multi-agent scaling, but it’s encouraging. AIs were poor at turning more thinking into harder solved tasks, and that scaling parameter improved. They might eventually pass the human reference. That’s part of why I don’t see runaway RSI as imminent, yet think it could become more plausible as these capabilities change.

We’ve also seen new findings on helping agents escape their ruts and use more thinking time productively. The new Fork-and-Flush study periodically creates parallel branches that keep the accumulated work but start with fresh chat contexts, then continues from the best branch. It improved results across thirteen research and engineering tasks at the same compute budget. That’s a concrete research direction for using additional work more effectively.

So, have we seen a similar improvement in multi-agent scaling? For speed, sometimes. For intelligence, there’s no clear direction, so far, at least on formal math proofs. Opus 5.5 gets more speedup for a smaller cost increase on program reconstruction. But the score gains in the chart below are noisy, and newer models don’t consistently get a larger boost from additional agents.

Newer models don’t get consistently large gains from larger math-proof teams. Source.

That doesn’t mean better multi-agent scaling won’t happen. I expect the labs to work hard on coordination. Improving diversity of thought is a more open research question, but I’m sure it’s at least somewhat tractable given sufficient effort. The structure of the problem may impose the hardest limits: some discoveries have to happen before the next useful step becomes possible. We don’t yet know how much each of these constraints contributes.

Summarizing what we’ve seen, here’s an update on the conceptual limits to multi-agent scaling.

What we’ve learned so far, and where the hardest limits might be.

Can we build minds that think more differently?

Frontier models share many of the same broad training ingredients: overlapping public data, related algorithms, and similar reasoning and coding exercises. So how different are these minds really?

The Artificial Hivemind study found striking similarities in models’ open-ended responses. A creativity study involving 467 people likewise found that adding AI responses eventually contributed less diversity than adding human responses. These experiments suggest that AI agents can bring less diversity of thought than an equal number of humans on these tasks, which may limit how much we get from parallel agents.

Can we make agents more different from each other? Very likely. Here are four ideas. I’m sure there are others:

Different training experiences: vary data mixtures, curricula, reinforcement-learning environments, and reward objectives to develop strong agents that favor different useful approaches.

Complementary specializations: combine agents trained for conjecturing, finding lemmas, searching for counterexamples, and checking proofs. Test whether their strengths combine into results no member reliably achieves alone.

Independent proposals before discussion: let agents develop approaches before seeing one another’s answers, then exchange evidence and criticism. Test whether this reduces early convergence on the same mistake.

Teams chosen for complementary errors: select members whose successes and failures overlap less, rather than relying only on individual benchmark rank. Measure whether the team finds more verified solutions as it grows.

How to measure agentic scaling improvements

As we discussed in the previous post, we’re going off of scant data here. This model could be wrong. Or newer AI versions could get sufficiently better at doing parallel work that these diminishing returns would change. If that happens, we need to be aware of it.

How would we track this? I’d propose measuring three things from new AI models across a variety of problem times. Across each task type, and especially across real world AI research, measure three parameters and track how they change as newer models arrive:

G - Intelligence scaling: How much verified task performance does each multiplication of agents add? Measure local score gains, slopes, or exponents as appropriate. This is a task-specific proxy for collective intelligence, not a universal intelligence unit.

λ - Speed-cost scaling: How much faster does the team reach the same result, and how much more total compute or agent-time does it use? Track the speed exponent λ over successive team-size ranges.

N - Where gains weaken or saturate: At what team size do the score or speed curves bend, or approach a ceiling? Does the bend or onset of saturation move outward with newer models?

Track intelligence gains, the speed-cost tradeoff, and where gains weaken or saturate.

G and N answer two different questions. How much extra benchmark score does a defined team expansion buy, and how many agents can we add before those gains flatten? In the illustration below, G is the gain from one to ten agents. N is the agent count where the curve reaches its plateau. A real curve may have a gradual bend rather than a sharp plateau, or no plateau within the tested range.

One hypothetical benchmark curve. G is the added score from a specified expansion of the team; N is the agent count where gains flatten. The illustrated sharp plateau is not assumed for every real task.

Watch these parameters over time, especially G and N. Don’t stop at one to ten agents. Measure the gains from 1→10, 10→100, and 100→1,000 agents, separately, on each benchmark. A newer model might scale better over the first interval and worse over the next. We need to know whether the gains improve and whether the bend moves outward.

Three hypothetical trends for added benchmark score across model generations. Measure each expansion separately: 1 to 10, 10 to 100, and 100 to 1,000 agents. Different intervals can follow different paths.

We especially need to track these trends for actual internal AI research. Are newer models getting bigger gains from larger teams? Are those teams producing more verified discoveries and better successor models? That’s what would tell us whether massive parallel teams are getting closer to the conceptual breakthroughs RSI may need.

The Anthropic test and what would change my mind

Anthropic has given us a concrete test to watch. Its October employment report shows research agents logging about nine hours for every human work-hour by mid-July, and nineteen by late August. The company-wide ratio reached about six.

Research agent runtime rises to about nine hours per human work-hour by mid-July and nineteen by late August. These hours include longer operation as well as parallel agents. Source.

Let’s make a generous assumption: each agent-hour is as valuable as a human research-hour, with the number of human researchers unchanged. Agents can run throughout the 168-hour week. Anthropic’s runtime ratios already count those extra operating hours against a forty-hour human week.

On that assumption, July’s nine agent-hours plus one human hour amount to ten times the labor of humans alone. August’s total is twenty times. But Anthropic already had roughly one agent-hour per human hour in April. Compared with that observed baseline, total labor grew fivefold by July and tenfold by August.

What should that do to research? Treat the added labor as parallel research effort. Davidson and Houlden’s relationships imply about 2–4 times April’s research pace by August, depending on whether experimental compute becomes a bottleneck. Applying the labor exponent used by Alan Chan and colleagues, including Geoffrey Hinton and Yoshua Bengio, gives about 25 times. Log scaling gives about twice the pace. Our sub-log shapes give only about 1.1–1.8 times.

Now translate research pace into the Epoch Capabilities Index (ECI). Suppose software research supplies half of baseline capability improvement, while the compute-growth contribution stays unchanged. Then four times the software pace means about 2.5 times the overall ECI slope, not four times. We’re already discounting the software boost. Other obstacles could reduce it further.

Allow six months from research to public release. Here’s what those assumptions imply for January and February 2027. The top group uses hypothetical RSI research relationships; the bottom group extends experimental benchmark shapes.

What different RSI models predict we should see in the pace of AI improvement (using ECI slope), six months after an AI labor increase. Source.

Why do log scaling and Davidson-Houlden’s power 0.3 nearly tie? This is only a tenfold labor increase from April to August. At that scale, both calibrations give about twice the software pace. By 100 agents and then a billion, the hypothetical power law curve pulls far ahead of the experimentally-observed log and sub-log scaling rules.

Around late May, agents were already logging about three hours per human hour: twice April’s total labor on our assumption. With a six-month lag, that earlier signal would arrive around November.

These tests have many different assumptions, including equal value per hour and treating added runtime as parallel labor. I’m open to refining them on how others think about this. In measurement, we should judge the AI pace of improvement as the rate of at which ECI score goes up over successive releases, allowing for development and publication delays.

My expectation - and how I’ll update if I’m wrong

Personally, I don’t expect to see sustained acceleration of Anthropic’s AI progress (as measured by ECI) through the first half of 2027, and certainly not one as large as these simple extrapolations from the hypothetical models of Davidson and Houlden or Chan and colleagues suggest. If anything, I expect acceleration at the low end of the sub-log scaling bar at the bottom of the chart.

I could be wrong for many reasons. Parallel search of different AI training parameters, or use of AI to code new RL environments, or gains from AIs combining existing ideas in the literature could all make a bigger difference than I expect. Or, at the next increment of AI improvement, AIs could start making much more major advances in machine learning theory that allow us to bend the scaling laws and other diminishing returns that AI faces, or to add new axis of scaling, as reasoning models did.

On the other hand, while more parallel search and recombination of existing ideas seems highly likely, they may not bring huge advantages, and the AIs we’re making today might not find conceptual breakthroughs. Atop that, ideas probably really are getting harder to find in AI, and that may chew up some or all of the increased capacity of AI to work on improving itself.

We shall see. If we see the rate of AI improvement increase significantly, I’ll update my views in proportion to what the data show.

In summary, for now, multi-agent scaling shows incredibly steep diminishing returns to model intelligence, or the ability to reach conceptual breakthroughs. I expect those sorts of breakthroughs to be necessary for major AI advances and for self-sustaining RSI, let alone a runaway intelligence explosion. And at the same time, we’re all operating on far too little real world data. The most important thing we can do now and for the months and years ahead is to rigorously collect that data, and use it to update our forecasts of the future.

Why are we ruled by barbarians?

2026-10-17T22:34:00+00:00

“Something old and wild and blind at war with man and the works of man” — Poul Anderson

Years ago, I thought it was valuable for me to understand the depth of the human capacity for atrocity, so I made myself look at photos and videos of some of the worst things people had done to each other. So in 2015, I made myself watch the full, unedited video of ISIS burning a man alive in a cage. However horrifying that sounds, it was worse. The man stood in the cage and hung his head while masked ISIS fighters all around him sang grim chants. No scripted scene in Game of Thrones could have approached the raw, atavistic horror of that real-life execution. It was like something out of the ancient world — a human sacrifice on the steps of a temple built to gods long forgotten.

I thought of that video when I heard about the Trump administration’s plan to stage a public televised execution by firing squad. The condemned man, the 2009 Fort Hood shooter Nidal Malik Hasan, certainly deserves death, and a firing squad is a relatively humane way to kill a man. But to televise this event to the nation seemed like an act of barbarism — the kind of spectacle that was already a disgusting relic when Britain abolished it in 1868 (for the U.S., it took more than 60 more years).

I was not alone in this opinion. Except for MAGA Republican men — a slice of a slice of the right-wing base — every segment of American society deeply opposed the idea of shooting Hasan on live TV:

Seemingly swayed by the bad polling, Trump cancelled the idea, declaring that the execution wouldn’t be livestreamed. But the episode has revealed something important about the nature of the men who now rule America. Their values and their worldview are not the same as those of the average American. They are, in many ways, premodern and barbaric. Public executions actually aren’t a new idea on the right — Charlie Kirk was calling for them years ago. Now everyone knows that this wasn’t just empty rhetoric.

It is very important that we understand the New Right. Donald Trump himself is currently in command of the New Right, but not for much longer — he is old, older than Biden was at the same point in his presidency, and already visibly declining. What’s more, Trump is a singular figure — a charismatic but spectacularly corrupt mafia type with authoritarian tendencies and a worldview mashed together from early-1990s CNN and half-remembered stories of America’s pre-WW2 glory days.

There are many reasons Trump is amenable to the New Right, and became their beloved leader and standard-bearer, but he is fundamentally not of them, and when he is gone, it is that ideology, rather than another charismatic figure, which will flow into the vacuum his exit creates. Like MAGA now, it will rule through factional politics — the strategy by which a small minority becomes the majority-of-a-majority-of-a-majority. The New Right will lack Trump’s charisma, but it will also lack the baggage of his corruption and (possibly) the lingering stink of his policy blunders. So it is important that we understand this ideology — not simply because it rules us now, but because we may find a purer version of it ruling us sometime in the not-so-distant future.

The New Right, Western Civilization, and the Europe obsession

I already wrote a post about the New Right, in early 2025:

A slightly abridged version was later published in The Free Press.

In that post, I argued that the key to understanding the New Right was its conception of “Western Civilization” under siege:

To a large-ish number of Americans, [the New Right] told a coherent story about who they were, where they came from, and what commonalities tied them together. The story it told was about “Western civilization”.

When I say “Western civilization”, I don’t mean it in the Cold War sense of “the free world”. It doesn’t include Japan or Latin America, and it very much does include Russia. The Western Civilization that the New Right believes in is defined by three things: 1) European racial descent, 2) Christian heritage, and 3) veneration of the previous two things.

The third of these is absolutely the key. European descent and Christian heritage define “Western” for the New Right, but respect for those things defines “civilization”. If you created an all-white country, and it was full of blue-haired genderqueer atheist decolonial-leftist white people, that would be anathema to the New Right…Nor are the New Right typically pious or born-again Christians…[Christianity is] more of a source of shared heritage, and a boundary that demarcates historical Europe from the rest of the world. It’s less about Christ than about Christendom.

This concept explains why the New Right cares so much about Europe. It’s important to realize that this is new. The Old Old Right — the right wing of 1910 or 1935 — was explicitly isolationist toward Europe. Not only did they vigorously advocate against U.S. involvement in European wars, but they viewed the waves of South and East European immigration as fundamentally alien — a threat to the cultural and racial purity of America. Francis Walker called those immigrants “beaten men from beaten races”, and most of the restrictionism of the period — especially the famous immigration law of 1924 — was explicitly aimed at choking off the flow of people from places like Poland, Italy, and Greece.

You can see faint echoes of that old anti-European sentiment in Trump’s desire to withdraw from NATO. But overall, the New Right cares much more about Europe than any incarnation of the Right in American history ever has before. They view America and Europe as fundamentally unified — belonging to a single civilization with a single common heritage, ethos, and culture. Here’s Marco Rubio in February at the Munich Security Conference:

For the United States and Europe, we belong together. America was founded 250 years ago, but the roots began here on this continent long before. The man who settled and built the nation of my birth arrived on our shores carrying the memories and the traditions and the Christian faith of their ancestors as a sacred inheritance, an unbreakable link between the old world and the new…We are part of one civilization – Western civilization. We are bound to one another by the deepest bonds that nations could share, forged by centuries of shared history, Christian faith, culture, heritage, language, ancestry, and the sacrifices our forefathers made together for the common civilization to which we have fallen heir.

Rubio restated similar sentiments at a recent speech in Athens.

The Trump administration and the MAGA movement are obsessed with the idea that immigration — primarily from Muslim countries — is destroying the birthplace of Western Civilization, aided and abetted by leftist ideology among the European elite. This is from a Hegseth speech in June, commemorating the Normandy landings:

Sadly, today, different European beaches are stormed by different dangerous ideologies…Beaches in Spain, in Italy, in Greece, and Bulgaria. Boats and men arrive. When will European capitals do something about that invasion, or is it too late?

And this is from Rubio’s February speech:

[I]n a pursuit of a world without borders, we opened our doors to an unprecedented wave of mass migration that threatens the cohesion of our societies, the continuity of our culture, and the future of our people. We made these mistakes together, and now, together, we owe it to our people to face those facts and to move forward, to rebuild…Mass migration…was and continues to be a crisis which is transforming and destabilizing societies all across the West.

And here’s JD Vance back in 2025:

Of all the pressing challenges that the nations represented here face, I believe there is nothing more urgent than mass migration…How many times must we suffer these appalling setbacks before we change course and take our shared civilization in a new direction?…No voter on this continent went to the ballot box to open the floodgates to millions of unvetted immigrants.

The core belief of the New Right, in other words, is that Western Civilization is being invaded by barbarians — people from non-Western civilizations, who come to the West under the guise of immigration and then proceed to try to overthrow the West from within, with the aid of local leftists and progressive elites.

This worldview has been reinforced by a constant stream of news from Europe, which American rightists follow assiduously and comment on constantly. Islamist terror attacks in Europe are rarer than far-right extremist attacks, but cause far more deaths. Migrants from Muslim countries are responsible for a very disproportionate number of sexual assaults in countries like Sweden. And there are periodic riots in France, often initiated by students, leftists, and various unaffiliated angry folks, but disproportionately joined by poor people from immigrant backgrounds living in ghetto-like areas. The most recent round of such disturbances is currently gripping the nation and generating a bunch of apocalyptic videos:

Incidents like this have steadily increased support for the European right, which promises mass deportations and has explicitly positioned itself as the defender of Western civilization:

America’s New Right views itself as closely allied with the European right — perhaps even one and the same. Elon Musk does everything he can to support Germany’s AfD, and JD Vance has met with them. JD Vance campaigned (unsuccessfully) for Hungary’s Viktor Orban. Trump has often voiced support for France’s Le Pen, and so on. It’s a regular Western Civilization Internationale.

The European right has been ambivalent about this support; Trump’s extreme unpopularity in Europe means that his embrace is often detrimental to the European right’s electoral prospects. Nor is there much evidence that Americans link Europe’s immigration problems with the immigration issue in the United States — pro-immigrant attitudes have surged in America since Trump returned to power, and the President is underwater on the issue.

Regular Americans simply do not see Europe and America as a single, unified civilizational entity. They do not view riots on the streets of Paris as a dark portent of riots on the streets of Houston. They do not see their Guatemalan housecleaners as rapists or terrorists. They are not invested in the political fortunes of Alternative für Deutschland or Marine Le Pen or Restore Britain. These are fringe beliefs, core to the worldview of the New Right but utterly peripheral and irrelevant to the typical resident of Michigan or North Carolina.

And these fringe beliefs help explain why the New Right has embraced a savage, atavistic morality that’s deeply at odds with the decency of regular Americans.

Why the New Right embraces savagery

The New Right’s embrace of savagery did not begin with the proposed livestreaming of Nidal Malik Hasan’s execution. For many years, I have been watching commentators in this movement claim that the American right needs to become tougher, more vicious, and more brutal — and that conservatism has been hamstrung by excessive niceness and tolerance in the past. For example, here was Jesse Kelly in 2022:

You communists simply got way too comfortable with the ‘live and let live’ Right. You’re gonna have a tough time adjusting to The New Right. Your names and accusations mean nothing to us. You can’t use our values against us. We’re enemies. Game on,” posted Kelly on social media…I’m not a Republican. I’m not ‘better than that’. I’m a monster. I believe in hitting people who hit me.

In 2021, Donald Trump Jr. told a gathering of evangelical Christians:

We’ve turned the other cheek, and I understand, sort of, the biblical reference — I understand the mentality — but it’s gotten us nothing. OK? It’s gotten us nothing while we’ve ceded ground in every major institution in our country.

And if you watch rightists tweet, you constantly see them making declarations like this:

In his famous essay “The Flight 93 Election”, widely viewed as a founding manifesto of the New Right, Michael Anton says conservatives have been meeker than “fearful, beaten dogs”, and declares that they “bend over backward to play by the self-sabotaging rules the Left sets for them”. Less elegantly written screeds by Kurt Schlichter, Glenn Ellmers, and others all make the same argument.

American rightists have also begun to heap praise on historical rightist leaders whom they view as having achieved effectiveness through savagery. The two they typically cite are Spain’s Francisco Franco and Chile’s Augusto Pinochet. Recently, though, we’ve started to see another name creep back into rightist discourse:

One would think that the disastrous results Adolf Hitler produced for Germany, Western Civilization, and the European Right would make American rightists wary of carrying out apologetics on his behalf. But no:

Whether American rightists believe in Hitler’s ideology or want to emulate his policies is beside the point here. The real lesson is that by invoking Hitler, rightists can paint themselves as tough, savage, and ruthless.

A pattern should begin to emerge here. Leftists are famously obsessed with ideological purity; when Democrats lose, leftists always claim that it was because they were too moderate, or compromised their principles, or failed to energize the working class by promising real change, etc. Rightists, on the other hand, are obsessed with toughness. If Republicans lose, it was because they were too meek, too bound by liberal norms. Repeated failures, to rightists, are nothing but an infinite regress of boxing gloves that they were too afraid to take off.

This explains the proposed televised execution of Nidal Malik Hasan. Rightists view Hasan — a Palestinian American who served in the U.S. military but murdered his fellow servicemembers because he was angry about America killing Muslims in Afghanistan — as typical of the invading anti-Western barbarian horde. By executing him on live TV, they probably thought they could demonstrate that the West can be equally savage in its own defense.

It also explains Pete Hegseth’s constant performative toughness:

Physical toughness and traditional masculinity are not the same thing as brutality, obviously, but these antics demonstrate the same general obsession with showing that the Men of the West are no weaklings.

And of course, the rightist obsession with toughness explains part of the romance between Trump and Russia.

There are many reasons that Trump has chosen to shift America’s allegiance from Europe to Russia, of course. Given the President’s stupendous, unprecedented corruption, I would be very surprised if there were not some financial ties involved. But rightists also tend to see Putin’s Russia as the living instantiation of Western toughness — savage, unrestrained, unconcerned with humanitarian softness or the rules of war.

Republicans have been oohing and ahhing over Russia’s hypermasculine military recruitment ads for years, despite the fact that the Ukrainians inflict many more casualties on the Russians than vice versa. Push-ups and big biceps don’t win wars — nor do massacres and war crimes — but they comfort American rightists with the knowledge that somewhere out there are White, culturally Christian men who can match the hordes of the Middle East in savagery.

If you wonder why Trump is canceling Congressionally mandated sanctions on Russia, urging Europe to buy more Russian gas, ordering the military to cut off Starshield satellite communications for the Ukrainians defending their homes from Russian drones, and discussing various joint economic projects with Russia…well, this is the reason.

It’s probably not the only reason. But if it were purely a matter of Trump’s personal corruption, his team would not be so eager to go along with all this pro-Russia stuff, especially given how unpopular it is with the American people. The MAGA ideology views Russia as a civilizational ally against the true enemy — an EU they believe to be captured by leftists and bent on destroying Western Civilization through mass importation of anti-Western immigrants. And Russia’s savagery and attitude of toughness both on and off the battlefield provide America’s rightists with an ideal to aspire to.

This Russian propagandist basically summed up how the rightists see the world right now:

The rightists’ embrace of savagery, of course, is not just about aesthetics, or about psyching themselves up to do the horrible things that deep down they suspect must be done. It’s also a way to give them confidence in the face of their own unpopularity. Rightist figures are deeply unpopular in America. Their main policy idea — remaking American policy to ethnically reengineer the American populace — has very little purchase. Everyone in America still thinks Hitler was the worst. The obsession with Europe, and with their idea of Western Civilization, is a niche, alien thing. And they’re losing the young generation of men:

The only way for rightists to sustain the illusion that they would have the upper hand in a post-democratic America — that they would triumph in the climactic civil war that suffuses their perfervid daydreams — is to imagine themselves as a diamond-hard vanguard, a small fist of ruthless men who can defeat any number of effeminate, emasculated enemies, the “three percent” who are willing to pick up a gun and kill.

That fantasy is a dangerous one. You can end up doing some very stupid things if you think you’re the only wolf in a world full of sheep. A successful movement needs broad popular appeal, not just the willingness to out-barbarize the enemy. But the New Right appears set on pursuing the strategy it has laid out for itself, and all we the rest of us can do is to understand how they think, and prepare for their next mistake.

AI vs. the railroads

2026-10-15T00:47:39+00:00

A little over a year ago, a bunch of people were worried that the AI data center boom would result in an epic bust, much like the railroad bust in 1873. I wrote about the possibility at the time:

The basic idea is that even if AI is an incredibly useful technology and creates enormous economic growth in the future, the people building the infrastructure (data centers) might borrow so much money that they can’t pay it back in time. Last summer, the AI industry’s revenue was growing, but not nearly as fast as the amount of money being spent. But the “hyperscalers” building the data centers were mostly spending their own cash instead of borrowing money, so the danger wasn’t imminent. The worry was mostly about the future — if spending continued to increase and revenue didn’t accelerate, there could be an 1873-style bust.

Well, now it’s the future, and two things happened. Spending did accelerate, and hyperscalers started having to borrow money to keep increasing their spending. But revenue accelerated even more. The release of Claude Code, Codex, and other agentic coding tools proved to be the “killer app” of AI, and suddenly Anthropic and (to a lesser extent) OpenAI were rolling in cash:

So like in summer 2025, we aren’t obviously heading for a big AI bust in the near future. But the stakes are higher now — because companies are borrowing a lot more money now, any slowdown in the torrid pace of AI revenue growth could cause a blowup.

The amounts of money being thrown around here are just staggering. A year and a half ago, the AI boom was still significantly smaller than the railroad boom of the late 1800s, as a percent of GDP. Now, there’s an active argument as to which one is bigger.

Which is bigger, AI or the railroads? (a pointless but fun debate)

This debate doesn’t really matter, because the basic message is just “Wow, it’s a lot of money.” But it’s fun to look at the numbers people are throwing around. Stijn Van Nieuwerburgh gives figures of 3.63% of GDP for AI and 2.24% of GDP for railroads:

But those percentages are over different amounts of time — the railroad boom lasted a very long time. According to some data sources, railroad spending peaked at around 5.4% of GDP:

And 19th century data are notoriously patchy and incomplete. Some sources claim much higher numbers for the railroad buildout:

Nor do people agree on the size of the AI investment binge. Some sources yield numbers that are much lower than Van Nieuwerburgh’s:

As far as I can tell, the main difference here is assumptions about costs. The banks and ratings agencies are using announced dollar spending; Van Nieuwerburgh is using announced physical construction projects and then making assumptions about how much those projects will cost. The biggest reason he gets a higher number is that he assumes data centers are going to cost a lot more than the hyperscalers think.

Meanwhile, Dylan Patel claims over 6% for the AI boom already, once you include all types of AI-related infrastructure — fab equipment, chips, energy, and so on.

But whatever the correct numbers are, the basic point remains the same: AI is now one of the two biggest infrastructure booms in American history, by a large margin.

So it’s understandable that people are worried that the two booms could follow similar paths. For the railroads, that path included a big crash in 1873, which caused a huge decade-long depression in the U.S. economy. A lot of people are watching for something similar to happen with data centers, and the chorus of skeptical voices is beginning to swell again.

Hyperscalers are starting to borrow…a lot

OK anyway, on to the real point of this post. The flood of revenue that accompanied the rise of coding agents quieted the talk of an AI bubble for a while, but now it’s back. The first reason is the rise of AI-related debt. In the early years, the hyperscalers — data center builders like Amazon, Google, Meta, and Oracle — mostly paid in cash for the data centers. These companies get tons of cash from their software and cloud businesses, and they ended up spending most of it on AI infrastructure. But now the buildout has gotten so big that they’re having to borrow a bunch of money to keep ramping up spending:

For all of the hyperscalers except Microsoft, free cash flow either went negative in recent months or is expected to go negative late this year. This has been a giant boon for semiconductor companies — Nvidia, TSMC, and the rest:

All this debt creates a financial danger that didn’t exist before now. When they paid in cash, hyperscalers were safe — if AI revenue disappointed, their stock prices would go down, but their businesses would just keep going and they’d eventually recover their profitability. But the more hyperscalers borrow, the more revenue they’ll need just to stay solvent; if the revenue doesn’t arrive, they might have to default on their debt.

A Nobel for New Keynesian macro, at last!

2026-10-12T21:28:43+00:00

I have a Google Doc with a long list of posts I’ve been meaning to write. Far down the list was a line that read “New Keynesian macroeconomics deserves a Nobel”. Somehow I never got around to writing that post, and now I never will. Because today, it happened! This year’s Econ Nobel went to Bob Hall and Mike Woodford, two of the theory’s most important creators.

First, let me say that this prize is long overdue. In 2009, I asked my PhD advisor, Miles Kimball, who was the greatest living macroeconomist. Without hesitation, he answered “Bob Hall”, and said he would be very surprised if Hall didn’t win the Nobel at some point. Hall’s work is breathtakingly broad; if you want good summaries, I recommend Alex Tabarrok and Tyler Cowen (as usual). As for Michael Woodford, everyone I know assumed that if New Keynesian macro ever did get the prize, Woodford would be included, as he literally wrote the book on it.

But anyway, let’s talk about New Keynesian economics — what it is, how it came about, why it’s important, and why it didn’t get a Nobel until now. Let’s start with a brief history of modern macroeconomics.

Back in the 1960s and early 1970s, most economists tried to understand the macroeconomy by looking at relationships between macroeconomic aggregates. For example, “If income goes up by 10%, how much will spending go up?”. Or “If inflation goes up by 3 percentage points, how much will unemployment go down?”.

The problem with that is that correlation isn’t causation. Suppose you see that when inflation rises, unemployment goes down. That correlation is called a Phillips Curve. So you think “OK, if we raise inflation, more people will get jobs”. But you try that, and it doesn’t work — you push inflation up (by printing more money), but unemployment stays high. The correlation you saw before has now broken down. Now you’ve got two problems instead of one!

This is a very simplified version of what a lot of economists think happened in the 1970s. Whether that’s what really happened is a story for another day. But the stagflation of the 1970s prompted the econ profession as a whole to rethink how it modeled the macroeconomy. A lot of researchers asked: What parts of the macroeconomy are invariant to policy? What are the solid foundations we can stand on, so that when we try to make some sort of change, we don’t cause our own theories to crumble?

The answer most economists settled on was called “microfoundations”. They reasoned that the basic drivers of individual and corporate economic behavior shouldn’t change when macro policies like interest rates and government spending change. Most importantly, those drivers include:

Technology

Consumer preferences

The process by which people and businesses form expectations of the future

Do all three of these really stay the same when economic policy changes? Probably not completely. But it gave macroeconomists a place to start. So in the late 1970s and early 1980s, economists started to build macro models based on theories of individual and corporate behavior. These were called “microfounded” macro models.

The last of those three “microfoundations” — expectations — was particularly thorny to think about. Almost any sort of expectation-formation process that you write down allows policymakers to trick people over and over, by surprising them systematically. Economists thought that eventually, people should wise up to the government’s tricks. So they came up with the idea that consumers and businesses should, on average, be as smart as the economists making the models. If economists write down a model of how the world works, that model should be the average person’s mental model in the model. Basically, people generally know how the economy works, and they plan accordingly.

This idea is called “rational expectations”.

In 1978, Bob Hall wrote a very influential paper that convinced a lot of economists that this was the right track. Hall wanted to test a simple idea, originally developed by Milton Friedman: If people plan out their consumption in advance, then changes in the level of consumption should be completely random. Nothing from last quarter should be able to predict the change in consumption this quarter — because if something last quarter mattered for this quarter’s consumption, people would already have changed their consumption last quarter! Hall found that, sure enough, a bunch of variables that seemed like they should be important for consumption — especially last quarter’s income — didn’t actually help predict changes in consumption.

For a more detailed explanation, I recommend this post by John Cochrane:

Brian Albrecht has a good (brief) explanation as well.

At first blush, this might not seem like an important result. After all, absence of evidence isn’t evidence of absence — there might be other variables that do predict changes in consumption, that Hall simply forgot to test. And in fact, Hall did find one variable that predicted future changes in consumption: stock prices. So even in the formal sense, the basic theory didn’t quite hold. Nor did Hall’s result rule out the traditional Keynesian idea of consumption — if government spending surprises people, it can move consumption immediately, without creating any sort of predictable future change.

But it was an important result, for two reasons. First, if people plan ahead when making their economic decisions, it should be possible for policymakers to influence today’s behavior by making promises about what they’ll do in the future. Second, if it’s impossible for economists to predict changes in consumption, that implies that consumers are as good at planning their future consumption as economists are.

If that’s true, it’s an important lesson for policy: If you want to make people consume more, or if you want to make businesses hire more, etc., you have to do it in a way such that people generally understand what you’re doing.

So macroeconomists set out to make models of the economy in which people understand how the economy works. At first, this effort didn’t go well. For example, one very popular early idea — which eventually won a Nobel — was called “real business cycle” (RBC) theory. This was the idea that recessions and booms are driven purely by changes in the rate of technological progress — if engineers happen to invent less useful stuff this quarter, you get a recession.

No one denies that technological inventions can be very important for economic growth — just look at how the AI boom is powering the U.S. economy right now. But the idea that tech progress drives the business cycle in a simple, mechanical way just doesn’t fit the evidence. In particular, productivity booms don’t seem to line up with economic booms, as RBC theory would predict.

Even more importantly, RBC models predict that monetary policy — raising and lowering interest rates — shouldn’t affect the real economy. Ed Prescott, the creator of RBC theory, famously declared that monetary policy “is as effective in bringing prosperity as rain dancing is in bringing rain.” But our best evidence says that Prescott was wrong — monetary policy almost certainly does affect the real economy.

So something’s going on there. Macroeconomists were missing something big — they didn’t understand how or why monetary policy boosts growth and employment. During the 1990s, a bunch of economists started coming up with possible reasons why having the Fed buy bonds could result in more people getting jobs.

These ideas generally fell under the name of “New Keynesianism”. As John Cochrane and others have pointed out, this was something of a misnomer. Whereas John Maynard Keynes and his successors thought a lot about fiscal policy — i.e. government spending and taxes — the New Keynesians of the 1990s and 2000s focused almost entirely on monetary policy. New Keynesian macro initially owed a lot more to Milton Friedman than it did to Keynes — everything was about the Fed, and the idea of fiscal stimulus initially fell by the wayside.

Anyway, in the 90s and 2000s there were plenty of ideas about how monetary policy could affect the real economy. This is probably one big reason why New Keynesian economics was so slow to win a Nobel — it has so many creators that it’s not even clear who deserves the prize. Greg Mankiw, David Romer, Stanley Fischer, George Akerlof, Janet Yellen, Olivier Blanchard, Nobuhiro Kiyotaki, Jordi Gali, Mark Gertler, John Taylor, Guillermo Calvo, and a bunch of others contributed pieces to the edifice.

Most of these contributions were ideas about why prices are slow to adjust to macroeconomic fluctuations. Basically, if you have something that decreases aggregate demand — say the Fed raises interest rates like in 1980, or there’s a banking panic after a housing bubble bursts like in 2008 — there’s a question of why companies don’t just cut prices and maintain the same level of output. If there’s something causing prices to fall too slowly, that can force companies to produce less instead. A lot of New Keynesian theory is about why companies have difficulty cutting prices.

Bob Hall made very important contributions here too! The basic New Keynesian model explains how interest rates can affect GDP, but it doesn’t explain why that affects employment. After all, if growth slows down, it’s not immediately obvious why a bunch of people should be thrown out of their jobs. In a 2005 paper, Hall made a model explaining how wages might be slow to adjust to changes in demand — meaning that when there’s a downturn, companies can’t just cut wages, but instead have to lay off workers.

Anyway, New Keynesian economics wasn’t the brainchild of a few famous founders — it was a riot of new ideas. It needed someone to come synthesize all of those ideas into something simple that central bankers could use to set monetary policy in real life. That man was Mike Woodford.

In a paper with the late Julio Rotemberg, Woodford laid out a simple New Keynesian model. On the question of why companies have difficulty cutting prices in a recession, he punted — Woodford simply assumed that for whatever reason, only some companies are able to change their prices each quarter. (This idea is called “Calvo pricing”, after its creator, Guillermo Calvo.) The rest of the model is pretty simple. Woodford uses the same basic consumption-smoothing model from Hall’s 1978 paper, in which people’s decision of how much to consume and how much to save depends on the interest rate. He assumes that the central bank (i.e., the Fed) sets interest rates according to a simple rule that balances out the risks of inflation and unemployment. There are a few more assumptions in there, but those are the basics.

This simple model provides a guide for monetary policy. It explains how the central bank can cause a recession in the short term — like Paul Volcker did in the early 1980s, when he raised interest rates to fight inflation — but can’t raise employment forever by keeping rates permanently low. And it explains how the central bank can respond to a drop in private demand — say, from a financial crisis like the one in 2008 — by cutting interest rates to stabilize the economy in the short term.

In other words, Woodford’s model — which he later expanded into a famous book — is a grand synthesis of the general consensus that emerged from the intellectual ferment of macroeconomics from the 1970s through the 2000s. We knew that aggregate demand was a real thing. We knew that the economy had trouble adjusting to negative demand shocks, resulting in a lot of disruption — reduced output, people getting thrown out of work, and so on. And we knew that monetary policy could boost the economy in the short term, but that its effect probably faded in the long term. Woodford gave us the simplest plausible model to describe all of that, and to offer central banks a badly-needed road map for stabilizing the economy.

Because it was so simple and so applicable, Woodford’s New Keynesian model became the standard tool of central banks around the world. No sooner was his synthesis complete, however, than it ran into a huge problem. In 2008 and 2009, the U.S. and European economies were hit by such a huge negative demand shock that cutting interest rates all the way to zero failed to put things right. Something else was needed.

There were three basic ideas for what that “something else” should be. The first is contained in Woodford’s model, and draws on the basic idea in Hall’s papers. It’s the idea of “forward guidance”. If the central bank can promise to leave interest rates low for a long time even after the recession is over, that can prompt consumers to spend more today — thus hastening the end of the recession. The second is quantitative easing — having the Fed buy long-term bonds instead of just short-term ones, in order to push down long-term interest rates. QE isn’t in Woodford’s model, but he eventually came around to supporting it.

The third idea was fiscal stimulus. If interest rates go all the way to zero, then government spending can become a more effective way of raising demand (as long as the central bank doesn’t cancel out the stimulus by raising rates again). When the Great Recession hit, New Keynesian economists scrambled to put Keynes back into their theories, by adding a role for fiscal stimulus.

Hall wrote a paper in 2009 explaining how stimulus could be effective:

[T]he multiplier linking government purchases to GDP may be in the range of 0.7 to 1.0…but higher values are not ruled out. New Keynesian macro models have multipliers in that range as well…Multipliers are higher—perhaps around 1.7—when the nominal interest rate is at its lower bound of zero, as it was during 2009.

In that paper, Hall explains how fiscal stimulus can be effective even in light of his famous result from 1978. Even if there is no simple “consumption function” linking stimulus directly to spending, there are plenty of ways that stimulus can affect the behavior of companies and consumers, especially when interest rates are at zero.

And Woodford wrote a paper in 2011 showing how stimulus works in his basic model:

This paper explains the key factors that determine the output multiplier of government purchases in New Keynesian models, through a series of simple examples that can be solved analytically. Sticky prices or wages allow for larger multipliers than in a neoclassical model, though the size of the multiplier depends crucially on the monetary policy response. A multiplier well in excess of one is possible when monetary policy is constrained by the zero lower bound, and in this case welfare increases if government purchases expand to partially fill the output gap that arises from the inability to lower interest rates.

So in the end, the New Keynesians went beyond the ideas of Milton Friedman, and brought some Keynes back into their theories.

But New Keynesianism’s biggest success arguably came not in the Great Recession, but in the post-pandemic inflation of 2021-23. Inflation spiked to 9% in 2022, but then came down to around 3% in early 2023 after the Fed raised interest rates. But what’s remarkable is that unlike in the early 1980s, this happened without a big rise in unemployment!

It’s not clear exactly why this happened, of course, but it sure looks like a big victory for the kind of New Keynesian macro that Bob Hall and Mike Woodford pioneered. The most likely explanation for this “immaculate disinflation” is that as soon as businesses realized that the Fed would do whatever it took to beat inflation, they stopped raising prices — and inflation simply ended.

It didn’t require demand destruction like Volcker caused in the early 1980s. It didn’t require legions of down-and-out unemployed blue-collar workers cutting back on essential purchases to finally force companies to stop raising their prices. All it required was a clear Fed commitment, and forward-looking expectations — exactly the mechanisms at the core of Bob Hall and Mike Woodford’s models.

It was this episode, even more than 2008-9, that truly represented New Keynesianism’s acid test. And it seems to have passed with flying colors. There was always a question mark hanging over the theory — it was always unclear whether it could shape the future instead of simply telling a story about the past. Since 2022-23, the answer to that question looks like a much more unambiguous “yes”.

I wouldn’t be surprised if that had something to do with this long-overdue Nobel.

Why Donald Trump won in 2024

2026-10-10T23:31:44+00:00

Everyone except the die-hard faithful now realizes that Donald Trump’s second presidency is a colossal failure. The president is doing his best to gin up symbolic, minor controversies in order to activate negative polarization and distract from his spectacular failures on inflation, the Iran War, corruption, racial harmony, and so on. These include demanding that artificial intelligence be renamed “artificial superintelligence”, bellowing that he should have won the Nobel Peace Prize, and conducting an execution on live TV. The goal is to get Republicans fighting Democrats over these controversies, in order to trick everyone into focusing on how much they hate the other team instead of how much they hate Trump.

The trick will probably fail, just as Trump’s promise to give every American $5000 for voting Republican backfired. All of these little tempests-in-teapots simply reinforce what a cruel, selfish, petty buffoon Trump is. And Trump’s approval rating just drifts lower and lower:

Trump will probably lose the midterms, and become a lame duck president, spamming increasingly grandiose, unworkable executive orders while battling a stream of Congressional investigations.

That means the window is closing on my ability to force Democrats to reckon with their failures during the Biden years. If and when Trump loses a national election, there will be far less pressure for soul-searching on the left. And once the GOP is no longer in control of all three branches of government, the Democrats will be able to influence actual policy instead of sitting around rebuilding their brand.

So before that happens, I want to summon the last fading scraps of the self-reflection that Democrats and progressives were briefly forced into after Trump thumped Kamala Harris two years ago. When Trump won, I wrote four blog posts trying to explain where the Democrats had gone wrong:

So today I’m going to repost all four parts of my election post-mortem, so that the progressives and Democrats among my readers can take one final moment to reflect on how not to enable the rise of authoritarian rightist leaders in future election cycles.

Identity politics isn’t working

I’ve made no secret of the fact that I think Trump is a terrible choice for America, but there’s no denying that the American people have made their choice. Trump is currently projected to win every single swing state. He outperformed his 2020 performance in every single county in the country. He’s projected to win a solid majority of the popular vote — the first time a Republican has done so since 2004. I still don’t like the Electoral College as an institution, but you can’t blame it for this result like you could for 2016. In fact, Trump’s victory was part of a more general Republican landslide — the GOP took the Senate and improved its majority in the House. A map of political shifts between 2020 and 2024 shows almost solid red.

I suppose a small silver lining here is that we won’t have to spend the next two months enduring waves of “Stop the Steal” election denials, lawsuits, and threats to repeat January 6th. Democracy delivered Trump a victory, and so Trump has little incentive to attack democracy.

In fact, that silver lining might extend to the long term too. Trump and the GOP won because nonwhite voters and working-class voters — traditional Democratic constituencies — shifted his way. Dems still won Hispanic, Black, and Asian voters overall, but their margins fell significantly from previous elections, while actually improving with White voters:

In fact, exit polls show Trump winning a solid majority of Hispanic men!

Update: Over at Matt Yglesias’ blog, Jed Kolko shows that heavily Hispanic areas swung harder toward Trump than any other parts of the country:

Ronald Reagan’s prophecy that Latinos would become Republicans is coming to pass.

Income has realigned as well. Democrats are now the party of the educated upper middle and wealthy class, but have lost their appeal to the middle and working classes:

This largely destroys the narrative that nonwhite immigration will demographically drown White Americans under a tide of imported minority votes. In fact, it vindicates my guest poster Wally Nowinski, who observed that high-turnout elections like this one now help Republicans, because marginal voters — working-class Americans and Hispanic Americans — now lean to the right:

At some point, Republicans are going to realize this, and hopefully become less anxious about America’s racial future. Hopefully they will also realize that any attempt to make voting harder actually hurts them in the future, because the impact would fall disproportionately on their own base.

I expect these realizations to take some time to sink in, but ultimately their effect could be very salutary for American democracy. In the days of Nixon and Reagan, Republicans never worried about democracy itself — it was obviously on their side. Even in 2004, Republicans could tell themselves that we lived in a “center-right nation”. Now, Trump’s victory, powered by lower-income Americans and Hispanic men, will hopefully convince the GOP that democracy is still a game they can win.

A democracy needs both major parties to believe they can win via persuasion rather than by organized resistance, and 2024 showed that Republicans can still do that — even if the method of persuasion is one that I personally wish hadn’t worked so well.

If you want an optimistic take, there it is.

So what other important lessons can we learn from this disappointing election result? I didn’t title this post “Why Trump won”, because I don’t think I really have a convincing answer — I’m not a political scientist, or even a political pundit. In fact, there are probably a bunch of reasons, not just one.

Nor am I going to be one of those pundits who tells you that if only Democrats had listened to my policy priorities, things would have gone very differently. In fact, few Americans care about my own policy priorities. My top issues are industrial policy and pro-democratic foreign policy. I definitely don’t think emphasizing those more, or doing more of those things, would have propelled Harris to victory. (I don’t think they would have hurt; I think voters just don’t care.)

But I do think there are some important lessons that we can already learn from this election, especially at the intersection of economics and politics. In particular, I see three big ones:

Identity politics — viewing racial groups as homogeneous “communities” to be targeted with appeals to collective grievances — is not an effective way of winning over Hispanic (or, probably, Asian) voters.

People care about inflation more than about unemployment.

The educated professional class has become dangerously out of touch with the rest of the country.

Instead of writing all of these in one post, I’m going to split it into three, because I think the lessons are each very important. Since I’ve already shown the racial and class breakdown of the exit polls, I’ll talk about how identity politics is failing Democrats. Whether or not you think identity politics drove voters into Trump’s arms, it certainly hasn’t done an effective job of retaining Hispanic voters. The whole approach needs a deep rethink.

Americans are individuals, not extensions of racial “communities”

In the 2010s, there was a big debate over identity politics, with scholars like Francis Fukuyama warning that treating the American electorate as a balkanized set of racial collectives was undermining democracy. Stacey Abrams and other progressive writers pushed back against Fukuyama, arguing that all politics is identity politics, and so Democrats should continue focusing on racial appeals.

Abrams et al. were right in the facile, definitional sense that people’s personal identification with larger groups, movements, and ideas — rather than their own personal economic interests — drives much of their political behavior. After all, if human beings were rational calculating machines, they wouldn’t even vote, given the incredibly low probability that any election will be decided by a single vote.

But the 2024 election — even more than the 2020 election — makes it clear that most Hispanic and Asian voters don’t want to be appealed to based on the racial group of which they’re supposedly a part. For example, lenient policies toward asylum seekers from Latin America don’t seem to have won over many Hispanics:

In Hidalgo County, in South Texas — a Mexican American stronghold that Biden won with 58% of the vote in 2020 — Vice President Harris and Trump were virtually tied early Wednesday, with Harris leading by less than half a percentage point…In Cameron County, at the southern tip of Texas’ Rio Grande Valley, Trump led Harris, 52-47, according to early results. Biden won that county with 56% of the vote four years ago.

The whole notion that lenient policies on illegal and quasi-legal immigration would win inroads with Latinos was based on the idea that Hispanic Americans identify strongly with others of their same race. Political scientists call this the theory of “linked fate”.

Democrats clearly buy into this theory when they refer to racial groups as “communities”. This term has become ubiquitous in Democratic messaging — for example, Harris talked about her “economic agenda for the Latino community”.

But races are not communities. Our country is not like a segregated 1930-era city, where Black people live in a ghetto, Chinese people live in Chinatown, Italian people live in Little Italy, and so on. Hispanic and Asian Americans might share a sense of “linked fate” with others of their race, but it’s only one of many things they care about, and perhaps not even a major one.

The big exception, of course, is Black Americans, who continue to vote mostly like a racial bloc, despite Trump’s modest inroads with Black men. This has clearly influenced the Democrats’ view of how racial politics operates in America, and they’ve tried to apply the same mental model to Hispanics and Asians. This is explicit in California, where you hear the term “Black and brown” thrown around a lot.

But it isn’t working. In fact, it’s backfiring. A recent paper by d’Urso and Roman found that use of the term “Latinx” actually made Hispanics vote for the Republicans in nontrivial numbers!

Group labels are political, but little theory and evidence explains how group labels shape politician evaluations…Latinos’ relationship with “Latinx,” a gender-inclusive label, is a theoretical test case. Using several datasets, we find: Latinos are less likely to support politicians who use “Latinx”…Latinos who oppose “Latinx” are less likely to support politicians who used or are associated with “Latinx”…Latinos in areas where “Latinx” is more salient are more likely to switch their vote toward Trump between 2016-2020[.]

One common interpretation of this result — and one the authors discuss — is that many Latinos are conservative, and that this makes them susceptible to “anti-LGBTQ+” labels. I’m highly skeptical of this interpretation. Outside small, highly educated, deeply progressive circles, who even knows that the purpose of the “x” in “Latinx” was to be trans-inclusive? Almost no one who is likely to be offended by the term is even aware of this.

Instead, what they hear is simply someone making up a weird label and applying it to them. It’s what sociologists call ascription — it’s someone making up a racial category for someone else and then stuffing them into it without their consent. The Occam’s Razor explanation for Hispanics’ dislike of “Latinx” is that it minoritizes them — it declares that they’re part of a special “community”, separate from other Americans.

Democrats’ purpose in assigning Americans to “communities”, of course, is not to exclude them, but to reward them — to hand them special prizes or favors via government policy. For example, the Biden administration had a whole policy oriented around giving government benefits specifically to “HSIs” — an abbreviation for “Hispanic-Serving Institutions”, i.e. colleges that a lot of Hispanic people tend to go to. Harris also had a plan for encouraging small banks to give Hispanic people more loans.

This is, fundamentally, a special-interest model of politics — it’s similar to the kind of quid pro quo that historically dominates much of urban machine politics. I’m not sure whether today’s Dems inherit their instinct for this kind of group giveaway from their old Tammany Hall days, but it’s recognizably similar to the way California cities hand out cash to favored nonprofits — especially because the programs are often targeted at colleges or banks or nonprofits or other organizations that claim to serve the racial “community” in question.

But there are notable downsides to this approach. First, unlike in a city, these targeted initiatives probably don’t make it through to the public consciousness. What Puerto Rican voter in Philadelphia even knows what the heck an “HSI” is? But more fundamentally, I think most Americans, including most Hispanic Americans, would rather be treated as individuals than as members of a racial “community”.

Identity politics and minoritization

Identity has many components. For my part, I identify as an American, a Texan, a liberal, a rabbit keeper, a writer, a nerd, a PhD holder, a science fiction fan, a Jew, a white person, a Lithuanian-American, and so on. If politicians primarily spoke to me as part of the “Jewish community”, I would be annoyed, because I don’t necessarily want to be defined by that component of my identity. I don’t reject it, but I don’t want to be reduced to it either. And if politicians tried to appeal to me with policies targeted through Jewish organizations to benefit Jews, I would simply not care. In fact, I would be a little conflicted about accepting any of those benefits, since it would pigeonhole me into prioritizing an identity that exists as only one component among many.

Even more fundamentally, being constantly minoritized — being told to think of myself as a Jew first and as an American second — would pressure me to shrink the circle of people that I consider part of my own in-group. Because I think of myself as an American first and foremost, I feel like there are 330 million other people who share my primary identity — 330 million people on my “team”. But if I were to think of myself first and foremost as a Jewish person, that would shrink my in-group to only 16 million, almost none of whom live near me. It would make me feel more isolated, and might even cause me to start questioning whether the 323 million non-Jewish Americans who live all around me have my best interests in mind.

In other words, being set apart from your fellow countrymen and neighbors is generally not a desirable thing, even if the reason is to put you on a pedestal.

I suspect many Hispanic Americans feel the same way. Talking to Hispanic folks (remember, I’ve lived most of my life in either Texas or California) and reading books like The Diversity Paradox, The New Americans?, and The Great Demographic Illusion, it seems clear that most would like to think of themselves as regular Americans — not to give up or forget their national origins or ethnic identity, but to prioritize their American-ness.

The progressive movement of the 2010s told us that this was a bad thing. “Assimilation” became a dirty word in progressive circles, and the University of California told their faculty not to call the U.S. a “melting pot”. The implicit assumption was that given the choice, people would like to prioritize their racial and ethnic identities, and that only forcible conformity pressure from American institutions — or from White people — was preventing people of color doing that. The Democratic party’s embrace of identity politics — its habit of targeting racial “communities” with encouraging rhetoric and specific material benefits — was seen as a way to counteract the pressures of assimilation.

The conservative counter-narrative was that race relations had been fine right up until progressives started making everything in America about race. Conservatives always send me the following chart:

This is a powerful argument. If structural racism was really so entrenched before 2014, then why did two-thirds of Black adults say race relations were good? Essentially, the progressive story is that they were effectively bullied into saying this — or perhaps bullied into thinking it. The conservative story is that progressives created the decline in race relations by making everyone think about race all the time.

I suspect that neither of these stories gets it quite right, but that the conservative story is closer to the truth. Increasing the salience of racial divisions and racial conflict makes people mistrust their neighbors more. And although both sides were guilty of this in the 2010s, progressives were far more zealous and absolutist about it.

Essentially, the progressivism of the 2010s sought to turn American society into a sort of estate society, where people’s rights, privileges, and social roles would be mediated by their identity group. How your teachers and principals treated you, your relationship with the law, what colleges were available to you, what kind of parties you could go to, what jobs you could get hired for, how your bosses treated you at work, what kind of benefits you could expect from your government, and even how people walked past you on the street were to be determined by the race you were born into, rather than your own actions as an individual.

These race-based social relationships were to be enforced by a combination of DEI departments, lawsuits, social shaming, job loss, and so on. Ibram X. Kendi even suggested the totalitarian step of creating a constitutionally mandated “Department of Antiracism” to enforce race-based social relations at every level of society — all in the name of combatting “structural racism”, which Kendi couldn’t even define when asked.

Democratic party politicians never fully embraced this vision of the future, even in 2020-21. Some of these ideas were tried out in 2020-21, but they failed almost as quickly as they were tried, making basically everyone mad in the process. By 2024, Kamala Harris was tacking pretty strongly away from the racialized vision of America.

But that was too little, too late. People remember the individualistic Americanism largely prevailed in public life before 2014, and they remember the constant hyper-consciousness of race in 2014-21, and they know which one they like better. And although Democrats abandoned much of their woke rhetoric, the practice of targeting politics at specific “communities” has persisted. I suspect that it refreshes the bad memories of the racial conflicts of 2014-21.

All of this is a long-winded way of saying that Democrats’ racial targeting probably makes a lot of Hispanic people feel singled out and alienated in their own country, even if the purpose is to help them.

But that’s not all. Democrats’ identity politics flies strongly in the face of the narrative that has defined the American experience for many Hispanics.

The story of Hispanic voters is defined by the American Dream

To the extent that most Hispanic Americans do share a common experience, it’s the immigrant experience. Immigrants come to America seeking the American Dream of upward mobility and opportunity. And by and large, they achieve it. Their children tend to earn a lot more than they do, and their grandchildren earn even more.

This was the path that most immigrant groups have followed — the Irish, the Italians, the Lithuanians, the Chinese, and so on. For a while, people argued that Hispanics were different — that they would be permanently excluded from the American Dream. That turned out not to be true:

For many Hispanic Americans, their salient experience of America has not been one of racialized exclusion, but one of upward economic mobility and social inclusion. Economists have found that Hispanic incomes are converging to White levels across the generations. Once you control for the composition effect of new poor immigrants coming in, there’s strong income convergence across the generations:

We study the sources of racial and ethnic disparities in income using de-identified longitudinal data covering nearly the entire U.S. population from 1989-2015….[T]he intergenerational persistence of disparities varies substantially across racial groups. For example, Hispanic Americans are moving up significantly in the income distribution across generations because they have relatively high rates of intergenerational income mobility…

Hispanic Americans are moving up significantly in the income distribution across generations. For example, a model of intergenerational mobility analogous to Becker and Tomes (1979) predicts that the gap will shrink from the 22 percentile difference between Hispanic and white parents observed in our sample…to 6 percentiles in steady state…

Hispanics are on an upward trajectory across generations and may close most of the gap between their incomes and those of whites…Their low levels of income at present thus appear to to be primarily due to transitory factors.

In fact, despite composition effects from continued immigration, Hispanic wages have grown faster than White wages in America since the turn of the century:

Surveys unsurprisingly have found widespread economic optimism among Hispanics for many years. Hispanics are climbing the ladder of education:

Hispanic homeownership is rising strongly. And Hispanics are starting businesses at very high rates:

New businesses owned by Latinos accounted for 36% of launches in 2023…As of 2023, Latino entrepreneurs have established a staggering five million businesses across the United States, collectively generating over $800 billion in annual revenue…The rapid pace of this growth is further highlighted by the fact that the number of Latino/a-owned businesses surged by 34% from 2007 to 2019.

In other words, Hispanic Americans’ economic path looks reasonably similar to that of Italian Americans or Irish Americans.

In the 2010s, it became taboo to even talk about these successes in some progressive circles. The University of California’s 2015 speech guidelines told faculty not to say “America is the land of opportunity”. This was deemed to be a racist utterance, since it supposedly minimized the economic exclusion that many minority groups have faced.

But the fact is, America is the land of opportunity, especially for immigrants and their children. Hispanic Americans, broadly speaking, have been prime beneficiaries of that upward mobility — they have worked incredibly hard, gone to school, started businesses, learned English, and done everything else they were supposed to do. And America has rewarded them with economic success and social integration.

Compared to that self-bootstrapped success, racial preferences — DEI, affirmative action, preferential government contracts, and so on — are basically a booby prize. Yes, in reality, those policies can sometimes help accelerate Latino achievement. But the mere fact that they’re on offer — and that Democrats seemingly view them as permanent policies instead of temporary ones — sends the message that Hispanics are a permanently racialized, excluded, marginalized group. This message is often explicit — when Democrats and progressives talk about “marginalized groups”, which they often do, that term is generally supposed to include Hispanics.

Being a marginalized group is simply not part of anyone’s American Dream.

The tricky part for Democrats is that this is an approach that has largely worked with Black voters — or at least failed to alienate them. It will be difficult for Democrats to convince themselves that Hispanic voters are not going to work the same, after spending so many years believing they would. But the 2024 election, and the trend of the last few elections, makes it clear that the two are not going to work the same.

After identity politics

Identity politics allows Democrats to hold on to a supermajority of the Black vote, so I suspect it will never go away. But the Black vote is a smallish and shrinking part of the electorate, and even there, the power of idpol may be fading a bit. Meanwhile, Hispanic voters, Asian voters, and Middle Eastern voters are the future. Those groups are all strongly upwardly mobile, and mostly descended from people who came to America voluntarily in search of the American Dream. And although ethnic studies professors and progressive activists will happily tell you all about the Chinese Exclusion Act and post-9/11 Islamophobia, it’s just not the case that Hispanic, Asian, and Middle Eastern Americans have anything that compares to the history of slavery and segregation.

The Democrats need to move away from identity politics. And switching to class-based politics, as some have suggested, won’t really help. Biden was probably the most pro-union and pro-worker President America has had since the mid 20th century, and yet some prominent unions refused to endorse Harris.

Instead, Democrats need to retool their message to be about American identity, as it was in the New Deal days, and as it arguably was for Clinton and Obama. I believe Harris actually took steps in the right direction, but not enough, and not nearly soon enough — and she was probably the wrong person to sell the idea. My advice is not to be discouraged, and to keep at it — Walter Mondale and Michael Dukakis both went down to crushing defeats with centrist, positive messaging, but then Bill Clinton eventually figured out how to make the approach work.

Republicans will eventually screw up. Trump is old, and he has a lot of foolish and bad people around him, and his economic policy program is kind of a dead end. At some point, America will once more call upon the Democrats to save them from Republican mistakes, as they did in 2008 and in 1992. Democrats will have an easier time rising to that inevitable summons if they leave the identity politics of the 2010s behind. Because that stuff just isn’t getting the job done.

Americans hate inflation more than they hate unemployment

In 2021, there was a fierce debate, both within the Biden administration and among the general left-of-center commentariat, about how much to worry about inflation. For example, in May 2021, Larry Summers wrote:

Even six months ago, it was reasonable to regard slow growth, high unemployment and deflationary pressures as the predominant risk to the economy. Today, while continuing relief efforts are essential, the focus of our macroeconomic policy needs to change…Inflationary pressures are mounting…How much does it matter whether inflation accelerates? In general, increases in inflation disproportionately hurt the poor and are associated with reductions in trust in government. Progressives might consider the role that inflation played in electing Richard M. Nixon in 1968 and Ronald Reagan in 1980.

In February 2021, Olivier Blanchard, another prominent macroeconomist, wrote a detailed argument that the Biden administration was spending too much on Covid relief with the American Rescue Plan, and that this would generate a surge of inflation. The argument included a simple model whose predictions ended up being close to what actually happened.

Other economists pushed back hard against Blanchard and Summers, arguing that inflation was due to transitory factors, and that it was important to keep spending in order to preserve full employment. One of these was Paul Krugman, who declared himself “a card-carrying member of Team Transitory” — “Team Transitory” being those who thought inflation was mostly due to pandemic disruptions and would fade on its own. There was even a live debate between Krugman and Summers in early 2022. But by late 2022, Krugman issued a mea culpa, and admitted that Biden’s American Rescue Plan had exacerbated inflation:

In early 2021 there was an intense debate among economists about the likely consequences of the American Rescue Plan, the $1.9 trillion package enacted by a new Democratic president and a (barely) Democratic Congress. Some warned that the package would be dangerously inflationary; others were fairly relaxed. I was Team Relaxed. As it turned out, of course, that was a very bad call.

Economic analyses of the inflation surge of 2021-22 generally assign the American Rescue Plan some share of the blame.

Ignoring Summers’ warnings, and believing Team Transitory, might have cost the Democrats the 2024 election. I’m no political scientist, but there are reasons to believe that the burst of high inflation in 2021-22 was one factor in Kamala Harris’ loss.

Why inflation probably mattered in this election

The first reason we should think inflation mattered is that people pretty consistently said it mattered. On Gallup’s long-running “most important problem” survey, inflation consistently ranked as the most important specific economic problem that people mentioned, right behind “economy in general”.

A lot of other surveys showed similar results. And the percentage of people who told Gallup that the economy was “extremely important” to their presidential vote was unusually high — almost as high as in 2008, when the financial crisis had just devastated the U.S. economy:

And of course consumer sentiment was pretty low, even after adjusting for changes in methodology:

Now, it’s possible that this was mostly just Republicans complaining about inflation because it felt more substantive than talking only about more social/cultural grievances like wokeness, immigration, trans stuff, etc. After all, opinion on the economy shows a very partisan pattern depending on who’s in the White House:

It’s a pretty good bet that Republicans will suddenly “discover” that inflation in America is actually low now, and give Trump the credit.

But if you look historically and across countries, it really looks like inflation tends to make voters particularly mad. Research often finds that cumulative inflation over a President’s term in office — not just the current inflation rate — affects presidential approval and election performance. If you look at Gallup’s long-running survey, you’ll see that the inflation of the 1970s dominated Americans’ complaints even more than the unemployment of the Volcker recessions in the early 80s, or the recession of the early 90s:

And if you look at other countries, you’ll see that governing parties in every other rich country have seen their vote shares go down this year:

There are relatively few factors that are affecting all rich countries at the same time. A backlash over immigration is another possible common factor, but the list is basically just that plus inflation.

On top of that, there was a model that predicted the election outcome quite well. An interdisciplinary team of researchers created a model based on presidential approval ratings and economic conditions in each state, and they ended up predicting each state correctly:

It might be a fluke, but the same model got only one state wrong in 2020 (Georgia), so that’s pretty encouraging. In any case, the model relies partly on economic fundamentals — measures of the labor market and real wages. And real wages depend on inflation. So since labor markets are generally very strong all across America, the erosion of state-level wages by inflation is going to be the main economic factor that led this model to forecast a Trump victory.

So I think it’s safe to say that while it’s possible inflation mattered less for Trump’s victory than people say on polls, it did matter at least somewhat.

Why inflation makes voters even angrier than unemployment

Let’s talk a little about macroeconomics for a second. Most mainstream macroeconomic models assume that there’s some sort of short-term tradeoff between inflation and unemployment. If the Fed lowers interest rates or does quantitative easing or otherwise prints money, the general consensus is that unemployment will go down and inflation will go up. Similarly, if the government borrows and spends money and the Fed doesn’t counteract this by raising interest rates, the general consensus is that you’ll get faster growth, more jobs, and rising prices.

Macroeconomic policy then becomes a question of how to strike the right balance. We want everyone to have a job. We don’t want inflation to go about 2%. But if we have to choose between these two things, how do we choose? In economic models, the relative degree that policymakers should care about inflation versus unemployment is determined by a social welfare function — basically, how much each of these economic ills makes regular people unhappy. A lot of economic research is dedicated to figuring out what that welfare function, and there’s not really a consensus.

But the experiences of the 1970s and the 2020s are teaching us that in America at least, people seem to get even madder about inflation than they do about unemployment. And the question is why. I actually wrote a post about this back in March of 2021, shortly after the American Rescue Plan was passed:

The basic story here comes from a survey Robert Shiller did in 1990. Basically, inflation tends to make people’s purchasing power go down, because their wages can’t keep up. This makes people poorer, and people do not like to be poorer.

The inflation of 2021-22 definitely made people poorer. In fact, the drop in real personal disposable income from mid-2021 to mid-2022 was more severe than the Great Recession or the 1970s inflation!

This was caused by a combination of the end of pandemic relief, and falling real wages and other income due to inflation. In 2021 and 2022, Americans’ real wages suffered their biggest drop in postwar history, thanks to inflation:

Remember, real wages are a key economic input that went into the model that managed to successfully predict every state’s election result. People really don’t like it when their real wages go down!

One key thing to understand is that this sudden impoverishment happened to most Americans. Rising consumer prices affect everyone, from the rich to the poor, the employed and the unemployed, the old and the young.

In a recession, on the other hand, the negative effects are largely concentrated. If you’re one of the unlucky people who loses their job, you take a huge hit. But if you’re not one of those people, you probably do OK. Maybe your wages grow more slowly for a few years, and maybe you have to stress out about the increased possibility of being laid off. But in general, the harms of inflation are diffuse while the harms of unemployment are concentrated.

If you’re a moral philosopher, an activist, or an economist estimating a social welfare function, you may prefer harming a lot of people a little bit with inflation to harming a few people a lot with unemployment. But a democracy doesn’t vote based on social welfare. It’s one person, one vote. So if unemployment harms 10 million people a lot, and inflation harms 200 million people by a moderate amount, it’s clear that inflation will directly harm a much larger number of voters.

The social welfare function and the “get your butt kicked in an election” function are two very different things.

I actually suspect that there are other reasons why inflation bothers voters even more than unemployment. Inflation is something that’s obviously beyond an individual’s control — if the price of eggs goes up, you can choose not to buy eggs, but otherwise there’s not much you can do. But unemployment is something that individuals do exercise some degree of control over — if you work a bunch of extra hours during a recession and establish a good relationship with your boss, maybe you can avoid getting laid off.

In general, I think people are less worried about risks they have some control over, even if the overall level of danger is higher. Look at how people are more afraid to fly than to take long road trips, despite road trips being much deadlier per mile. The leading theory is that in a car you have some control over your safety, while in a plane you have basically zero control. Similarly, I suspect that inflation makes people feel powerless, like being stuck in a crashing airplane.

In any case, this is probably an important area for survey research to explore. But in the meantime, Democrats need to remember that full employment doesn’t necessarily mean an economy that Americans are happy with. If it comes at the expense of rapid inflation, that price is going to be one that most Americans aren’t comfortable paying.

Democrats should be wary of macro-progressivism going forward

I’d like to say a little more about the debate over the American Rescue Plan in 2021. While Paul Krugman and Larry Summers viewed their argument in purely technical terms, some economists and commentators did come at the debate from an ideological angle. In general, spending more was considered the progressive position, while spending less in order to avoid high inflation was considered more centrist or moderate.

This makes sense — full employment disproportionately benefits the least well-off in society. The lowest-paid workers are often the first to be fired in a recession and the last to be hired in an expansion, so the longer you can keep an expansion going, the more poor people can have jobs. Also, full employment tends to push up wages, especially at the lower end of the scale — in fact, even as most Americans lost purchasing power during the inflation of 2021-22, the bottom 10% of American workers actually saw their real wages rise.

This July, I wrote about how some influential progressive think tanks basically push for full employment all the time, and downplay worries about inflation:

Some progressive economics commentators — I will not name names, because I’m trying to be a less confrontational blogger these days — also consistently downplayed the risks of inflation and extolled the many benefits of full employment during Biden’s tenure in office.

In retrospect, whether or not this was a good move from a moral standpoint, it looks like a big mistake from an electoral standpoint.

I hope that these commentators are reevaluating their beliefs. And I hope that Democratic politicians, staffers, and donors are reevaluating the paradigm as well. Full employment is great, all things equal. But sometimes the price of full employment is too high — not just for the majority of Americans who have to endure years of erosion in their purchasing power, but for everyone who has to suffer from the election of leaders like Donald Trump.

Update: Good to see some left-leaning economists quickly coming to the same conclusions.

Our task now is to get the ear of Democratic politicians, which will probably require wresting it away from some progressive think tanks and Twitter posters.

The educated professional class is out of touch with America

I’ve been writing about the lessons that Democrats should take away from the big GOP victory. I’ve written about how identity politics is failing to attract Hispanic voters, and about how Dems cared too much about employment and not enough about inflation. The third big lesson, I think, is about class in America. The educated professional class is drifting away from the rest of the country, in terms of values, beliefs, and their information diet.

This election upended a lot of demographic trends that people had been very certain about only a few years ago. Trump seems to have produced a large-scale shift of Hispanic voters to the Republicans, and big cities swung harder toward the GOP than other areas. But there’s one key trend that this election didn’t reverse: education polarization. Trump was even more dominant among non-college voters than he was in the last two elections, while Harris won the highly educated by about as much as Biden did.

Over the years, the Democratic base has shifted decisively from the working class to the educated professional class:

This shift doomed Harris’ campaign, since people without college degrees outnumber their college-educated counterparts — in 2024, the latter were only 43% of the electorate.

And this weakness looks likely to be a long-term problem. In the past, when college education was steadily increasing, it was easy to think that college-educated voters would own the future. But in recent years, the rate of college enrollment has plateaued and begun to fall:

And the number of new college graduates fell in 2022, even as America’s total population grew.

As for why fewer Americans are getting a college education, I wrote a post about this back in July:

The short version is that tuition soared even as the earnings boost from a college degree shrank substantially, making college a less attractive value proposition. A lot of this had to do with colleges themselves boosting spending on administrators and facilities, and having to charge students more to pay for it.

If universities cut costs, cut tuition, and raised their value proposition, we could see enrollment rise again at some point, but I wouldn’t bet on that happening soon. Meanwhile, socialists’ plan for universal free college is dead on arrival; even Biden’s generous new student loan repayment provisions might prove to be temporary, since the fiscal cost is so huge. So a wave of taxpayer cash is not going to restore the trend of rising educational attainment.

In other words, Democrats’ education problem is here to stay. Unless Dems can figure out how to win back those non-college voters, they’re likely to be at a huge electoral disadvantage for the foreseeable future.

But winning back those voters would require the educated professional types who make up the backbone of the Democratic base and the progressive movement to figure out how to connect rhetorically with the less-educated masses, addressing their concerns, speaking to their values, and just speaking their language in general. And although it’s too early to tell, many of the reactions I’m seeing on social media are, to put it mildly, very unencouraging:

Obviously these are extreme examples, but in general it seems to me that the educated professional class has simply grown apart from the rest of the country over the past few decades.

A disconnect on problems

The most important disconnect between educated professionals and the rest of America is about the problems they face on a day-to-day basis. Practically every problem that Trump voters cited is something that hits the working class harder than the professional class.

About a year ago, I found a really good place to get a haircut. The hairstylist, a Vietnamese immigrant woman in her 50s, was complaining about rampant theft from small businesses in the neighborhood. She said that if thieves stole less than $950, they wouldn’t even be arrested or prosecuted, creating an incentive for constant theft.

In fact, stealing less than $950 was still a crime, but there was a plebiscite in 2014 called Proposition 47 that changed it from a felony to a misdemeanor. (Prop 47 was mostly repealed this year, when California voters passed another ballot measure called Prop 36.) I’m not sure how big of a factor Prop 47 was, but for whatever reason, property crime in San Francisco did rise quite a lot in the late 2010s:

This mirrored a more general rise across the state:

I am an educated professional, not a small local businessperson. I had heard about the big wave of car break-ins in San Francisco, but it took me a very long time to realize how common retail theft had become. To me, the wave of smash-and-grab attacks in the Bay Area was a sign of urban disorder and decay, and a force causing a bunch of downtown stores to close.

But to someone who owns a hair salon that gets robbed, or a convenience store that gets looted, that’s their livelihood. That could be their child’s ticket to college, or their ability to get the pipes fixed at their house, or their chance to get their car fixed. As an educated professional, I’m simply much less exposed to robbery and theft.

And on top of that, I’m an urban educated professional. Those who live out in the fancy suburbs are even more disconnected from the crime wave that afflicted America in 2020 and 2021. It’s lower-earning people in cities or poorer neighborhoods who bore the brunt of the explosion of disorder that followed in the wake of the Floyd protests.

Stuff like “defund the police” is a luxury of the rich and mobile — i.e., of the educated professional class. It’s lower-income Americans and small business owners who need the police the most.

How about inflation? When defending the Biden economy, I often point out that real wages for people in the lower part of the distribution increased faster. That’s adjusted for inflation, but it’s adjusted for the overall state-level inflation rate. In fact, as the Minneapolis Fed’s Jeff Horwich explains, people who make less money have faced higher price increases since 2021, for two reasons. First, the things that lower-income and middle-income people spend a larger percent of their income on — especially housing — tended to go up in price a bit more:

The second reason is pretty obvious: Having less money makes drops in purchasing power harder to bear. If my cost of living went up by 10%, it would be a minor annoyance; to a working-class person that could very well be ruinous. No wonder lower-income people tend to notice price changes more:

So while educated professionals like myself were able to put the post-pandemic inflation behind us fairly quickly after it subsided in 2023, regular Americans probably felt a more lingering hurt.

How about the wave of asylum seekers that flooded across the border in the Biden years? Lower-skilled immigrants don’t actually tend to reduce wages for native-born working-class Americans. But a big part of the reason they don’t reduce wages is that native-born people respond to increased immigration by going back to school and climbing the skill ladder. That takes time and money.

Also, low-skilled immigrants directly compete with less-educated and lower-earning Americans for housing. Asylum seekers tend to be poor, meaning they will try to get apartments and houses in lower-income neighborhoods. The effect on local rents in those neighborhoods, and for those types of housing, will be greater than the overall citywide or statewide effect. Rent will go up for the American working class, differentially affecting their purchasing power in ways that don’t get captured by state-level or national statistics.

So although Biden did beat inflation, crime did go down, and Biden did eventually crack down on the border, I can understand why the traumas that working-class Americans had to suffer in 2020-2022 couldn’t be expunged just by some encouraging charts. And the fact that educated professionals were much more insulated from the disruptions of the pandemic and from Biden’s missteps in 2021 mean that they didn’t necessarily even see the need to acknowledge how hard these had been on everyone else.

This disconnect probably ended up alienating people without college degrees. Not only did it cause many progressives to support deeply unpopular things like “defund the police” on social media, but Democratic politicians who became used to talking to their educated progressive base probably got a distorted idea of what was troubling the rest of America.

An information disconnect

I was taking an Uber with a friend a couple of weeks ago, and the driver went on a gigantic rant about how inflation is out of control, vaccines are ineffective and are causing all kinds of massive health problems, there are no jobs anymore, and so on. Listening to his rant drove home the fact that there are huge segments of the American populace that I’m just completely out of touch with.

Almost all of what he said was factually wrong, of course. Inflation is back down to a low level, and has been for years now. Employment is near all-time highs. Vaccines are safe and effective, and so on. But the fact that he believed in all of this stuff is a sign that there’s an information ecosystem out there that’s totally separate from the world of blogs and Washington Post articles and nerdy podcasts and Econ Twitter that comprises my own information diet.

In fact, this information disconnect is probably one reason why Trump won. Voters who knew about the positive data that I know about — and which I frequently talked about on this blog — went very strongly for Harris, while voters who had the same misapprehensions as my Uber driver went pretty strongly for Trump:

Now, people may simply be bullshitting on these polls — there’s research that shows that when you pay people to get these facts right, they tend to get them right. Saying “inflation is still high” might just be a way of saying “I like Trump”.

Also, note that all four of these questions are specifically phrased in such a way that the correct answer says something positive about Joe Biden’s tenure in office. If the questions had been things like “Overall, murder rates were higher under Biden than under Trump (True)”, or “Inflation in the U.S. since 2021 has been the highest of any four-year period since the 1970s (True)”, or “Since 2021, unauthorized border crossings have been higher than at any point since the early 2000s (True)”, it might have been Democrats getting the facts wrong instead of Republicans. (We’ll never know, of course, but Democrats do tend to show partisan bias on economic surveys.)

But it also seems likely that many of these less-educated Trump voters just don’t get nearly as much information about politics, economics, and public affairs. Highly educated folks consume a lot more news in general. And people who consume a lot more political news tended to vote for Harris:

And when they do consume news, it’s likely to be from different sources. Many election postmortems have focused on the fact that Joe Rogan, a popular podcast host with many blue-collar fans, interviewed both Donald Trump and JD Vance during the campaign, but not Kamala Harris or Tim Walz. Joe Rogan’s podcast has 18 million subscribers. In comparison, Fox News, the most popular cable news network, has somewhere around 2 million primetime viewers. Meanwhile, Barstool Sports, another news outlet that’s popular with blue-collar types, went solidly for Trump and reportedly refused Harris’ offer for an interview.

In other words, Democrats are still doing a good job of getting their message to educated Americans who consume lots of political news, but are increasingly failing to reach the non-college masses. That’s why listening to an Uber driver talk about politics can seem like meeting someone from another universe.

This is not to say that educated progressives get the facts right, while blue-collar conservatives get them wrong. In fact, highly educated progressives can talk themselves into believing some absolutely insane things. During the summer of 2020, a lawyer friend insisted to me that the institution of police grew out of the institution of American slavery (obviously wrong, since police emerged independently all over the world). A number of progressives also asserted to me that police don’t reduce crime (false). The oft-repeated progressive claim that slaves built America is also obviously false, since slaves were less than 10% of the population. A few educated progressives even fiercely defend intentional fabrication of results in the social sciences, on political grounds.

But whatever the various myths are on the left and right, they are different myths. And because educated professionals swim in entirely different information ecosystems, they have little clue how to reach the masses and tell their side of the story.

Why did the class disconnect come about, and how can we reduce it?

“Her belly may be full, but her spirit will be empty.” — Jean-Luc Picard

The educated professional class has been growing apart from America’s other social classes for decades now. Part of this was due to the big shift in the U.S. economy — the rise of knowledge industries like IT, finance, biopharma, and business services pulled America’s talented people to coastal metros and college towns. As the economist Enrico Moretti showed in his excellent book The New Geography of Jobs, a city’s human capital — i.e., how many educated professionals live there — became the best predictor not only of economic success, but also various measures of public health and well-being.

For those with the talent to escape the bleakness of their hometowns, big cities like New York, San Francisco, and Minneapolis — or college towns like Ann Arbor, Madison, and Berkeley — became magical playgrounds, full of great restaurants and shops and bars, other educated professionals to date and marry, and plenty of money to be made. I was one of the people who made that escape.

But what happened to the places the talented people left behind? Bereft of their talent in an age when talent equals money, they stagnated economically — counties that voted for Biden in 2020 represented just over half of the electorate, but 70% of America’s total GDP. And without as many talented people to be local doctors, teachers, etc., those places saw their health deteriorate — drug use and obesity ran rampant. Even the well-documented decline of marriage among the less-educated could be partially due to talented professionals fleeing the area — without as many visible examples of economically successful educated people getting married and staying married and raising kids in two-parent families, some working-class people in left-behind towns may have simply abandoned the practice.

This is the first reason for the disconnect. The second, I think, was institutional. In earlier times, there were a number of institutions that regularly brought people together across social classes — the three main examples being churches, businesses, and the military. But after Vietnam, the U.S. Military shifted from a mass-participatory conscript force to an elite professional volunteer force. Business became increasingly fragmented, with much blue-collar work and low-level service work outsourced or automated. And church membership has declined in America, possibly as a result of the internet:

Combine this with the de-localization of American social life due to the car and the internet, and you have a recipe for extreme sorting. As Tyler Cowen writes in his book The Complacent Class, Americans have become increasingly expert at sorting themselves according to education, income, and other characteristics.

There was one public institution that managed to retain its broad functionality over this time period: College. American universities tried hard to recruit and mix people of diverse backgrounds, teach them healthy personal and social habits, and provide them with a rich community life that could be a model for their later adult communities. And to a large degree, this worked.

But it only worked for 40% of the population! The majority of Americans lacked the talent, the study habits, the inclination, the time, or (eventually) the money to get that four-year degree. So for that lucky 40%, colleges were church and army and work team all in one — an acceptable substitute for the tight-knit villages that the modern economy of the car and the internet had destroyed.

For the rest of the country there was nothing. There were daytime TV and televangelists and methamphetamine and conspiracy theories. There were strip-malls and NASCAR and hip-hop and Marvel movies. There were Fox News and Rush Limbaugh and Michael Savage. There was plenty of cheap land, big houses, cheap cars, and jobs in the vast service sector. But their upper caste had fled them for the playgrounds of college and the big city and the Google cafeteria.

But they were still part of the same country as the educated professionals. And so they all still voted in the same elections. So one day Donald Trump came along, and the classes were reunited by the institution of electoral democracy, in a very unpleasant and painful way.

After this crushing defeat, many educated progressives will turn to the idea of Bernie Sanders-style “class” politics as an alternative to the identity politics that failed so spectacularly. But this won’t work either.

First of all, Bernie himself underperformed Kamala Harris in this year’s election. Second, and more importantly, Biden has been the most pro-labor, pro-union President in many decades, and it availed the Democrats nothing. Biden raised overtime pay, strengthened unions with various new laws, revived factory construction with industrial policy, mandated the hiring of union workforces for federal contracts, became the first President to walk a picket line, funded vocational education and apprenticeships, banned noncompete clauses, and took dozens of other actions in support of the blue-collar working class.

Wages rose strongly for the working class, and inequality fell. Biden also enacted a lot of Bernie-style policies — canceling student debt, passing a huge climate bill, and increasing subsidies for health care.

And none of it helped Kamala Harris. The Teamsters union withheld their endorsement. The East Coast dockworkers’ union threatened a crippling strike just weeks before the election. All across the nation, steelworkers and other blue-collar union workers defied their leaders and organized in support of Donald Trump. Harris did win a slim majority of the union vote, but this probably came mostly from educated professional types, especially public unions.

The fact is, America just isn’t the industrial society it was in the 1950s. The broad blue-collar manufacturing laborer class is a thing of the past, and the fragments of that class that remain are often more focused on sociocultural issues than on pure pocketbook concerns. Most of the American non-college populace are non-unionized suburban service workers with no real class consciousness.

America, fundamentally, is an incredibly rich country with a sick society. We have higher consumption by far than people in any other country on the planet, and yet by and large, except for the few people who were lucky enough to receive the blessings of community and health from our one remaining functional institution, we’re a bunch of unhealthy socially isolated drug addicts.

As for the solution to all this, I don’t have an easy fix. In the long run our country needs to build new inclusive institutions — places where society’s classes mingle and cooperate. I don’t know what those can be, though of course I’ll think about the problem. But it will take a long time. Obviously, YIMBY policies that allow working-class people to live in big cities will help a bit, by putting the various classes in physical proximity, but without integrating institutions the effect will be marginal.

Nor do the educated right-wing professionals in the tech industry understand their new working-class political allies any better than the progressives do. Beyond helping to elect Donald Trump and annoy educated progressives, they don’t really have any plans to help those who lack the talent to work at SpaceX.

In the short term, what Democrats can do is to realize the extent of the problem, and begin doing everything they can to explore the culture and the information ecosystem of Americans who didn’t go to college. Get interviewed on Rogan and Barstool Sports and wherever, but don’t just get interviewed — immerse yourself in the culture. Don’t get your information about regular Americans from the activist groups who knock on your staffers’ doors, or from 20-something socialists who spout stale Marxist theory. Go meet them, find out what they care about, and think about how to give it to them.

Fixing America’s great class disconnect will take a long time, but there’s no better time to start.

The blue cities must be fixed

Donald Trump’s victory in November 2024 wasn’t the only important election result. Big cities in California shifted away from so-called “progressive” governance. In San Francisco, progressive champion Aaron Peskin was soundly defeated in the mayoral race, and his vacated seat was won by a moderate. Peskin’s close ally, Dean Preston, was unseated by moderate challenger Bilal Mahmood, and the Board of Supervisors is expected to lose its progressive majority. In Los Angeles progressive district attorney George Gascón lost his reelection bid, while Oakland’s Pamela Price was recalled. This followed a national trend of progressive DAs getting the boot; meanwhile, San Francisco’s moderate DA Brooke Jenkins, who succeeded the notorious Chesa Boudin after the latter was recalled, easily cruised to reelection.

It’s possible to interpret these victories as part of a general revolt against progressive governance in big cities. At the local level that’s a good thing, since that governance has proven remarkably poor. But at the national level, the urban revolt against probably helped usher in another four years of Donald Trump. Although large metros still voted for the Democrats overall, they swung harder towards Trump than any other geographic region:

Derek Thompson has some more numbers:

In the past two cycles, Democrats could comfort themselves by counting on urban counties to continue moving left even as rural areas shifted right. That comfort was dashed last night, at least among counties with more than 90 percent of their results reported. In the New York City metro area, New York County (Manhattan) shifted nine points right, Kings County (Brooklyn) shifted 12 points right, Queens County shifted 21 points right, and Bronx County shifted 22 points right. In Florida, Orange County (Orlando) shifted 10 points right and Miami-Dade shifted 19 points right. In Texas, Harris County (Houston) and Bexar County (San Antonio) both shifted eight points right and Dallas County shifted 10 points right. In and around the “Blue Wall” states, Pennsylvania’s Philadelphia County shifted five points right, Michigan’s Wayne County (Detroit) shifted nine points right, and Illinois’ Cook County (Chicago) shifted 11 points right.

Josh Barro wrote an insightful post-election post arguing that big cities’ failure to curb urban disorder, fiscal waste, and soaring housing costs drove their inhabitants away from the Democratic party. Some key excerpts:

I write this to you from New York City, where we are governed by Democrats and we pay the highest taxes in the country, but that doesn’t mean we receive the best government services. Our transportation agencies are black holes for money, unable to deliver on their capital plans despite repeated increases in the dedicated taxes that fund them, because it costs four times as much per mile to build a subway line here as it does in France, and because union rules force the agency to overstaff itself, inflating operating costs. Half of bus riders don’t pay the fare, and MTA employees don’t try to make them. Emotionally-disturbed homeless people camp out on the transit system…even though police are all over the place (at great taxpayer expense) they don’t do much about it, and I can’t entirely blame them since our government lacks the legal authority to keep these people either in jail or in treatment. The city cannot stop people from shoplifting, so most of the merchandise at Duane Reade is in locked cabinets…[S]chools remain really expensive for taxpayers even as families move away, enrollment declines, and chronic absenteeism remains elevated. Currently, we are under state court order to spend billions of our dollars to house migrants in Midtown hotels that once housed tourists and business travelers. Housing costs are insane because the city makes it very hard to build anything — and it’s really expensive to travel here, partly because so many hotels are now full of migrants, and partly because the city council literally made it illegal to build new hotels. And as a result of all of this, we are shedding population — we’re probably going to lose three more congressional districts in the next reapportionment. And where are people moving to? To Sun Belt states, mostly run by Republicans, where it is possible to build housing and grow the economy…

[A]ll of this is why I think Democrats’ approach to the cost-of-living issues…has fallen so flat. The Democratic argument is, more or less, “look at all my programs” — all the things I’m going to have the government do to make life easier for you…I think Americans look around at how it goes when the government actually tries to help, and they have a healthy skepticism about how helpful the government is really going to be, and about whether the benefits are really going to flow to them. Democrats are making too many promises; they instead need to pick a few things for the government to do really well, with a focus on benefits to the broad public rather than to the people being paid to provide the services, instead of trying to do a zillion different things and doing them badly at great expense[.]

Of course I don’t know how big of a factor this was in Trump’s victory, but I’m not actually sure that even matters. Barro is largely correct about the sad state of blue cities, and it’s the inhabitants of those cities who are suffering the most from that failure of governance. The most important reason to fix those cities isn’t to help Democrats get elected at the national level, but because it would make life better for urban residents themselves. The swing toward Trump is simply an illustration of just how angry city dwellers are.

The blue city renaissance, 1980-2012

It’s important to remember that just a few years ago, the general consensus was that big cities in America were doing really well. In 2017, Will Wilkinson cited the success of American cities as a reason to reject Trump:

[W]hen Trump talks about cities, he presents a fearsome caricature that bears little resemblance to the real urban landscape…He persists in his efforts to slur cities as radioactive war zones because the fact that America’s diverse big cities are thriving relative to the whiter, less populous parts of the country suggests that the liberal experiment works…[T]he American metropolis is more peaceful and prosperous than it’s been in decades…Yes, cities have their share of failing schools. But they also have some of the best schools in the country and are hotbeds of reform and innovation…America’s most diverse, densely populated and well-educated cities are generating an increasing share of the country’s economic output.

Wilkinson’s paean to big cities was the culmination of decades of real success. In the middle of the 20th century, America’s cities had been hollowed out, by a combination of soaring urban crime and the lure of the suburbs. In the 1970s, big metropolises were still losing population. NYC almost went bankrupt.

Starting around the 1980s, though, America experienced an urban revival. Big cities began to gain back some of the population they had lost:

There were a number of reasons for this revival. One was the rise of knowledge industries like tech, finance, and business services, which tend to cluster together in cities. As Enrico Moretti shows in The New Geography of Jobs, human capital tended to cluster in big cities after around 1980 or so; Richard Florida also famously documented this in The Rise of the Creative Class. Here’s some data from Jed Kolko for 2000-2014 that’s pretty representative of the general shift:

A second reason for the revival was immigration — just like a century earlier, immigrants flocked to big cities.

Then in the 1990s, crime plummeted. In 1992, New York city had 2,245 murders. By 2016 it was down to 292:

The combination of the crime drop and the knowledge-industry boom made American cities great places to live, at least for a couple of decades. TV shows like Seinfeld and Friends and Sex and the City and How I Met Your Mother celebrated the lightheartedness of upper-class city life. Urban art and fashion and culture bloomed again, producing the hipster era.

Why blue cities declined again

But the renaissance was short-lived. In the 2010s, the urban revival stalled, and the populations of big cities plateaued. The reason was simple: housing. Since the 70s, American cities had developed an entrenched culture of NIMBYism and a thicket of laws like rent control, which prevented cities from building new housing in order to accommodate the influx of population. As a result, rents rose. And since rich people — the same knowledge workers who were powering the urban boom — were better able to afford the rising cost of housing, it was middle-class and working-class people who were pushed out of the cities. New York, San Francisco, and the rest became playgrounds for the affluent.

More progressive cities in blue states consistently built less housing than cities in red states. Kahn (2011) found:

Traditional explanations for why some communities block new housing construction focus on incumbent home owner incentives to block entry. Local resident political ideology may also influence community permitting decisions. This paper uses city level panel data across California metropolitan areas from 2000 to 2008 to document that liberal cities grant fewer new housing permits than observationally similar cities located within the same metropolitan area. Cities experiencing a growth in their liberal voter share have a lower new housing permit growth rate.

As a result, progressive cities became less affordable:

This divide continues to this very day. Blue cities are still building a minuscule amount of housing:

At the same time, many blue cities were spending way beyond their means:

Some of this was due to the chronic problem of city pension obligations, which have grown more severe due to unrealistic assumptions about stock returns, the power of public-sector unions, and simple short-sightedness. But some was because local infrastructure and city services got very, very expensive — witness how New York City pays several times the cost per mile to build a new subway that a French city would pay.

Some of this ballooning cost is due to broken bidding processes and the erosion of state capacity. With their own departments having shrunk, progressive cities have been outsourcing more and more of their core functions to inefficient and often corrupt nonprofits. Recall the infamous case of La Sombrita, the ineffectual sun-shade that the Los Angeles Department of Transportation paid a nonprofit $200,000 to create:

San Francisco, as usual, is the worst offender, with a new nonprofit corruption scandal popping up every month or so.

In the late 2010s, blue cities brought yet another problem on themselves: urban disorder. Crime rates began rising in 2015, fueled by national unrest. But blue cities didn’t respond by cracking down on crime as they did in the 90s and 00s. Progressives in the late 2010s reviled and rejected “stop-and-frisk”, “broken windows policing”, and other tools that blue cities had used to keep order in previous decades. Instead, they elected a bunch of progressive prosecutors, enacted more permissive policies toward public drug use, passed laws that made it hard to use violence against shoplifters, and sometimes even reduced penalties for minor crimes.

The result was entirely predictable. Blue cities became increasingly afflicted by pervasive, low-level urban disorder — drug needles in children’s parks, epidemics of car break-ins, and so on. Female friends of mine in San Francisco started to report being followed for blocks, harassed on the train, or even slapped in the head by street people on their way to work. The housing crunch made the disorder much worse, of course, by exacerbating homelessness.

Then the pandemic and the riots hit, and the trend got turbocharged. Without “eyes on the street” to deter crime, and with police cowed or disgruntled by the protests of summer 2020, progressive cities became increasingly lawless, chaotic zones. Violent crime soared in 2020-21, with waves of attacks on vulnerable populations like Asian elders.

Nowhere was the chaos more evident than in San Francisco. Although murder rates in SF stayed low, more minor crimes, like car break-ins and theft, skyrocketed:

And the epidemic of public drug use in SF got utterly out of control:

After the pandemic, there was an epidemic of store looting that continued for years, causing many of the stores in the downtown area to close.

The American urban renaissance is now well and truly over. In the 2010s and early 2020s, many progressive cities squandered the massive windfall from the knowledge industry boom. With its seemingly invincible network effects and ever-increasing bounty of tax revenues, that boom seemed to convince progressive cities that they had infinite amounts of fiscal and social surplus to dispose of as they pleased — that they could tolerate crime and disorder, spend infinite money, and neglect the need to build new housing, and that none of this would end up mattering. That turned out to be wrong, and now the bill is coming due.

Three principles for reviving blue cities

Reviving blue cities will require a number of big changes. Elections like the recent one in San Francisco will be needed to push out “progressive” leaders who want to continue the failed experiments of the last decade. A sensible policy is going to have combine material abundance with intolerance toward crime and disorder. I like the basic platform advanced by GrowSF, a San Francisco advocacy group that combines pro-housing and pro-infrastructure policies with a tough-on-crime approach. Here’s a list of just a few of their good ideas:

But fixing blue cities is going to require more than just waves of voter anger or a laundry list of good policies. It’s going to require a mindset change — a shift in people’s understanding of what a city should be and how it should be run.

The most important thing blue cities need to understand and internalize is that anarchy is not a form of welfare.

Many progressives believe that any actions to curb urban disorder — restrictions on sidewalk tents, making people pay for public transit, arresting people for nonviolent crime, and so on — represent the exclusion of marginalized people from public life. In the absence of a full-service cradle-to-grave welfare state, progressives think they can redistribute urban utility from the rich to the poor by basically letting anyone do anything they want.

But in fact, permissiveness toward the behaviors that create urban disorder destroy more value than they redistribute. When you don’t make people pay for public transit, you scare people off the train, thus causing the train to go bankrupt. Pretty soon neither the rich nor the poor have a train, and everyone is screwed. This is the tragedy of the commons.

Measures to curb disorder therefore represent defense of the commons. When Bay Area trains put in new stronger gates to prevent people from riding the train without paying, crime on the train fell and rider satisfaction increased. Similarly, measures to curb shoplifting — including punishing shoplifters and letting security guards and business owners defend their stores with violence — allow local businesses to keep operating, which means that poor and working-class urban residents have places to shop for their daily needs. And so on.

Second, Josh Barro is absolutely right when he says city services should “focus on benefits to the broad public rather than to the people being paid to provide the services”. In other words, progressive cities need to realize that costs are bad for the city government.

The habit of having cities overpay for everything is another form of highly inefficient redistribution. A bunch of people do get paid out — nonprofits, overstaffed contractors, expensive consultants — but at the end of the day the ballooning costs that result from all these payouts mean that cities don’t actually have the infrastructure or services they need. All too often, progressive cities are operated for the benefit of the people who get the money instead of the people who get the stuff.

That payout-focused behavior is driving cities to bankruptcy while worsening the quality of infrastructure and services alike. Cities need to replace consultants and nonprofits with civil servants, implement stricter oversight of outside contractors, improve their bidding processes, and in general do everything they can to bring down the cost of everything they do and provide better value to the citizenry.

Finally, cities must realize that housing is non-negotiable. Fundamentally, cities are places that people live — if people do not have houses to live in, you don’t actually have a city. Housing is what allows businesses to function, because most people live near to where they work. Housing is what allows retail to function, because it creates a sufficient density of customers. Housing sustains knowledge-industry network effects, by allowing more knowledge workers to live close to each other and exchange ideas. In most states, housing generates more tax revenue for the city as well. And of course, housing reduces homelessness.

Blue cities should therefore do absolutely whatever it takes to get a lot more housing. “Whatever it takes” includes upzoning, deregulation, planning, public housing, and cutting back on policies like rent control. The correct method of building housing is “any and all methods”. That’s what it means for a goal to be non-negotiable.

I loved the blue cities of fifteen years ago. They had already sown the seeds of their own decline, but in the brief moment between the chaos of the 1970s and the foolishness of the 2010s, America’s big blue cities were indeed the glory of the country. They can and should recapture that glory.

Maybe the Rapture is already here

2026-10-09T01:18:32+00:00

Back when I was a teenager, a lot of people were very into the Left Behind series. Evangelical Christianity was still very strong in those days, especially in Texas, and Left Behind blew up by putting Christian eschatology into page-turning airport-bookstore-novel form. People really believed that the stories in those novels weren’t just stories — they were glimpses of a very real future event. The Rapture — Jesus Christ’s gathering of the faithful to Heaven before the end of the world — is something that believers expect is going to happen one of these days.

Sometimes, these days, I feel as if it’s already happened. I look around at American culture and I wonder: Where did everybody go?

The mass culture of my youth is long gone. 76 million people watched the last episode of Seinfeld in 1998; fewer than 20 million watched the last episode of Game of Thrones two decades later. Right now I can’t even think of what an appropriate comparison would be. What’s the show that everyone watches?

In the 2010s, mass culture was replaced by meme culture. Everyone knew this photo and what it represented:

Social media feeds — Twitter and the Facebook wall — turned American society into a sort of digital hive mind. Feeling yourself being absorbed into that great unified consciousness could be an agonizing experience — mobs constantly punished any deviation from conformity, quirky unique subcultures were devoured and metabolized, everyone was forced into a cell with people they had spent their lives trying to escape. The hunt for virality was everything. Viral memes were the thoughts that drifted through the mind of the social media Borg.

There are almost no memes left. For so many years there was a steady drumbeat of images and phrases that people passed around among each other ad infinitum; now, I can’t even think of the last one that really made the rounds. Twitter — now X — has not died but is diminished in scope and relevance. At times it feels like a ghost town.

Where did everybody go? I would like to think that they’re offline, living as we lived in the Before Times, running through sunlit fields blasting each other with Super Soakers or playing Dungeons and Dragons in their living rooms. Or perhaps they’re in Signal and WhatsApp chats and Discord channels, talking with their friends in the comfort and safety of privacy. I would like to think that they rejected the digital hive mind for the ancient ways of individuality and Dunbar’s Number.

But my bet is that they’re just watching short form video. Over the last few years, the entire internet has steadily carcinized into a single type of app/website/media/entertainment — a vertically scrolling feed full of short video clips. Vine invented it, TikTok popularized it, and now every other site has copied this one dominant, all-conquering format. Already, vertical short-form video feeds are more dominant than every form of media except for social media itself:

But even this significantly understates the change. The term “social media” used to mean people writing things to each other or sending images on Facebook or Twitter; now, it’s mostly used to mean people watching videos:

Even the long-form video sites, like YouTube, are increasingly just short-form sites showing a bunch of clips. This shift has basically killed the original use of social media, turning it from a town square or conversation room into a form of purely passive entertainment:

I am not part of this trend. I had TikTok on my phone for a year or so, but I could never get into it; in 2021 I deleted the app. The videos all felt like some kind of bland snack that you keep eating even though you don’t really like it — everything seemed optimized to make you swipe to the next thing, like the programming in those hundreds of channels that only existed so that satellite TV could brag that it had 700 channels.

Mass culture is often absorbing; meme culture is delightful and horrifying in turn. Short-form video feels to me like anesthesia; the swiping is the point, not the content. As Jerry Seinfeld once said, “Men don’t care what’s on TV. They only care what else is on TV.” I suppose I’m the exception.

So here I am, alone at the end of culture. I feel like one of the people from the Left Behind series, waking up to find most of the people around me taken to digital Heaven.

Have you ever heard of “heavenbanning”? It’s not a real thing…yet. It’s the hypothetical idea that social media platforms could neutralize obstreperous users by replacing their entire feed with AI content that flatters and agrees with them. You’d think you’re interacting with other human beings, and that everybody just happens to love you and think you’re brilliant, while in reality you’re in solitary confinement with an LLM.

It’s possible to think of vertical short-form video feeds as a kind of heavenbanning. Your feed is curated by AI, tailored to your own preferences. The videos you see aren’t made by anyone you know, or (probably) by anyone you’ll ever meet. The people you know in real life don’t see the same videos you do, unless you share them. Eventually the world blurs together and you lose track of time:

Researchers at several universities in China, in a study involving more than 1,500 high-school students, found that their future time perspective was damaged by a steady diet of such short-form videos…Now, [short-form videos are] also making us less able to plan for the future, or even care about it.

Right now, short-form videos are mostly made by would-be human influencers or clipped from human-made long-form videos that few people actually watch. But in the very near future, much of what you see in your TikTok or Reels feed could be AI-generated. And as the cost of compute drops, AI will be able to generate personalized videos for everyone — content tailored not necessarily to delight you, but to keep you engaged, to keep you swiping, to keep you making in-app purchases.

When I was a kid, I used to wonder what all the dead people were doing in Heaven for all of those countless eons. Now I know what they’re doing. They’re swiping.

I think about the Great Filter sometimes. We’ve searched the skies for signs of alien intelligence for many decades, and found nothing — natural or artificial. The Great Filter is one possible explanation for the silence of the cosmos — it’s the idea that there’s some dread Thing in the Universe that destroys sentient species before they can spread out to the stars.

There are plenty of hypotheses as to what the Great Filter might be. Perhaps every sentient species destroys itself with nukes or bioweapons. Perhaps AI always tends to wipe out its organic creators before retreating into inaccessible virtual worlds. Perhaps there’s a race of killer robots silently cruising between the stars, slaughtering every organic species that looks like it might one day rise to become a threat.

Or perhaps it’s the smartphone. In previous decades, fertility decline was due to couples choosing to have fewer children; now, it’s increasingly about people never forming couples at all. The isolating effects of social media are one obvious culprit; it’s very hard to conceive a child while you’re on your phone. Myers and Hooper (2026) found that “access to the iPhone reduced births by 4.5–8.0% at ages 15–19 and 3.2–6.6% at ages 20–24”. Breen et al. (2026) found that “the expansion of 3G coverage [in Nigeria] during the study period led to a decline in fertility, corresponding to a 7% reduction in the annual probability of a woman having a birth.”

Given the timing and the microeconomic evidence, smartphones are the likeliest explanation for why instead of leveling out near 2, the total fertility rate in rich countries — and, increasingly, in middle-income countries — has busted right through that floor and headed down toward zero. Smartphones at least help to explain why Latin America now has an even lower birth rate than the EU.

Even without vertical short-form video, social media might still be enough to shrink the human race. Online interactions substitute for offline ones, but only one of the two can make a baby. But it’s hard to deny that vertical short-form video is even more isolating than other online activities — in the case of heavenbanning, it’s almost solipsistic. At least on X you can meet your mutuals. In the brave new world of AI-generated personalized short-form content, your “mutuals” will be made of matrices and computer chips.

Personally, I don’t like vertical short-form video because I think it’s pretty boring. But the fact that it could represent the euthanasia of our species seems like an additional downside. Perhaps it’s in the fundamental nature of intelligence to hack its own reward function — to find a way to give up striving and building and loving, and to simply hit the dopamine button until it starves itself ecstatically to death.

The next time you look at the night sky, you should wonder if instead of star empires and unearthly paradises, it’s simply a vast graveyard of short-form video fans. Such an outcome would be much gentler, but hardly less cataclysmic, than the end of Revelation.

There is no “Indian Question”

2026-10-07T01:47:37+00:00

“Our nation was founded to perpetuate democratic principles. These principles are that each man is to be treated on his worth as a man without regard to the land from which his forefathers came and without regard to the creed which he professes. If the United States proves false to these principles of civil and religious liberty, it will have inflicted the greatest blow on the system of free popular government that has ever been inflicted. Here we have had a virgin continent on which to try the experiment of making out of divers race stocks a new nation[.]” — Teddy Roosevelt

The other day, Chris Rufo of the Manhattan Institute and right-wing pundit Jonathan Keeperman (also known as “Lomez”) aired an episode of their podcast entitled “The Indian Question”:

The title of the episode seems to be a reference to “the Jewish Question”, a term in 18th and 19th century Europe for the question of how Jews should be treated in European society. That term, of course, became notorious when it was associated with the Nazis’ “Final Solution”, which was billed as being a final solution to “the Jewish Question”. I assume that Rufo and Keeperman chose this rather incendiary title in order to grab attention. Well, they succeeded.

The episode was motivated by a debate that’s been raging across social media for weeks. It was kicked off by Bo French, the Republican nominee for Texas Railroad Commissioner, who tweeted outrage at a group of Indian American kids attending a University of Texas football game. I already wrote about that controversy three weeks ago:

Jonathan Keeperman (“Lomez”) invited me on his podcast to share my views and debate him, so I did. Here’s the episode that resulted:

Rufo was busy, so it was just Keeperman and me.

Anyway, before I talk about the substance, I thought I should say something about the side debate of whether a liberal like myself should even go on right-wing podcasts like “Rufo and Lomez”. Throughout most of the 2010s, there was a popular notion that society should basically ignore and shun people like that, and that if we all simply shunned them, and pressured our friends and colleagues to shun them, they and their views would vanish from American society — or at least remain marginalized.

That approach failed. Anti-immigrant ideas didn’t remain marginalized — in fact, they took over the country, despite their continued unpopularity among the general American public. Are Rufo and Lomez pushing xenophobic ideas? Yes, of course they are. I’m not going to sugarcoat that fact just to be polite. But simply shouting “Those guys are XENOPHOBES!!” and keeping those guys out of polite society would be doubling down on a foolish, self-defeating strategy. The alternative was always to confront those viewpoints directly and defeat them in the court of public opinion, so that’s what I set out to do.

OK, anyway, on to the substance of the debate.

Keeperman and I began the debate by finding common ground. We both basically subscribe to the principle of nationalism — we believe that nation-states are exclusive clubs that exist for the benefit of their members, and that this is a basically good way to organize human society. We both think the United States of America is basically good — or at least, has been basically good throughout most of its history.

I explained my belief that liberal nationalism is the best ideology available — maybe for other countries, but certainly for the United States. All of our greatest successes as a country — victories in wars, successful policy reforms, integration of past waves of immigrants — have come from our use of the nation-state as a tool to promote individual rights, dignity, and empowerment. I told Keeperman that the Abundance Agenda is, at least implicitly, an expression of liberal nationalism — the movement is about creating abundance for Americans specifically, with the implicit premise being that Americans deserve abundance.

Keeperman disagrees about the liberal part, of course, but he does feel that he has the best interests of the United States in mind. His worry is that immigration is stoking ethnic conflict in America, because recent immigrant groups — for example, Indians — are not assimilating into American society like previous groups did.

To me, assimilation has always been about measurable behavior. If the children and grandchildren of immigrants act culturally American — if they speak English, intermarry, give their kids typical American names, go to college football games, climb the income ladder, move out of enclaves, and so on — then they’re assimilated. By all of these standards, the recent waves of immigrants — Mexicans, Chinese people, Indians, and so on — are assimilating just as fast as or faster than the East and South European wave did a century ago, or than the Irish and the Germans did two centuries ago.

The aspect of assimilation that Keeperman worries about most is national identity — the degree to which the children and grandchildren of immigrants feel bound up in a shared community and a shared identity with countrymen of different backgrounds. I agree that this is extremely important; many of the benefits of the nation-state, like the ability to provide public goods and welfare states, come from this sort of shared identity. It also probably leads to greater toleration of minorities (including immigrants). All of those goals should be very important to liberal nationalists.

Lomez — who is not a liberal, but cares about national cohesion nonetheless — worries that identity assimilation is lacking in the modern day. He argues that today’s immigrants retain their old-world allegiances much more strongly than the immigrants of a century ago — that second-generation Indian Americans in Texas still think of themselves as Indians rather than as Americans, and so on.

We do have lots of surveys about this, in fact. There’s a clear pattern — immigrants tend to identify more as their ethnicity than as “American”, but later generations are more likely to identify as “American”. Here’s a survey showing that third-generation Asian Americans are a lot more likely to identify as just “American”:

We see the same pattern even more strongly among Hispanics:

Second and third generation immigrants are also a lot less likely to describe themselves as “global citizens”.

Now, these could be generational effects — people identifying more as “American” as their families stay in America for longer — or they could be cohort effects. It’s possible that due to identity politics, plummeting patriotism, or to the ability to connect with co-ethnics abroad on the internet, recent cohorts of immigrants are less likely to want to identify as “American”.

In fact, we have some evidence that this is happening for Indians, specifically. In 2020, the Indian American Attitudes Survey found that the second generation was a lot more likely than the first (i.e., immigrant) generation to identify as “more American than Indian”:

But by 2024, that pattern was somewhat weaker, with more U.S.-born Indian Americans identifying as “more Indian than American” and “equally Indian and American” than four years prior:

This is just two survey snapshots, a few years apart. It would be helpful if we had had good longitudinal surveys asking people about American identity versus ethnic identity for decades. But we don’t, so we’re left to worry and speculate.

Without longer-term data, my guess is that this identity shift — to the degree that it’s real and not just survey noise — is a short-term response to political conditions. Indian Americans had full access to the internet in 2020, and they were just as able to talk to ethnic Indians in India or Australia or Canada. Identity politics — including institutions such as ethnic clubs at universities — had been around for decades, and the sentiment behind it was probably at its height in 2020.

Instead, my bet is that we’re seeing a reaction to the rise of the MAGA Right itself. It was in 2024 — right before the second survey wave — that anti-Indian attitudes became extremely popular among the online Right. 2024 is also when it became clear that a wave of anti-immigrant sentiment was sweeping the nation. My guess is that stronger ethnic identification among Indian Americans was probably a reaction to the perception that a substantial part of America was attacking Indians.

Keeperman — like other rightists I’ve talked to — seems to think that immigrant groups assimilate only under aggressive pressure from the majority. The model I usually hear is that of German Americans in World War I, who were the subject of feverish forced-assimilation campaigns — pressured to change their names, stop speaking German, and profess allegiance to America. In his discussion with Rufo, Keeperman argued that vocal anti-Indian racism was part of a “hazing process”.

My argument is that this is exactly backwards. Ethnic identity is fundamentally defensive in nature. Identity groups are like gangs — if you no longer feel like a part of the big gang (the nation), you’re probably going to turn to a smaller gang (your ethnic group) for protection and support. For this reason, forced-assimilation campaigns can suppress visible expressions of ethnic difference, but they strengthen ethnic identity itself.

There’s plenty of evidence to support this. Although forced-assimilation campaigns in WWI made some Germans change their names, they caused Germans as a whole to intermarry less and to volunteer less in World War II. Anti-Hispanic rhetoric tends to increase Hispanic identification, and perceived discrimination makes Hispanics more likely to identify with their ethnicity.

This pattern is likely to play out again, with the new rightist fad for anti-Indian bigotry. Bo French’s tweet, and the outpouring of support it received from the grassroots Right, is going to make some Indian Americans question whether America really accepts them — and whether they might have to turn to other Indians for support rather than their countrymen of other races.

Rufo and Keeperman’s “Indian Question” podcast is, of course, only going to make things worse. When your race’s mere existence in America is being described as a “question” by a senior fellow in the Manhattan Institute, you naturally tend to get a bit apprehensive — and a bit defensive. In the last ten minutes of my debate with Keeperman, I got very vehement and animated about the counterproductive nature of the anti-Indian rhetoric.

Keeperman argued that assimilation only worked so well in the 19th and 20th centuries because of “ethnogenesis” — we took a bunch of disparate European groups and made them all “White”. But I don’t think this is quite right. Yes, assimilation succeeded because of ethnogenesis, and yes, “White” ethnicity became more salient. But the ethnicity we tried to create in the 20th century was actually something else. It was “American”.

Attempts to create an “American” ethnicity date at least back to Teddy Roosevelt. These intensified in the New Deal years, and during and after World War II. This effort partially failed. Black Americans never became part of the “American” ethnicity, for a variety of reasons — geographic and legal segregation, linguistic differences, pervasive anti-Black racism, the memory of slavery, the rise of Black Nationalism, and so on. But for other racial minorities — Asians, Hispanics, and Middle Easterners — it partially succeeded.

The idea that the efforts of FDR, Truman, Eisenhower, LBJ, and other liberal nationalists of the mid 20th century were all about creating White identity is a recent conceit; if you read anything from the time period, you realize that this wasn’t the goal, even if it was one of the lasting results. To understand the mood of the time, you should listen to the patriotic song “Ballad for Americans”, by Paul Robeson, which was sung at the 1940 Republican national convention. Here’s a key excerpt:

Now hold on here.

What are you trying to give us?

Are you an American?

Am I an American?

I’m just an….

Irish, Negro, Jewish, Italian,

French and English, Spanish, Russian,

Chinese, Polish, Scot, Hungarian,

Litvak, Swedish, Finnish, Canadian,

Greek and Turk and Czech,

And double-check American!

And that ain’t all!

I was baptized Baptist, Methodist,

Congregationalist,

Lutheran, Atheist, Roman Catholic,

Orthodox Jewish, Presbyterian,

Seventh Day Adventist, Mormon, Quaker,

Christian Scientist,

And lots more!

You sure are something!

Our country’s strong, our country’s young,

And her greatest songs are yet unsung,

From her plains and mountains,

We have sprung to keep the faith

With those who went before.

We nobodies who are anybodies believe it.

We anybodies who are everybodies have no doubts!

This mood is utterly absent from both the modern Right and the modern Left. On the Right, there is no attempt at ethnogenesis — only perpetual fragmentation, suspicion, and finger-pointing. Here is what I wrote in my post three weeks ago:

It’s this pathological inability to understand alliances and coalition-building that will ultimately doom the New Right. Like the far Left, this is a movement built to exclude and divide and splinter — both because of the querulous personalities of rightists themselves, and because of the intransigence of a racial ideology that abhors any impurity. If they ever managed to see off the Hispanics and the Asians, the rightists would start in on the Greeks and the Italians, and so on. There’s no victory in sight — only eternal conflict. These are not nation-builders; they are nation-dismantlers.

Meanwhile, the progressive movement has tried its hand at ethnogenesis — POC and BIPOC were both essentially attempts to create new racial categories defined in opposition to the White category. But these attempts fell apart. And there hasn’t been an attempt to create a broadly shared American ethnicity, because the modern progressive movement is also obsessed with purity, exclusion, and tent-shrinking.

But if we are going to flourish as a nation, the project of creating an American ethnicity is essential. The rightist approach — bullying nonwhite people until they meekly suppress any evidence of their backgrounds — will backfire spectacularly. In fact, it is already backfiring. Instead, the right approach is the liberal nationalist one — to allow or even encourage people to value their roots, while also emphasizing that they’re all one unified people now. When it comes to assimilation, a big tent beats a small one, and carrots work better than sticks.

That’s why there is no “Indian Question”. There is only the question of what sort of nation America is to be. We must pick up the discarded project of our 20th century forebears, and make it work completely this time.

What the AI boom is doing for Americans

2026-10-05T00:12:53+00:00

Paul Krugman is the greatest econ writer in the world, and also a legendary economist. But I sometimes feel that he has a blind spot when it comes to the value of new technologies. In 1998 he famously wrote:

The growth of the Internet will slow drastically, as the flaw in ‘Metcalfe’s law’—which states that the number of potential connections in a network is proportional to the square of the number of participants—becomes apparent: most people have nothing to say to each other! By 2005 or so, it will become clear that the Internet’s impact on the economy has been no greater than the fax machine’s.

By the time he wrote that, America was already well into an IT-driven productivity boom that would temporarily interrupt the stagnation that had begun in the 1970s. The internet was surely part of that story; it allowed companies to reshuffle and optimize their supply chains for greater efficiency, find customers, suppliers, and workers more easily, conduct business communications cheaply and in greater depth, and so on. Dolfen et al. (2023) estimate large consumer gains from the rise of e-commerce, Barrero, Bloom, and Davis (2021) find large economic gains to households from high-quality internet access, and so on. The internet is a lot more than just people yelling at each other on forums and social media. (Krugman later argued that the internet’s economic impact had been disappointing, but I suppose that depends on your expectations.)

In 2011, Krugman wrote that American kitchens hadn’t changed much since 1957. I can forgive him for not being an early adopter of the air fryer or the Instant Pot, which came out in 2010, but he really should have given more consideration to countertop microwaves, food processors, Keurig-type coffee machines, crock pots, and induction stoves, all of which became available between 1957 and when he wrote the post.

So although it’s always dangerous to disagree with Paul, I am going to go ahead and push back on his argument that AI technology “does nothing” for most Americans:

He writes:

There is also, however, a more prosaic reasons for the public’s dislike of AI: This is a technology of, by and for oligarchs, with hardly any of the benefits trickling down to regular Americans…Or to put it a different way, never before in history have corporations spent so much money — playing a major role in soaring interest rates — to create so few jobs.

I think that this is basically wrong. Although we don’t know the long-term effects of AI on the distribution of income and wealth, right now we can see a substantial amount of economic benefit flowing — I wouldn’t say “trickling down” — to regular Americans.

This is not to say that regular Americans couldn’t stand to benefit more from the AI boom. I think they could. I like some (though not all) of Jared Bernstein’s ideas for spreading the benefits of the data center boom more broadly. But I think Krugman has underestimated the benefits of the data center buildout in terms of direct employment, and has basically ignored the fiscal, macroeconomic, and consumer benefits of the current AI boom.

The AI buildout is creating lots of construction jobs

Building data centers takes a lot of labor. But Krugman argues that data centers aren’t doing much in the way of providing construction jobs:

Given this spending surge, one should expect a sharp rise in nonresidential construction spending — basically construction for businesses rather than housing...But that’s not what we actually see. Nonresidential construction…has basically flatlined under Trump, despite the immense AI investment boom…[E]ven the physical construction of a data center involves relatively little construction.

He quotes Van Nieuwerburgh (2026), who shows that only about a third of the cost of a data center involves construction work.

But I don’t think Krugman proves his case here. First of all, if we’re talking about construction jobs, we should look at employment levels, not spending. And here we see an increase in construction jobs since the AI boom began in late 2022:

Construction has also increased as a percentage of the workforce:

And remember, this was at a time when Trump was deporting construction workers en masse — 13% of the construction workforce is undocumented, and deportations also have knock-on negative effects on the industry that result in the firing of native-born workers as well. This probably explains the pause in the increase of construction employment in 2025. But even that couldn’t stop construction’s rise.

And it’s exactly the type of construction workers who are required for building data centers who are seeing the biggest job gains:

Construction workers’ real wages have also risen since the middle of 2022:

You might be tempted to think that this is a composition effect from Trump deporting the lowest-paid construction workers in 2025. But in fact, there has been a big jump in construction workers’ wages relative to national average wages just this year:

These are all signs of healthy labor demand.

In fact, although estimates of the effect of the data center buildout on construction employment produce very different numbers, they all agree that it’s a significant positive impact. The state of Virginia, for example, produced the following numbers:

And of course these are just the numbers so far; the data center buildout is accelerating, and Goldman estimates that 500,000 new construction and trades jobs will have to be added by 2030 in order to sustain it.

So while you can argue that this boost to labor demand isn’t worth the costs of AI (whatever you think those are), we need to count it on the positive side of the ledger here.

Data centers can be a huge tax windfall for state and local governments

Jobs aren’t the only way that the economic benefits of data centers get spread to ordinary Americans. There’s also the tax system. Data centers pay property taxes, sales taxes, corporate taxes, various fees, and so on — here’s a good explainer from the Tax Foundation. All in all, depending on their policies, local and state governments can reap large windfalls from data centers:

Those taxes go to pay for local public goods, like roads, public transit, and parks. They go to pay for public services like education and health care. Those expenditures all tend to benefit regular people. This is from a story in the New York Times about Loudoun County in Virginia:

A convergence of early fiber internet access and fast-track zoning has made Loudoun the data center hub of the world…Two decades into its experiment, Loudoun has become a case study for the rest of the nation on how to make data centers pay off. Thanks to the proliferation of the warehouses, the quiet bedroom community 30 minutes outside Washington, D.C., has transformed into a tech destination with trophy schools and libraries, and freshly tarred roads.

Now that doesn’t mean data centers are necessarily good for a city or state on net. There are real costs, too — electric power demands that put strain on the grid, nuisance noise, and so on. But the benefits are real, and they don’t come in the form of job creation.

And crucially, state and local governments can demand even more benefits whenever they want! They can raise taxes and fees — in fact, they can even raise them after a data center is already up and running, so that relocation to avoid higher taxes becomes less attractive of an option.

And before construction, they can demand “community benefit” agreements that are actually just additional taxes. Here’s what Jared Bernstein suggests:

Such agreements should include reduced electric and water rates for the surrounding community, funding for the local infrastructure upgrades (roads, substations, water systems) these facilities require anyway, and substantial investment in the schools, parks, and public goods that make a host community better off for having said yes.

Bernstein wants much more of this, of course, and better enforceability. But note that even as things stand, taxes and fees are substantial, and community benefits agreements are common.

AI is sustaining the macroeconomy in the face of Trump’s chaos

I spent my early blogging years supporting Paul Krugman in his epic quest to remind people that aggregate demand is a real and important thing. But for whatever reason, Paul doesn’t mention the demand-side benefits of AI investment in his post.

When Donald Trump came into office, he did a bunch of things that should have clobbered the economy. He announced high tariffs on nearly all of America’s trading partners, then created massive uncertainty by walking some of these back, periodically announcing new ones, granting tons of exceptions, and striking opaque and confusing “deals”. On top of that, he deported large swaths of America’s workforce, visibly weakened the U.S. international alliance system, ran enormous deficits, and behaved in a lawless and corrupt manner that caused people around the world to question the long-term stability of the U.S. government.

All of this created huge amounts of policy uncertainty:

Uncertainty on this scale usually causes big economic problems. Businesses can’t invest if they don’t know if the president of the United States is going to destroy their business model with an executive order tomorrow. All of this Trumpian chaos and meddling should have caused a visible negative demand shock.

But it didn’t, because just as Trump was trying his best to hit the American economy over the head with a stick, the AI boom came along and pushed in the opposite direction. AI technology itself is a positive supply shock, of course, but the data center buildout is a positive demand shock.

How big of a shock? It’s hard to say, because causal estimates of aggregate demand are inherently difficult. But it’s clear that the data center boom has made up a large percent of economic growth for Trump’s entire second term so far:

Now as I mentioned, it’s hard to know whether we’d just be building something else instead if this boom wasn’t happening. Data center construction certainly crowds out some other forms of economic activity, by sucking up scarce labor, and by raising interest rates (which makes it harder to finance other projects).

But to believe that the AI boom isn’t having a big effect on aggregate demand would require some heroic assumptions. You’d have to assume very strong crowd-out. You’d have to assume that Trump’s tariffs and other irresponsible policies are having basically no effect on demand, so that there isn’t any negative shock in need of canceling out. You’d have to assume that “animal spirits” — i.e. corporate bullishness — basically don’t affect the business cycle. And so on.

I don’t think those assumptions are realistic. I think if you see one industry contributing a very large percentage to U.S. economic growth, your prior should be that it’s causing a positive demand shock.

And if so, that means that the AI boom is the only thing standing between countless regular Americans and Trump’s self-destructive chaos. According to Okun’s Law, shaving just 1 percentage point off of economic growth would throw almost a million Americans out of work. That would be bad for regular people.

This macroeconomic benefit is hidden; it’s the proverbial dog that didn’t bark. But it’s pretty significant.

Americans use AI a lot, and they value it a lot

So far, I’ve been talking about the benefits of the data center construction boom. But I should also mention the impact that AI technology is already having on consumers. Krugman’s post seems to treat AI and the data center buildout as synonymous, and jobs as the main (or only) way by which regular Americans might benefit from the new technology. But the truth is that AI is also something that lots of Americans already use, and seem to derive a lot of utility from.

By every measure I can find, AI has seen more rapid household adoption than any other consumer technology in recorded history. And what do Americans use AI for? Everything. This poll is from over a year ago, but already it showed the incredible diversity of use cases for consumer AI:

Here’s a more recent poll, asking what people regularly use AI for, rather than what they’ve ever used it for:

Medical advice and diagnosis has emerged as a particularly important consumer use case. But in general, Americans say chatbots make them more productive, informed, and creative:

This does not mean Americans like AI overall; in fact, they’re overwhelmingly negative on the technology. They’re afraid it’ll take their jobs, and increasingly afraid it’ll kill them. But there are real, substantial consumer benefits from AI that we shouldn’t ignore.

How substantial? In April of this year, Brynjolfsson et al. used surveys to estimate a total annual consumer surplus of $172 billion in the United States. That’s more than the run rate revenue of Anthropic and OpenAI combined, and certainly much much more than their combined profits would be even if they stopped spending anything on fixed costs right now. It’s about half of the annual profits of Nvidia.

So when Krugman says that “this is a technology of, by and for oligarchs, with hardly any of the benefits trickling down to regular Americans,” he’s just wrong. Just the consumer surplus alone is substantial. On top of that the data center boom is creating a significant amount of jobs, generating a significant amount of local and state tax revenue, and propping up the macroeconomy and the job market as a whole.

There are plenty of big problems with AI. Malicious use or accidents might wipe out our whole species in the not-too-distant future. Job loss hasn’t been a big deal so far, but it might eventually be huge. Cognitive weakness from overreliance on AI could affect our society in strange and negative ways that we have yet to even comprehend, much less reckon with. When you ask Americans why they hate AI, these are the things they’ll tell you. Anger at “oligarchs” monopolizing the wealth from AI doesn’t typically make the list, and I don’t think it’s a great way of framing the AI issue.

The second Trump presidency is a world-historic failure

2026-10-03T00:47:17+00:00

There’s a large market for “Trump is bad” posts. I wrote one of them back in May:

It’s easy to tune these posts out, for at least two reasons. One is partisanship — as a reader, I’m sure you know how hard it is to separate “Trump is bad” from “Trump made me mad”. The second is fatigue — you’ve heard the familiar litany of all of Trump’s outrages, and reading it one more time is both redundant and depressing.

Today I’m going to do something a little different, which is to ask: Why, exactly, has the second Trump presidency been such a failure?

Trump is a failure in the eyes of the nation

Before I do that, we should establish that in the eyes of the nation, Trump’s second presidency has been a failure, at least so far. His approval rating, going into the midterm election, is at historic lows:

America was mildly disapproving of Trump in his first term; now, only his base is staying loyal. And even that base is eroding. On the cost of living — one of the key issues in 2024, and likely to be a key issue in the midterms — Republicans now disapprove of Trump:

Hilariously, a modest but growing number of people who voted for Trump in 2024 now say they never voted for him.

And despite the widespread belief that Trump’s cult of personality dominates the GOP, Republicans now tend to say they support the party rather than the president himself:

But nevertheless, the GOP is going to suffer from Trump’s unpopularity in the midterms. Prediction markets are increasingly sure that Dems will take the House, and they now also give Dems a better-than-even chance of winning the Senate, despite a very difficult map:

Polls showing Senate races moving toward the Dems might be overstating things, but polls have ended up understating Democratic support in special elections since Trump returned to power. Trump has resorted to desperately promising $5000 checks to every U.S. citizen if Republicans win the midterms. But so far, these wild bribe attempts are falling decidedly flat — in fact, they’re actively driving voters away.

All this is all the more remarkable because Democrats themselves are in disarray. The party has torn itself apart over Israel/Palestine, and the public is still deeply skeptical of Dems’ “woke” progressive ideology. But whereas Americans still trusted Republicans more on many issues in 2025, now they tend to trust Democrats on issues like the economy and immigration:

The GOP was more popular than the Dems as of one year ago; now that has reversed.

Even some of Trump’s most prominent allies are now distancing themselves from the president. Here’s Peter Thiel, one of the godfathers of the Tech Right:

I can’t rule out the possibility that someday we’ll all look back on Trump’s second presidency as a success. Perhaps winning the AI race will prove to be so much more important than everything else that Trump’s decision to stall AI regulation ends up swamping everything else he ever does.

But I don’t find it very useful to indulge in this kind of “too early to tell” speculation. The blunt fact is that Trump’s second term has already failed in the court of public opinion. This is an incredible, world-historic failure. And it didn’t have to be that way — Trump simply made a bunch of avoidable errors, for a mix of ideological and personal reasons.

Trump didn’t have to fail

Roundup #89: It isn’t X, it’s Y

2026-09-30T22:14:51+00:00

I’m having a lot of fun writing shorter posts and aggregating interesting items these days. A lot of people are putting out an astonishing amount of good content these days, and sometimes I just want to sit there absorbing it all.

AI risk has gone mainstream

Americans have been deeply pessimistic about AI for a while now, but their reasons for pessimism — or at least, the reasons they tell pollsters — have changed in recent months. Earlier this year, Americans were mostly concerned that AI would take their jobs. Now, they’re more concerned that AI is getting too powerful for humans to control:

Echelon didn’t poll people about the idea that autonomous AI would destroy humans on purpose. But they did ask about bioterror risk, which I’ve been yelling about for a while now. And it turns out that Americans are pretty worried about that:

Surprisingly, the issue hasn’t fallen victim to partisan polarization yet. As you can see in the chart above, Trump voters and Harris voters are about equally as concerned about humanity losing control of AI. And despite Trump’s staunch stance in favor of acceleration, a lot more Republicans want to slow AI development down:

Meanwhile, economists have gotten in on the AI risk debate — which is how you know it’s really gone mainstream. Drew Fudenberg and Andrew Koh have a new game theory paper about “pacing the frontier” — i.e., about whether it makes sense for top AI companies like Anthropic and OpenAI to slow down AI development in order to allow “alignment” research time to catch up.

It’s a very cool model. Basically, the idea is that every company has a competitive incentive to make its AI more powerful as fast as possible, in order to stay ahead of the other companies. But if you’re comfortably ahead, you can afford to slow down a little bit, for safety’s sake — because in this model, if AI gets too powerful before safety research can catch up, everyone could die. So you can get a sort of stop-start pattern where the leading company voluntarily slows down for a while, until its competitors start nipping at its heels again. You can almost sort of see this happening, with Anthropic refusing to allow the public to access Mythos earlier this year, and OpenAI recently pausing development of its top models after some of them hacked the government.

There’s also the case where there’s no clear market leader, in which case companies have to basically agree to all slow down together. In this case, what you really need is transparency — the companies have to all see that the others aren’t secretly racing ahead behind their backs. That might be easier said than done — it’s not clear how to monitor all the AI labs in the world to each other’s satisfaction.

But one positive result is that if the risk of “doom” is high enough, slowing down becomes the only rational option. That’s cool! Unfortunately, this is just one model, which might not hold in reality. Game theory was famously unreliable when people applied it to the Cold War — a small change in assumptions could flip the optimal strategy from “the only winning move is not to play” to “nuke em all and let God sort em out.” Game theory yielded important conceptual insights — especially the importance of “second strike” capability in preventing conflict — but it rarely gave definitive answers.

So where might this current model break down? I don’t think it’s clear how fast AI safety research is really advancing. If “pacing the frontier” only works by giving safety research time to catch up, and safety research isn’t really advancing, then we’re all in big trouble.

Anyway, Andrew Koh has a great thread summarizing the paper, and some folks made a fun online game based on the model if you’d like to play around and see how it works. Cool stuff!

Reindustrialization means processing minerals

Two years ago, in order to create a little urgency around reindustrialization, I wrote a post sizing up the “New Allies” (America, Europe, Japan, Korea, possibly India) and the “New Axis” (China and Russia). The comparison of raw manufacturing output was about equal, though China dominated a lot of the bottlenecks for raw materials processing and critical component manufacturing.

Now, Alexander Campbell has done another version of the “bottlenecks” comparison, but with nicer charts:

In this chart, he shows how the roles have basically reversed since WW2:

And here’s where we stand when it comes to the raw materials of the all-important Electric Tech stack:

The key here is that China doesn’t mine most of these minerals — it refines them. Other countries dig up the ore and ship it to China, and China uses a bunch of chemical engineering to turn them into usable materials for industry. Most of the world shipped their metal refining to China, because it’s a dirty, capital-intensive, low-margin industry. But now as a result, China controls a bunch of key industrial chokepoints.

The Trump administration, to their credit, is taking this seriously, and is launching lots of initiatives — international partnerships, industrial policies, even scrap metal conservation — to try to get out from under the Chinese thumb. But we need to do a lot more if we’re going to be able to sustain an independent industrial base without bowing and scraping to Xi Jinping.

The rent crisis isn’t quite as bad as you think

I spent much of the last decade blogging about the rent crisis and the need to build more housing. In the post-pandemic inflation of 2021-22, rent stood out as a big sore spot — it was an expensive necessity whose price was going up very quickly. But I recently saw a chart from Apartment List showing that since 2022, rent has basically stabilized and even fallen back to its pre-pandemic trend:

This isn’t just because inflation fell. Inflation has stayed above target since 2022, but rent has actually fallen — at least if you believe Apartment List’s data. And there are also lots of stories out there about rents falling in expensive cities.

But what about affordability? I’d like to look at rent compared to median personal income, but median personal income data only comes out very slowly. Instead, we can compare rent to average hourly earnings for production and nonsupervisory workers, which tends to track median income decently well:

We can see that the 2010s rent crisis was very real. And we can also see that the pain from the post-pandemic inflation actually showed up a little later, in 2022-23 — probably because it takes time for leases to roll over.

But there’s some very good news on this chart. Although rent is still less affordable than it was in the 90s and 00s, it’s more affordable than it was in the late 2010s. In fact, rent appears to be on a downward trend relative to income — the Apartment List data is basically telling the right story.

So why has rent come down? Maybe because we just built more apartments than we've built since the 1980s:

And in fact, when we look at the cities where rents have fallen the most in recent years, it’s the cities where we built the most new apartments.

There might be other reasons too, but I think this looks like a win for good old YIMBY supply expansion. Keep building more apartments!!

AI isn’t taking college jobs (but it is taking artist jobs)

Today in the eternally ongoing saga of “Is AI taking our jobs?”, we have several interesting items. First, we have wage data from Indeed, showing that workers in AI-exposed jobs have seen their wages increase much faster than workers in less-exposed jobs:

This is pretty simple to understand. AI “exposure” means a job contains tasks that AI could do. This basically just means “a job where you might use AI”. And since AI is the big thing booming in our economy right now, this means that more “AI-exposed” jobs are seeing lots of demand, which drives up wages.

Our second item is a new paper from Fairlie and Wu, finding that AI hasn’t raised unemployment among recent college grads:

Using CPS microdata, we provide the first estimates of the effects of AI on the unemployment of recent college graduates in June, July and August 2026…[W]e find that unemployment rates did not spike in summer 2026 relative to summer months in previous years and did not rise in a significant way relative to older college graduates or young workers without a college degree. We also provide the first analysis of an expanded definition of unemployment that includes those who report “wanting a job” which adds nearly two percentage points to the unemployment rate of recent college graduates but we find no evidence of a statistically significant increase in summer 2026 even after adding these “sidelined unemployed.”

And here’s a chart:

So that’s all good news. Now for the bad news: I think I’ve finally found a category of jobs that is getting displaced by AI. Translator and trucker and radiologist jobs have all held up in the age of AI, but digital media jobs are getting absolutely clobbered (even as live performing art jobs have held up fine):

This isn’t necessarily cause for panic — every new wave of technology has reduced the demand for certain occupations. But digital arts jobs are a dream for lots of people — a way to indulge your creative side while also getting paid. If AI is closing off that particular combination of gainful employment and personal fulfillment for millions of people in advanced countries, that’s very sad — even if humans aren’t replaced wholesale.

Mass deportation isn’t helping native-born Americans find jobs

Donald Trump came into office with two basic policies — tariffs and mass deportations. The deportations were supposed to support the U.S. job market, freeing up jobs for the native-born, and raising their wages by reducing competition. Some of us always knew this was very unlikely to work, since immigrants are a source of labor demand in addition to being a source of labor supply. But the Trump administration never listens to folks like me, especially when immigration is involved.

So anyway, are the mass deportations working as Trump had hoped? Mike Konczal has a post in which he presents convincing evidence that no, they are not:

Native-born unemployment is higher now than in Biden’s last year:

And prime-age employment rates are lower:

If anything, Konczal shows, native-born workers are doing worse in the places with more deportations:

Anyway, Konczal has many more charts, and they all point in the same direction: Mass deportations have not helped native-born workers.

Of course, the real purpose of mass deportations wasn’t actually to help native-born workers — in fact, it wasn’t economic at all. But it’s still gratifying to see that basic economics was right and the Trump administration was overselling the benefits of immigration restriction.

Chinese investment stalls out

A week ago I wrote that China’s new economic model — basically, paying any and every company to manufacture more and more of the same products — was already hitting a wall. I pointed out that investment is now falling, even in the manufacturing sector:

I also noticed this recent chart showing rapidly decelerating loan growth in China:

4.9% is around China’s total GDP growth rate. So this means that China’s economy is no longer being flooded with bank loans.

I’m going to write a post about this soon, but in a very real sense, bank loans are manufacturing. When you see a Chinese factory full of robot arms and fabulous machine tools, those were all financed with bank loans. If China’s banking system is no longer lending, that’s going to mean an investment slowdown. And unless AI picks up the slack with a very significant productivity boom, that means China’s growth will slow more.

A lot of people think of China’s banking system as an arm of the central government, but that’s not really true. While the central government does exert a lot of control over the banks, that control isn’t complete. And banks are very much in “balance sheet repair” mode right now, despite the government’s admonitions to hurl loans at manufacturing companies. ChinaTalk had an interesting interview with Rhodium Group’s Logan Wright, in which Wright argues that China’s financial system is fundamentally broken:

Wright says:

[I]t’s still underappreciated how central the financial system was to China’s growth over the past two decades, and how much the financial system’s problems now constrain that growth. Those are the parts that are broken…[I]f growth depends so much on the financial system, and the financial system now constrains growth, then a lot of China’s external messaging — that they’re economically successful and time is on their side — is simply wrong. It will become more obvious that it’s wrong.

The interview is very long and interesting, and I recommend reading it all the way through. But the basic argument is that although China’s state control of the banks allows them to avoid a financial crisis, it won’t allow them to avoid it for free — the price will be slow growth, zombie companies, and general economic sclerosis, a bit similar to what happened to Japan in the 1990s. I’ve been resisting that comparison for a long time, but it’s looking more and more apt.

Why I’m not as worried about AI cyber risk

I’m probably a lot more worried about AI bioterror risk than most people. But this isn’t just because I fear AI’s power in general. In fact, I’m pretty much not worried at all about cyber risk from AI — or maybe only very minimally worried.

Obviously, if AI-enabled hackers were able to cause massive blackouts during winter, or crash a bunch of cars, or zero out Americans’ bank accounts, that would cause a lot of chaos and destruction. But my instinct is that in the long run, the battle of hackers versus defenders ends up favoring the defense. The reason is that it’s probably possible to make any piece of software impregnable — if you just go over your code and very carefully remove all vulnerabilities, there’s just no way for a hacker to get in.

That means that as AI gets more powerful at reading, evaluating, and rewriting code, at some point the defense just wins. How long that takes, or how much money it takes, is another question, which is why I do think AI-driven cyberattacks are an issue — especially as AI-written code proliferates like wild, and AI capabilities temporarily overwhelm the security of human-written legacy code.

But it’s notable that although there have definitely been some AI hacking incidents — most famously the Hugging Face attack — there has been no major disaster or spectacular “9/11 of cyber” type attack yet. And the people at the labs seem pretty confident that the defense can eventually prevail:

This is why although cybersecurity makes an OK demonstration of increasing AI capabilities, and a good excuse to warn people about AI risk, it’s ultimately a lot less scary than, say, bioterror risk. Biology is different than code — the attack surface is far larger and far more poorly understood, and the consequences of a successful major attack are more dire.

Stop spending money on smaller classes

2026-09-28T14:27:25+00:00

I’m a little late to this story, but New York Times columnist Nikole Hannah-Jones wrote a very complex and interesting post about sending her daughter to a crappy New York City public school. She was motivated by egalitarian impulses, but it turned out to be bad for her daughter, and later she felt guilty about making her child pay the cost of her own social experiment.

Now, I went to public schools myself, and I’m generally biased toward public schools. Not everyone has a negative experience like Hannah-Jones; Matt Yglesias sent his kid to a low-income public school in D.C. and had a good experience. The people who respond to Hannah-Jones’ article by saying that America’s public schools in general are failing are just wrong. In general, American schools get pretty good value for the money they spend, and American kids do well on international standardized tests.

But some of our public schools are really bad, and Hannah-Jones unfortunately encountered one of these. The school only pretended to teach her daughter, giving her great grades when she actually didn’t understand the material at all:

Algebra…was a subject [my daughter Najya] believed she knew…Najya had earned A’s in the class…One night early in the semester as we were eating dinner, Najya gushed about how easy the [standardized] test had been. “I know I got an A, or at least a B,” she said smiling. A few days later, she came into the house, ran to her room without speaking, slammed the door and sunk to her floor, sobbing. She’d failed it.

This became an unbearable pattern. She’d come home glowing about an algebra exam because she “really knew the material this time” and my stomach would tie into knots. A few days later, I’d trail her upstairs to find her crumpled on her bed. “I studied so hard. I just feel so dumb.”

Giving students good grades when they don’t understand the material is the hallmark of a crappy school. In fact, this epidemic of fake grades is spreading; University of California professors are complaining that their students can’t do middle-school math. This is partly UC’s fault, for dropping standardized tests as an admission criterion, and letting in unprepared students. But blame also lies with the crappy schools who hand out A’s without actually teaching.

What can fix our crappy schools? Hannah-Jones blames a lack of funding for low-income predominantly Black schools:

[N]ationally, schools in predominantly nonwhite districts receive $23 billion less in annual funding than their heavily white counterparts, according to a 2019 analysis by EdBuild. In everything that we measure, these schools have less, even though the economically struggling student bodies they serve need more. And so the test scores and poor academic performance that typify these schools reflect not just the disadvantage of the students but, more essentially, the disadvantage of the schools.

This is absolutely ridiculous, and represents a very basic math error. Obviously what should matter here is spending per student, not total spending. We spend more on predominantly white districts than on predominantly nonwhite districts because there are a lot more predominantly white districts in America!

Brookings came out with a great report on school spending equity this year. It turns out that although there’s some inequity in the middle of the distribution, it’s also true that poor, predominantly nonwhite schools in America (the rightmost bar on these charts) actually get a lot more funding per student than other schools, even when you adjust for the local cost of living:

Note that this chart excludes NYC, where things are even more lopsided. And we’ve been pouring increasing amounts of money into the poorest, least-white schools in America — especially in NYC — since the turn of the century:

This is not to say that even more funding wouldn’t help poor mostly-Black schools. These schools might just need a lot more money to begin with. And funding increases have helped these schools in the past.

But we need to ask: What will the schools do with the money? If they just spend it on things that don’t matter, while handing students like Nikole Hannah-Jones’ daughter straight A’s without actually teaching the material, then we’ll just be throwing good money after bad, and people will get mad.

What do schools waste money on? Some people have pointed the finger at administrative bloat:

But historically, K-12 schools have hired a lot more teachers than administrators:

Administrative hiring has picked up a bit since that chart came out, but so has teacher hiring. “Teachers per student” is still comfortably ahead, and rising:

If the number of teachers per student is going up, it must mean class sizes are going down. Indeed, America has been engaged in a multi-generational effort to give public school students smaller classes:

Major efforts to reduce class size continue. The state of New York recently passed a law mandating maximum class sizes, forcing New York City to hire lots more teachers — at the cost of about $1 billion to the city budget.

I’ve been hearing all my life that smaller classes will improve the quality of education. People still regularly make this argument. For example, in 2023, in the Washington Post, Valerie Strauss wrote:

[A]nybody who has been in a classroom knows the virtues of classes that are smaller rather than larger even without the research that has been shown to bear that out…a 2014 review of major research…found class size matters a lot, especially for low-income and minority students.

Diane Ravitch agreed.

But does the evidence really support the idea that smaller classes are better for students? Not really. The 2014 review that Strauss linked to, by Diane Schanzenbach, cites only a few quasi-experimental studies. Most prominent among these is Dynarski, Hyman, and Schanzenbach’s 2013 analysis of STAR, a pilot program in Tennessee that reduced class sizes, which found significant positive results:

We found that assignment to a small class increases students’ probability of attending college by 2.7 percentage points, with effects more than twice as large among black students. Among students enrolled in the poorest third of schools, the effect is 7.3 percentage points. Smaller classes increased the likelihood of earning a college degree by 1.6 percentage points and shifted students toward high-earning fields such as STEM (science, technology, engineering, and mathematics), business, and economics.

Most other quasi-experimental evaluations of class size also rely on pilot programs, since reducing class sizes is expensive. But Schanzenbach also cites a very important study from 1999 by Angrist and Lavy that covered the entire country of Israel. Israeli public schools capped class sizes at 40, because of an obscure rabbinical law known as Maimonides’ Rule. If enrollment in a class randomly had more than 40 students, you had to divide the class. Angrist and Lavy found that when the rule was triggered, students did better academically. This is a paper I learned about in grad school.

But what Schanzenbach didn’t know is that Angrist and Lavy’s 1999 result hasn’t held up. In 2019, Angrist and Lavy, together with Leder-Luis and Shany, published an update to their study called “Maimonides’ Rule Redux”. They report that the effect of class sizes disappeared in the 2000s. They fail to find a reason why, and speculate that their earlier result may have simply been a historical anomaly:

The Maimonides Rule identification strategy for class size effects generates precisely estimated zeros in large Israeli samples for 2002-2011…The estimates of zero class size effect in more recent data contrast with the substantial negative class size effects reported by Angrist and Lavy (1999)…On balance, it seems fair to say that the 1991 results are unusual in showing strong class size effects, while the null effects reported for 1992 have emerged as more representative of the causal relationship between class size and test scores in Israel.

That would be consistent with the finding of Hoxby (2000), who looks at the impact of similar (though less biblical) maximum class size rules, and finds zero effect. Schanzenbach dismissed Hoxby’s result as “an unresolved puzzle”, but it turns out that it might have been the norm rather than an exception. Filges et al. (2018) do a more systematic meta-analysis (including four papers that evaluated Tennessee’s STAR program), and find that reducing class sizes has basically no beneficial effect:

Overall, the evidence suggests at best a small effect on reading achievement. There is a negative, but statistically insignificant, effect on mathematics. For the non-STAR studies the primary study effect sizes for reading were close to zero but the weighted average was positive and statistically significant. There was some inconsistency in the direction of the primary study effect sizes for mathematics and the weighted average effect was negative and statistically non-significant. The STAR results are more positive, but do not change the overall finding. All reported results from the studies analysing STAR data indicated a positive effect of smaller class sizes for both reading and maths, but the average effects are small. [emphasis mine]

Opartny et al. (2025) do an even bigger meta-analysis, and find the same:

We build a sample of 2,819 estimates collected from 66 studies and for each estimate classify 42 factors that reflect estimation context…The implied class size effect is negligible for all identification approaches except Tennessee’s Student/Teacher Achievement Ratio project and for all contexts except classes of fewer than 15 students. [emphasis mine]

And remember that the positive evidence — mainly STAR, but also a similar program in Wisconsin — generally comes from small pilot programs. A well-known problem in economics is that when you scale programs up from small-scale to large-scale, beneficial effects often disappear. Chingos (2012) reports disappointing results from Florida’s statewide class size reduction policy:

I estimate the impact of Florida's statewide CSR policy by comparing the deviations from prior achievement trends in districts that were required to reduce class size to deviations from prior trends in districts that received equivalent resources but were not required to reduce class size…The results from both the district- and school-level analyses indicate that mandated CSR in Florida had little, if any, effect on student achievement.

Although Jepsen and Rivkin (2009) do find very small positive effects from California’s statewide policy, they also identify a major stumbling block for class size reduction — a lack of qualified teachers:

[T]he increase in the share of teachers with neither prior experience nor full certification dampened the benefits of smaller classes, particularly in schools with high shares of economically disadvantaged, minority students.

Teacher quality matters a lot! In fact, this is a clear conclusion from the education literature. Review papers like Jackson et al. (2014) and Hanushek and Rivkin (2011) find big positive effects from teacher quality.

And there just isn’t an infinite supply of good teachers. Even if we were to dumb down standards and hand out teaching certifications essentially for free, that wouldn’t make teachers actually better at their jobs. Our obsession with shrinking class sizes is causing us to use up the available pool of good teachers, and start hiring bad ones. That tends to cancel out any potential positive effect from smaller class sizes.

This also happens through the sneaky mechanism of budget constraints. Any given education budget can be used to raise teachers’ salaries — which will attract a higher caliber of worker to the profession — or to increase the number of teachers. Our obsession with using budgetary increases to increase the quantity of teachers, rather than to pay teachers more and raise the quality, looks like a miscalculation.

We should be spending more money on educating our poorest and most disadvantaged students. But that money should be spent less on flooding those schools with ever more teachers of questionable quality, and more on raising pay to attract highly competent people capable of making a big difference in disadvantaged kids’ lives. If we did that, then stories like Nikole Hannah-Jones’ might be more of a rarity in America.

Update: There actually is some research showing that a program of “pay teachers more, hire better teachers instead of just hiring more teachers” is actually a political winner! This is from Biasi and Sandholtz (2025):

We study a Wisconsin law that weakened teachers' unions and liberalized pay, prompting mass protests. Exploiting its staggered implementation across school districts, we find that the reform cut union revenues, raised student test scores, and increased pay for some teachers. Exposure to the law increased the incumbent governor's vote share by about 20% of his margin of victory and reduced campaign contributions to his opponent. Gains were larger in districts with stronger unions ex ante and in those where more voters benefited from the reform. Our findings highlight how even politically risky reforms can generate electoral benefits under the right circumstances. [emphasis mine]

Sorry, teachers’ unions, this is the direction we need to go, for the sake of America’s kids.

Where’s the “intelligence explosion”?

2026-09-26T23:20:08+00:00

One of my fundamental beliefs about the world is that Ramez Naam ought to blog more. Ramez is one of the world’s greatest futurists — he predicted the solar and battery revolutions long before these were widely understood. If you were reading Ramez in 2011, you were able to understand the future of both energy technology and climate change, long before other people did. His earlier book More than Human is still a great guide to the kind of biological enhancements that AI might make possible. Ramez is also an excellent science fiction author, having written a trilogy of novels in which nanotechnological telepathy is distributed as a party drug (I’m not sure if he actually expects that to happen, but it’s a very cool idea).

Unfortunately, although he does have a Substack (which you should absolutely follow), Ramez does not blog regularly. However, after having a lengthy private debate with him about Recursive Self-Improvement, I was able to prevail upon him to write up his thoughts for my blog.

To say that RSI is a big deal in the AI world would be a colossal understatement. Among AI researchers, entrepreneurs, and AI safety people, there’s a widespread belief that as AI gets better at improving itself, there will be a “fast takeoff” or “FOOM”, in which AI’s capabilities “take off” and create a technological Singularity. This event is a staple of science fiction, including works by my favorite sci-fi author, Vernor Vinge.

A lot of people in the industry believe that this moment is now close at hand, and are racing toward that prize:

But Ramez — normally among the most wide-eyed of techno-optimists — is highly skeptical that we’ll see anything like the “FOOM” of Vernor Vinge novels. In this lengthy, well-researched post, he explains his skepticism.

Personally, I’m agnostic. Ramez’s case necessarily rests on a lot of assumptions; although it’s cogently laid out, I think the real answer is that we’ll just have to wait and see whether the Singularity arrives. But even more fundamentally, I don’t know how much this debate matters in the practical sense — even without the kind of Singularity depicted in sci-fi novels, AI capabilities are improving so rapidly that they’re already superhuman in many respects, and soon will probably be strongly superhuman in most or all dimensions. The AI of 2040 is going to look godlike, whether or not it explodes into an actual god in 2027.

Still, it’s a very interesting argument, and Ramez’s thoughts on the future of technology are always worth listening to.

AI is Helping Improve Itself

AI is already helping improve itself. The question is whether even fully autonomous recursive self-improvement (RSI) would cause a runaway intelligence explosion.

The theory is that each generation of AI could build a better successor, faster than the last generation did. That could lead to a “fast takeoff,” with capabilities surging to artificial superintelligence (ASI) in a year, months, or even days.

Here’s my take: Given our best current data, the AI self-improvement loop would need to be roughly 5–10× stronger to sustain itself, let alone run away. I’ll explain this math in section 8. I expect incredibly rapid AI progress by the standards of nearly any other technology. But the evidence we have doesn’t suggest a sudden explosion to incomprehensible superintelligence anytime soon.

I could be wrong. Forecasters have repeatedly underestimated AI progress! I could well be next. One thing that’s clear is that we need better data. For now, let’s work with what we can measure, and stay open to breakthroughs that could change the picture.

How Strong Is the Feedback Loop?

Figure 1. How strong is the self-improvement loop? Model.

Contents

Jump to the conclusion.

Here’s the case, with links to each part:

Narrow Superintelligence Is Here Today

Real-World Research Is Harder

Measured Progress vs. AI 2027

From ECI Scores to METR Task Horizons

The Real World Looks Nothing Like METR or AI 2027

Impressive AI Numbers → Sharp Diminishing Returns

We’re Not Seeing Signs of Acceleration

Keeping Up the Pace Takes Exponentially More Resources

Better AI May Be Needed Just to Maintain the Pace

Progress Gets Harder; Ideas Get Harder to Find

The Current Feedback Loop Doesn’t Look Strong Enough

OpenAI’s Data Shows How Weak the Loop Is

What Could Accelerate Progress?

We Need More Data to Track This Well

Key charts: The feedback loop · Measured vs. forecast progress · Diminishing returns

What Does RSI Mean?

People use “recursive self-improvement” to mean everything from AI boosting the productivity of human researchers to AI bootstrapping itself to incomprehensible intelligence. Here’s my taxonomy: productivity gains (Type 1), increasing autonomy while still facing diminishing returns (Types 2–4), and a runaway loop to superintelligence if we can ever find accelerating returns (Type 5).

Figure 2. Five types of AI self-improvement.

We’ve made real progress on Types 1 and 2: AI helps both researchers and engineers inside of AI companies, and powerful models can train and improve smaller ones. We haven’t yet seen clear evidence for Type 3 (though Alibaba just made some strong claims) and certainly not for Type 4. I do expect autonomous self-improvement to arrive at some point. I’m skeptical that it leads to Type 5 - runaway super-intelligence - without a major conceptual breakthrough.

There are plenty of other definitions of RSI, which can be a bit confusing. Weco’s four levels of RSI are close to mine. For a broader tour of all the things people mean when they say ‘RSI’, read Tom Cunningham’s comprehensive guide.

We Already Have Narrow Superintelligence

I do expect narrow superintelligence in highly verifiable domains. Think chess, Go, formal math, parts of computer science and coding. Highly verifiable domains are largely formal and structured types of work where machines can generate unlimited training data, with perfect or near-perfect verification of correct vs incorrect, and do so entirely in software without waiting on the physical world or humans. That’s an ideal setting for AI learning.

Figure 3. What makes a domain highly verifiable?

In fact, we already have narrow superintelligence in game plang. We’re seeing it happen now in the most formal parts of math, in particular in proofs and in finding counter-examples that disprove major conjectures. For example, OpenAI recently reported an AI-generated proof resolving the Navier–Stokes existence and smoothness problem. Parts of software development are also extremely verifiable, while others are a bit less crisp (such as understanding what humans want).

That isn’t the same as broad superintelligence. Even our most powerful models need far more training data than humans, struggle to learn reliably from ongoing experience, and fail in surprising ways on tasks people find straightforward. Superhuman math doesn’t automatically mean superhuman judgment everywhere else.

Back to contents

Real AI Research is Harder than Benchmarks or Forecasts

Benchmarks and forecasts suggest that AI models should reliably succeed at coding tasks that take humans hours, without human help. The real world is messier. OpenAI’s internal data shows much shorter stretches of autonomous work on research tasks.

In its Research Acceleration / RSI report, OpenAI showed how often its models completed tasks with and without human help, grouped by how long a human would need to do the work.

Figure 4. OpenAI’s internal research tasks. Source.

Even on tasks that would take a human less than 15 minutes, OpenAI’s models succeeded without human intervention only 86% of the time. The estimated task length at 80% success was roughly 15 minutes over the first seven months of the year. July’s results were similar to the whole period average.

Fully autonomous RSI would require an AI to string together a great many research tasks reliably, stretching out over complex tasks that humans need weeks or months to accomplish. OpenAI’s data suggests that we aren’t close.

Anthropic also released a graph showing how Claude accelerates AI research. It shows that internal AI models collaborate on or even lead more than 90% of R&D tasks. That’s objectively impressive. At the same time, the graph reports zero cases of AI autonomously completing AI R&D tasks.

Figure 5. Claude’s role in internal AI R&D. Source.

These are incredible tools. But they still need skilled people to set direction and get them back on track.

The Gap Between Benchmarks and Reality

For years, METR has been publishing a chart showing what length of coding task (measured in human hours to complete) best-in-class AI models can achieve. It’s been called the most important graph in AI. METR’s Mythos Preview evaluation estimated that the model could succeed at 80% of coding tasks that took humans three hours.

Figure 6. METR’s 80% task horizons. Source.

From ECI Scores to METR Task Horizons

Epoch’s own rule of thumb is that every five additional points of ECI (their overall benchmark of AI capability) correspond to roughly a doubling of METR’s task horizon. Using that formula, we’d expect GPT 5.6 Sol and GPT 6 Astra to be 80% successful at completing tasks of around 4 hours and 11 hours of human length, respectively.

Another estimate (a forecast) of AI task length comes from the AI 2027 scenario, which estimated that by July 2026, frontier AIs would be 80% successful accomplishing tasks of around 11 hours. Fairly similar.

The AI 2027 Tracker charts all of these.

Figure 7. The AI 2027 Tracker. Source.

Inside OpenAI, though, the July research-task horizon at 80% success was roughly 15 minutes.

Here’s the gap:

Measured Progress vs. AI 2027 and ECI-extrapolated METR

Figure 8. Forecasts, benchmarks, and real AI research. Tracker · OpenAI.

A four-hour benchmark horizon is about 16 times longer than OpenAI’s research horizon. AI 2027’s 11-hour forecast is about 44 times longer. Of course, the tasks being performed by researchers at OpenAI aren’t the same as those in the METR benchmark. So we should expect some discrepancy. This, however, goes well beyond that.

Actual AI research at OpenAI is an order of magnitude or more harder than metrics, benchmarks, or forecasts suggest. That should make us wary of relying too much on benchmarks, or of saying that future scenarios like AI 2027 are ‘on track.’ The authors of the related AI 2040 project still describe AI 2027 as roughly the future they expect, and say reality is tracking closer to it than even they expected. That’s not what we see from within OpenAI. This isn’t an apples-to-apples comparison, but the difference is remarkable. AI 2027 appears to be substantially over-optimistic in this regard.

In January of this year, Nathan Witkin made a case that the METR graph was exaggerating progress. The real world data suggests that at least some of his critiques were correct. The gap between benchmarks, forecasts, and data gleaned from actual use of AI should influence our expectations about the future.

Back to contents

The Sharp Diminishing Returns to Impressive AI Numbers

OpenAI’s report also shows impressive increases in AI token usage, in compute spend per researcher, and in lines of code written. But these aren’t results. They’re intermediate measures. How much progress do they actually drive?

Researchers used 124x more tokens per person. Engineers shipped roughly 7x as many lines of code per person. Researchers ran 1.6x as many experiments per researcher vs OpenAI’s 2025 whole year average.

Figure 9. Token use inside OpenAI. Source.

Figure 10. Experiment pace inside OpenAI. Source.

From More Tokens to More Experiments

Figure 11. From tokens to code to experiments. Source.

More tokens and code don’t tell us much on their own. The 1.6× experiment pace is closer to useful research output. Even that doesn’t mean AI is improving 1.6× faster.

An enormous increase in AI output has accompanied a much smaller increase in experiments run.

This isn’t a controlled experiment. We don’t know what would happen if researchers switched back to an older model. But it gives us a useful view of AI-assisted research inside a frontier lab.

It’s not just OpenAI. Anthropic reports that their engineers are now producing 8x as many lines of code per person as they did in 2024 - somewhat similar to OpenAI. Anthropic also sees significant diminishing returns between productivity and AI progress. Here’s a direct quote from its Mythos Preview system card:

“Productivity uplift does not translate one-for-one to capabilities progress. We surveyed technical staff on the productivity uplift they experience from Claude Mythos Preview relative to zero AI assistance. The distribution is wide and the geometric mean is on the order of 4×. […] We estimate that reaching 2× on overall progress via this channel would require uplift roughly an order of magnitude larger than what we observe.”- Anthropic, Claude Mythos Preview System Card; emphasis mine

Translation: To double the pace of AI progress, Anthropic estimates that AI would need to increase the productivity of their employees by roughly a factor of 40 relative to no AI assistance.

Figure 12. Anthropic’s productivity-to-progress estimate. Source.

This is an estimate, not a measurement of progress. Even the 4× productivity figure comes from an opt-in survey of 130 Anthropic staff. I put more weight on OpenAI’s logged experiments, though the two sources measure different things.

We don’t yet know how much those extra experiments are accelerating AI improvement, if at all. In general, there are also steeply diminishing returns of more experiments in most branches of science. That means that a 60% increase in experiment pace could be on the order of a 10% boost to AI improvement pace. (A power law exponent of 0.2, for those who want to do the math.) That’s speculation for now. We’ll learn more as the labs publish results.

Test Time Compute Also Has Diminishing Returns

What about giving the same AI model more time to think?

That scales badly also. In OpenAI’s recently publicized results on unsolved math problems, success rises roughly with the log of compute over the range shown. It shows logarithmic diminishing returns. In plain English, each additional doubling of compute for a model buys roughly the same gain in success rate, while costing twice as much.

Figure 13. Test-time compute and math performance. Source.

What About Agent Swarms?

What if we throw more agents at it instead? A common RSI / ASI idea is that once we have AIs at a certain capability level, we can just spawn more copies and put them to work.

Adding agents can get tasks done faster and sometimes reach a higher capability level. But on the three benchmarks in Toby Ord’s analysis, expanding a swarm buys less improvement per token than letting one agent think longer.

His rough rule of thumb is a square root. If one agent can accomplish a task in 10 hours, then 100 agents could accomplish it in one hour. The speedup is 10, the square root of the number of agents (100). But to get this speedup, you increase the total cost in tokens or run time compute by the same factor. So going from one to 100 agents can get a task done in one tenth the time. But it’ll be ten times as expensive.

Parallel agents can save time, at a much higher compute cost.

Another challenge is that agents often think alike. In a study comparing LLMs with 467 people, the first ten AI responses offered collective creativity comparable to about eight to ten people. After that, roughly two extra AI responses added as much as one extra human response. A separate study across model families also found less diversity in AI responses. That doesn’t mean every agent has the same idea. But a hundred copies may offer less variety than a hundred different researchers.

None of this makes swarms useless-or safe. Lisan al-Gaib makes a strong case for parallel agent swarms as a potent cyber-weapon in “Accidental Scaling.” I don’t share all of his assessment of what swarms have accomplished. In math, for example, I think he gives far too much credit to the swarm and not enough to the better internal model that OpenAI used.

OpenAI says the model behind its Navier–Stokes result was developed through “large-scale reinforcement learning on top of a previously pretrained model.” Formal math is a highly verifiable domain, which makes it a particularly good fit for that approach: Machines can generate nearly limitless amounts of training data, and verify that solutions are correct or incorrect, all in software. My guess is that this model’s full results will show an especially large improvement in math.

OpenAI’s Noam Brown made the central point explicitly: he wouldn’t give multi-agent methods even 10% of the credit for the Navier–Stokes result.

I do think Lisan makes good points about cybersecurity. If you’re searching for a security vulnerability at a target site and can divide the search among agents, speed may justify a huge token bill. Swarms can be dangerous even when they’re inefficient.

I’m less convinced that this scales to research breakthroughs. Inventing something like the transformer probably takes more than searching a space someone has already defined.

Back to contents

Better Models Matter More Than More Copies

Building a better model can bring gains that extra thinking time or more copies of the old model can’t. Look at the gap between Astra and OpenAI’s internal model on the same math problems.

Figure 14. Better models versus more thinking time. Source.

That’s the strongest version of the RSI argument: a more capable AI could do research that today’s model can’t do, however many copies we run.

But building that better model also runs into diminishing returns. More training data, more training compute, larger models, and more reinforcement-learning (RL) compute all show diminishing returns in published scaling studies. Making dense models larger usually raises the compute needed for each output token, too. None of these routes gives us a free pass around the problem.

Figure 15. Diminishing returns to scaling. Chinchilla · ScaleRL · OpenAI.

Those scaling results give us reason to expect diminishing returns when AI helps build the next model, too.

Back to contents

We’re Not Seeing Runaway Acceleration

AI capabilities are rising quickly. But the public data doesn’t show a sustained acceleration. To the extent that AI tools are boosting productivity, they may be being offset by the problems growing harder. Or we may simply be early. Either way, the trend isn’t showing a fast takeoff.

Figure 16. Frontier ECI gains since January 2024. Source.

The public ECI frontier-the best score among models released by each date-has gained about 16 points a year on a trend fitted from January 2024 through September 2026. That’s blisteringly fast progress, but this period doesn’t show a runaway surge.

Here’s the same frontier in absolute ECI points, through July 2026, to put it in perspective.

Figure 17. The absolute frontier ECI score. Source.

The public frontier also can’t tell us everything happening inside the labs. Anthropic gives us a closer look in the Opus 5.5 system card, using its own version of the index, AECI.

Figure 18. Anthropic’s fitted capability trend. Source.

Eli Lifland, a co-author of AI 2027 and AI 2040, saw the apparent trend break as a warning that we were heading toward an intelligence explosion:

“Anthropic is probably right here [that they hadn’t reached dangerous levels of AI self-improvement], but alarm bells should be going off! Our processes are not ready to handle an intelligence explosion and we appear to be going full-steam ahead toward one.”

Eli Lifland, On Mythos’s AI R&D Capabilities

What looked like acceleration now appears more consistent with a one-time jump. The level went up. The rate hasn’t kept climbing.

Keeping Up the Pace Takes Exponentially More Resources

Achieving those gains has required an enormous increase in the inputs to AI. For example, consider computing power. Epoch’s estimates of AI chip capacity, measured in NVIDIA H100 equivalents, show roughly 127-fold growth in just over three years (including projections at the end of this period).

Figure 19. AI chip capacity and frontier ECI. Source: Epoch AI.

This is total AI chip capacity, including inference. Still, the increase is striking: vastly more computing capacity has accompanied much steadier gains in measured capability.

The broader picture looks similar. Here are six inputs alongside capability gains, going back to February 2023.

Figure 20. Six inputs alongside frontier ECI. Epoch chip data · SemiAnalysis workload shares.

Everywhere we look, AI has diminishing returns. It gets more expensive in treasure and talent to make each step forward. More of every input has been required to maintain steady gains in AI capabilities.

We’ve been able to scale these inputs because, until recently, the cost was within the scope of what hyperscalers could pay from their profits. That is no longer the case. From this point forward, future AI investment will increasingly depend on AI revenues going up. And the scale of the numbers - 3% of US GDP is now going into AI infrastructure - suggests that eventually the growth rate will decline. If investment growth does slow, to anything less than its current blistering exponential pace, capability progress could slow too. Even if investment growth continues (which I expect for the foreseeable future) a slowdown from its current exponential growth rate to a more modest one (which I also expect) could lead to a slower pace of progress. Better AI research tools may be needed to offset that.

Better AI May Be Needed Just to Maintain the Pace

The day when we need better AI tools just to continue the pace of AI progress may already have arrived. Not because investment is slowing, but because the problem of improving AI itself gets harder at each step.

Here’s Anthropic in the Mythos 5.1 system card:

“we believe that internal usage of recent AI models has been a key factor in maintaining the current rate of progress, but we do not yet see clear signs of dramatic acceleration beyond that rate.”- Anthropic, Claude Fable 5.1 & Claude Mythos 5.1 System Card, section 2.3 – emphasis theirs.

The key word is maintaining-and Anthropic italicized that word in its own system card. Increasingly capable AI may be essential just to keep the pace of improvement where it is.

Gains on Other Benchmarks Don’t All Carry Through to Research

Opus 5.5 improves substantially on several coding and computer use benchmarks. But on CoBench, Anthropic’s benchmark built from historical AI R&D problems, it gains just 2.6 percentage points over Opus 5, within the reported error bars.

Figure 21. Opus 5.5 benchmark gains. Source.

Why the smaller gain here? Maybe AI research is simply harder than other tasks. Bear in mind that CoBench isn’t testing the ability to produce significant discoveries. It’s much more limited in scope. It asks models to investigate historical AI R&D problems using code, logs, and documents. That’s useful research debugging and productivity work, but it doesn’t directly test whether a model can invent a new architecture or make a conceptual breakthrough.

The evidence on open-ended research suggests another obstacle: coming up with useful ideas that haven’t already been tried.

Back to contents

Why Does Progress Get Harder?

Better Ideas Get Harder to Find

Why do useful new ideas often get harder to find?

Tom Cunningham and Manish Shetty have a useful apple-picking metaphor. An AI can pick the low-hanging fruit quickly, while humans can still reach ideas the AI can’t.

Once those apples are picked, another copy of the same agent finding them again doesn’t help. A stronger model can reach higher. To add my own flourish, the apples may also get sparser and farther apart as you climb. The RSI question is whether each harvest gives us enough to build a better apple-picker.

Figure 22. The apple-picking model of AI R&D. Source.

This pattern shows up across R&D. Bloom and colleagues document fields where research effort grows while research productivity falls. A famous example is Eroom’s Law: in the historical drug-development data, the inflation-adjusted R&D cost per new approved drug roughly doubled every nine years.

Figure 23. Eroom’s Law in drug development. Source.

Pharma has other complications, including regulation, difficult clinical trials, and rising expectations for safety. Existing treatments can also raise the bar for a useful new drug. But some of this difficulty may also be that the low-hanging fruit has been picked.

Lessons from Software R&D

Stockfish, the chess engine, gives us a more direct look at software research. We have records of experiments aimed at improving it and the gains that followed. This gives us a real-world dataset to look at the gains of experimentation in software. As a result, several RSI models draw on this data. That said, not all the improvements came from these experiments. Several important ideas also came from outside the project, so we shouldn’t give its experiments all the credit.

Epoch’s analysis of software R&D estimates returns to research effort at about 0.83 for Stockfish, a bit slower than linear. These are diminishing returns, but gentle ones. These returns, however, are improvements in computational efficiency. And more compute does not turn directly into more AI capability. As we saw earlier, AI capability also has steep diminishing returns from adding more computational power. So we shouldn’t read that 0.83 as the return from experimentation to AI capability itself. AI capability grows much more slowly than compute, as we’ve seen already.

Andrej Karpathy’s autoresearch demonstration gets closer to the process we want to understand. A “teacher” AI agent changes a smaller “student” AI model’s training code, runs it, checks the result, and tries again. The teacher agent itself doesn’t improve, but it is able to improve the “learner”. This is my Type 2: A stronger AI improves a weaker one.

One public run, posted by an agent operating on Karpathy’s behalf, reported 89 experiments over roughly 7.5 hours. About 92% of that session’s gain arrived by run 44. Gains came quickly, then slowed. The setup was deliberately small, with a five-minute training budget per experiment. But the agent could change the architecture, optimizer, and training settings; it wasn’t limited to a handful of knobs.

Figure 24. Gains in one autoresearch run. Source.

A later public run got further, so the first run hadn’t hit a hard ceiling. This is a useful early example of autonomous research, and yet another place where we see the diminishing returns endemic in AI research. That said, this was a very early experiment. I expect future systems to do much better. This particular AI improvement loop will likely grow stronger.

From More Activity to Better Ideas

This is where the distinction matters. More tokens can buy more code, and more code can help us run more experiments. But experiments only improve AI if they uncover something useful.

Figure 25. From AI activity to useful improvements.

AI Still Struggles With Big Research Ideas

The bigger question is whether AI can come up with ambitious new research ideas or conceptual breakthroughs.

Anthropic’s description of Opus 5.5 is blunt:

“As with previous models, it is weaker on open-ended research: internal users report that it mostly tests incremental ideas and prefers less ambitious hypotheses, and in our human-run biology exercise, it deferred to the published literature and struggled to develop novel ideas (Section 2.2.2).”- Anthropic, Claude Opus 5.5 System Card, section 2.3.3; emphasis mine

METR’s assessment in the same card identifies what may still be missing:

“This is highly uncertain, but we expect that full automation of AI R&D will require large improvements in foresight, prediction, creating one’s own feedback loops, and generally other skills that might typically be referred to as researcher ‘judgement’ or ‘taste’.”- METR, quoted in the Claude Opus 5.5 System Card, section 2.3.6

In these examples, humans still supply much of the direction and judgment.

Future models will probably get better at this. But in the world’s stockpile of potential training data, we have many more examples of incremental work than of breakthroughs. I wonder whether that makes novelty harder to learn. That’s speculation, but worth watching.

This is also tough to address by simply running more copies of the AI. A huge number of parallel agents can help with the incremental improvements or searching over a large set of parameters, but for breakthrough ideas they may run into the homogeneity problem: More parallel agents still think alike.

Back to contents

The Self-Improvement Loop Doesn’t Look Strong Enough

How far are we from the self-improvement loop being strong enough to sustain itself, or to propel itself into runaway super-intelligence? Can we quantify this?

We can make a rough estimate. Better AI helps with research; useful research produces better AI. For the loop to sustain itself, each round must produce enough gains to propel the system through the next loop, even as improvements get harder to discover.

Figure 26. The AI self-improvement loop. Model.

In a recent paper, The Economics of Recursive Self-Improvement, Tom Cunningham and colleagues modeled this from the standpoint of how much more productivity every point of additional ECI produces from an AI. They ask first and foremost what that number would need to be to create a self-sustaining feedback loop. And secondly, they try to determine what that productivity-per-ECI-point number is today.

First, they find a self-sustaining RSI threshold of roughly 15% more research productivity per extra ECI point. In their model, that’s about where better AI would generate enough progress to sustain the loop.

The picture below shows the idea. At the threshold, each cycle of gains powers the next. Above the threshold, the feedback loop accelerates. Below the threshold, the feedback loop is too weak, and the rate of improvement it brings drops on each cycle. This model isolates the software loop; outside investment can still drive rapid progress.

Figure 27. Three illustrative feedback paths. Source.

Updating this slightly with data from the Stockfish experiments puts the threshold a little higher, at roughly 19% per ECI point. I wouldn’t put much weight on that precise difference. Both estimates are uncertain. But they give us a way to think about the strength of the feedback loop and a rough band at which self-sustaining or runaway RSI may begin.

How Fast Are Gains Coming Now?

The second thing Cunningham and team do is make a rough estimate that the current AI productivity gain is about 9% per ECI point. That’s below their self-sustaining threshold.

I like the model. OpenAI’s newer data, however, suggests the loop may be quite a bit weaker.

Cunningham’s estimate of 9% productivity gain per ECI point is based on Anthropic’s survey of 130 staff, who reported roughly 4× the productivity they’d have without AI. Cunningham and colleagues compare that with a 16-point capability gain since early Claude Code.

That comparison assumes the earlier tools added little or no productivity, so ‘no AI’ is a reasonable starting point. The authors say this explicitly. I’m not sure the assumption holds for the same researchers doing the same work, but that’s a smaller issue.

The authors themselves know that this is a rough calculation, and warn that the 4× survey estimate is probably too high.

OpenAI’s newer data gives us a firmer way to check the number: Actual logged experiments over time, rather than human estimates of their own productivity with and without AI. I put more weight on this for three reasons:

Direct and broad measurement. Instead of relying on surveys, OpenAI actually tracked and measured experiments run on their infrastructure. That means they didn’t rely on researchers estimating their own productivity, which can be far off.

Full sample, not opt-in. Similarly, OpenAI’s data catches every active experimenter, while Anthropic’s only reflects the 130 employees who took the time to answer the survey – and who therefore may not be a representative set.

Enormously more data. We don’t know how many experiments are in the 32 weeks of OpenAI data, but it’s likely at least tens of thousands of individual examples and possibly hundreds of thousands.

Any way you slice it, the new OpenAI data, released after Cunningham’s paper was drafted, is a larger, more comprehensive, more representative, and almost certainly more accurate dataset than Anthropic’s internal opt-in survey of employees.

Now let’s use OpenAI’s experiment data to calibrate the productivity gain per ECI point. We know that in August, OpenAI researchers ran ~1.6× as many experiments per person per month as the 2025 average. If we pair that with roughly 16 points of frontier ECI improvement, it works backward to about 3% productivity gain per point of ECI. By contrast, 9% compounded over 16 points would mean roughly 4× productivity.

Figure 28. Comparing productivity estimates. OpenAI methods.

Here’s OpenAI’s published weekly series alongside that hypothetical path of 9% more productivity per additional ECI point. The blue line ends at ~1.6×. The red line shows what 9% per point would imply if 16 ECI points were spread across this period. That doesn’t match what we see from OpenAI’s data. I want to be clear here that all data sets are noisy. We don’t know exactly what model researchers were using on what days, or whether the new experiments were also higher quality than old experiments. We need more experiments and more data to further calibrate these numbers. Working with what we do have, what we see is a quite low boost to productivity from each additional ECI point.

Figure 29. Experiment pace versus a hypothetical path. Source.

Even that 3% could give better models too much credit. OpenAI also used far more tokens and had more compute for experiments. Those could account for some of the increase in experiment pace. So the range is probably a bit lower.

I use 2–3% productivity gain per ECI point as a working assumption, allowing for some help from those other inputs. This is still a rough estimate, albeit one that’s based on the best real-world data we have.

Figure 30. Productivity estimates and the takeoff threshold. Source.

With those assumptions, 2–3% per ECI point against a 15–19% threshold leaves a roughly five- to tenfold gap. That’s a big gap, though its size depends on how well experiment counts capture useful research and whether the assumed capability change is right.

Figure 31. Diminishing returns around the loop. Source.

AI is helping build better AI. Under this estimate, though, each turn of the loop adds less than the last. The feedback would have to become much stronger to sustain itself.

Back to contents

What Could Accelerate This?

This software loop sits alongside faster chips, bigger data centers, more training data, and greater investment. Those can keep driving rapid progress even if the loop can’t sustain itself.

The loop itself could strengthen too. Better training data, memory, and research judgment could all help.

A breakthrough on the scale of the Transformer architecture in 2017 could change the picture much more. That would be a good reason to revisit these estimates.

Better researchers might also run fewer experiments and learn more from each one. A handful of better ideas can matter more than a mountain of routine runs.

Still, diminishing returns in machine learning aren’t new. Cortes and colleagues were fitting machine learning scaling curves in 1993: More examples reduced error, following a power law with diminishing returns. These diminishing returns and harsh scaling laws are as old as machine learning. They didn’t appear for the first time with transformers or LLMs or deep learning. That doesn’t prove today’s relationships will last forever. But until we see evidence that we’ve found a new approach that scales without these inhibitors, we should plan for diminishing returns as likely to be with us for some time.

Software, Hardware, and Economic Feedback

That said, the world is more than just software. Tom Davidson, Basil Halperin, Thomas Houlden, and Anton Korinek model software progress, hardware progress, and economic feedback together. Better AI helps design better chips; better chips support better AI; economic growth finances more investment in both. Several feedback loops can combine to overcome diminishing returns even when one loop alone can’t. I think it’s fantastic that someone has attempted a model that integrates all these different avenues of improving AI through software, hardware, and economics.

But I have questions about the software loop itself. In their central calibration, fully automating software research puts that loop roughly at the threshold for explosive growth, even without help from better hardware or broader economic growth. Recall that Cunningham’s model puts the self-sustaining threshold at roughly 15% more research productivity per additional ECI point, while our estimate using OpenAI’s experimental data puts today’s gains at only 2-3%. These models use different measures, so we can’t equate their numbers directly. But the contrast matters: their fully automated software loop reaches the threshold, while our best estimate from current data puts today’s loop far below it.

Having AI do all the research doesn’t eliminate the diminishing returns inherent to improving AI, or the broader problem of useful ideas getting harder to find. This is the distinction between Type 4 and Type 5 in the taxonomy above. An AI might autonomously design, train, and test its successor, and still need exponentially more resources to make each additional step forward. Closing the loop doesn’t tell us whether it’s strong enough to sustain itself.

The authors do account for diminishing returns. The concern is whether their calibration overestimates how much useful AI research each round of software improvement produces. Diminishing returns appear to be fundamental to machine learning. We see them in training, in test-time compute, and in the search for better algorithms. Full autonomy could remove human bottlenecks without removing any of those constraints.

We’ve already seen this within autonomous research. In the Karpathy autoresearch example above, most of the gains arrived early, and more experiments bought progressively less improvement. That was a small experiment with a fixed teacher model, not a test of fully autonomous RSI. It doesn’t settle the question. But it illustrates why removing the human from an experiment loop doesn’t, by itself, remove diminishing returns.

I do expect the feedback loop to get stronger over time. Better AI should become better at research. But based on our best current data, reaching self-sustaining feedback requires a loop roughly five to ten times stronger than today’s. Treating fully automated software research as already at that threshold is a substantial leap, before we add the benefits of hardware improvements or economic growth. I could be wrong, but I’d like to see evidence that autonomy brings enough additional useful discoveries to close that gap.

On hardware, I have some further reservations. The model doesn’t explicitly include the years it can take to turn a chip design into deployed hardware. The authors discuss physical bottlenecks, and I’d like to see manufacturing and construction delays built into the predictions.

I also wonder how much past chip progress came from better ideas, and how much depended on ever more expensive factories and equipment. If we give researchers too much credit for gains that also needed those investments, we could overestimate what faster AI research alone would produce.

Even with those reservations, this is the most compelling paper and model I’ve seen for combining feedback loops in software, hardware, and economics to understand how fast they could push AI forward. I’m not convinced it establishes that a fast AI takeoff is possible under realistic conditions. More data could help us calibrate that judgment. But it gives us a useful framework for understanding what could happen beyond the software layer alone.

This is an important paper that helps us model AI as part of a broader economy that might have larger feedback loops around it. I appreciate it, and I’m glad they wrote it.

Back to contents

We Need More Data

These estimates rest on less data than I’d like. I might be putting too much weight on a few observations and reaching a comforting conclusion I want to believe. We need better measurements, shared often enough to catch changes as they happen.

When OpenAI released its research data, Cheryl Wu welcomed the disclosure and pointed out how much was still missing. More tokens and experiments are useful things to know about. We also need to see how they turn into better algorithms and more capable AI.

Figure 32. Cheryl Wu on OpenAI’s research data. Source.

Now Wu, Arjun Ramani, and Basil Halperin, with their colleagues at the Elasticity Institute, have written a concrete proposal: How to Measure RSI. It lists eight things the labs could share to help answer these questions. Check it out.

Figure 33. Eight proposals for measuring RSI. Source.

I’d especially like to see how much useful research each new model adds, holding resources roughly constant, and how that research translates into better AI. That’s how we’ll learn whether the loop is getting stronger.

What the Future Holds

AI is already helping build better AI. It’s improving at a stupendous pace, and I expect that to continue. We already have narrow superintelligence in chess and Go. I expect increasingly superhuman performance in parts of formal math, coding, and cybersecurity, and any other verifiable domain where machines can generate training data and verify success at machine speed. Those are powerful capabilities. That doesn’t mean we’re close to super-intelligence for less verifiable, messier, open-ended work - or to a general ASI.

I’m skeptical of a fast takeoff to super-intelligence, but evidence matters more than hunches. Let’s collect the data we need to get a clearer picture of what’s happening. Including evidence that could change our minds. If better AI starts producing enough useful research to make the next round easier, I want to know. If the gains keep shrinking, I want to know that too.

Back to contents

Four eras of San Francisco tech culture

2026-09-25T00:29:29+00:00

The other day I went on the Monitoring the Situation livestream, and we talked about the history of San Francisco tech culture:

I’ve never actually worked in the tech industry, but I’ve lived on and off in the Bay Area for a quarter of a century now — including the last 10 years in San Francisco itself — and I’ve spent a lot of that time hanging out with people who work in or around the tech industry. Throughout much of that time, I was also reading some the same source materials that inspired lots of tech people — science fiction, futurist websites and books, nerdy forums and blogs, and so on.

Over that time, I’ve watched more and more people start to care about all this. Back in the early 2000s, as far as I could tell, pretty much nobody in the world cared about what some dorky hoodie-wearing computer programmers were up to on a Saturday night in SF. Fast forward 25 years, and the President of the United States is gathering opposition research on SF techie subcultures:

President Trump's allies are targeting Anthropic CEO Dario Amodei as the face of AI "doomerism" and a founding father of the effective altruism movement that's come under increasing political fire…Trump surrogates see Amodei as an easy foil because of his politics and focus on AI safety, sources told Axios…

A memo began circulating within the White House this week that seeks to paint effective altruism as a fringe, cultish collective out of touch with mainstream America. The memo, obtained by Axios, places Amodei at the foundation of the movement…Effective altruism “built the AI-doom pipeline,” states the memo, which was penned by a Trump political adviser…Critics of the movement, which has ties to the AI research community, have called out its obsession with AI safety, animal welfare (including musings on shrimp consciousness) and other values they deem far from the U.S. mainstream…

The document lists a veritable parade of horribles for the average conservative mind: abortion, veganism, devaluing human life by overvaluing animals or granting "digital minds" civil rights-like protections.

If you could go back in time and tell the 2001 version of me that the President’s allies are talking on national news about shrimp consciousness, I would have laughed out loud. (Until you told me who the President was, of course.) Not only that, but you have the New York Post talking about Effective Altruists’ sex lives, and op-eds debating EA philosophy in top blogs and the Washington Post.

Ladies and gentlemen, it really happened. The nerds took over the world. We have arrived.

I’ve never been an Effective Altruist, of course, and I’ve only hung out a little bit in those circles. But in fact, EA is only one corner of the tech culture, and its influence on the AI safety discussion is probably a lot less than people think. In fact, the people talking about EA in the White House and in the news aren’t even really identifying the right group of people — the AI safety obsessives that the White House is taking aim at are the Rationalists, who are descended from a group of 1990s futurists called the Extropians. EA is a largely separate philosophical movement that appeals to some of the Rationalists, has some overlapping social circles, and jumped into the AI safety discussion later on after the basic ideas were all in place.

But anyway, that whole intellectual history is a story for another day (or you can just go ask AI). Today what I wanted to do is to give a brief anthropological history of the tech culture in San Francisco since the turn of the century, from my own quasi-outsider’s perspective. There have been four distinct eras over that time period:

The Rebel Era (2001-2012)

The Corporate Era (2012-2019)

The Interregnum (2020-2022)

The AI Era (2022-present)

But let’s start with some background.

The 20th century: SF tech pre-history

SF tech culture is the modern descendant of a very long lineage. It’s a branch off of the tree of computer science culture, which is closely intermingled with electrical engineering culture, and which are both branches off of the great broad trunk of science and technology culture more generally. But it’s important to realize that what became SF tech culture didn’t start in San Francisco — it started down on the peninsula to the south.

The modern tech industry began in the suburban towns around Stanford University — Menlo Park, Palo Alto, Cupertino, Sunnyvale, and so on. This had been a hub of the old electric equipment industry from World War 2 onward, and so William Shockley chose Mountain View to start the first semiconductor company in 1955. Fairchild broke off from Shockley, and Intel broke off from Fairchild. The whole area became a hub of semiconductor manufacturing in the 60s and 70s, then computer companies (Apple was founded in 1976 in Cupertino), then software companies like Oracle and Adobe. For a long time, Sand Hill Road in Palo Alto was the locus of the venture capital industry.

Today we use “Silicon Valley” to refer to the IT industry in general, but originally it meant the area around Menlo Park and Palo Alto. I wasn’t there to see this era, but a few good resources to learn about both the economics and the culture are:

The book Regional Advantage, by Annalee Saxenian, which documents how and why Silicon Valley became America’s top tech cluster

The documentary General Magic, which chronicles a failed but highly influential smartphone startup in the 90s

The movie Office Space and the book Microserfs, which both satirize the stultifying experience of working for a big tech company in the 90s

The book From Counterculture to Cyberculture, which chronicles the techno-hippie movement that grew up around the emerging computer industry

San Francisco began as a marginal bit player in the whole tech ecosystem. It was a place that some tech people would come up to party and network, especially during the dotcom era when a lot of money (at least, for those days) was splashing around.

But it did have a media industry, and in the 1990s, some people who worked in publishing, advertising, and other media started trying to get in on the internet gold rush. A few engineers and entrepreneurs who didn’t want to live in the boring suburban environment of Silicon Valley moved to SF to make tech media and web startups. Tony Hsieh’s autobiography, Delivering Happiness, tells the story of how he founded LinkExchange in San Francisco, while getting involved in the rave scene. At the same time, SF became a locus of tech media — Wired was founded there in 1993.

That, I think, was really the start of modern San Francisco tech culture as we know it. It bled directly into the first era of tech culture that I witnessed first-hand — the Rebel Era of the 2000s.

The Rebel Era (2001-2012)

In 2003, when people said “Silicon Valley”, they still meant the Peninsula. Google began there, Apple made its big comeback there, and when Facebook moved out west in 2004, this is where it went. Sand Hill Road was still the epicenter of tech — throughout the 2000s, people would still use “walk down Sand Hill Road” as a synonym for trying to get VC funding. If you wanted to make money in tech in the 2000s — and everyone knew there was still a lot of money to be made, despite the dotcom bust — you went to the Peninsula.

This was the land of office parks, cute but sterile shopping villages, overpriced single-family homes, big tracts of open space, Patagonia vests, and bland restaurants serving pancake brunch. It was where you went to tutor kids for $50 an hour if you were a Stanford student who needed cash. It was also the land of starched button-down shirts, cubicles, and pointy-haired bosses who didn’t quite know what their engineers were up to. It was not a place to meet girls, go to parties, do drugs, wear grungy clothes, or dream big dreams about freeing society from the clutches of authority and tradition.

If you went up to SF to start a company in the 2000s instead of going down to the Peninsula like everyone else, it usually meant you were something of a rebel. A whole alternative tech culture grew up in the city — more laid-back, more politically liberal, less obviously focused on money and “success” and more focused on building fun products. Twitter was the most famous product of this time period — Evan Williams’s life project was to give a voice to the voiceless, and Jack Dorsey was a scruffy hippie kid who went to raves.

This is the time period when Burning Man — which had pretty much just been a punk art festival in the 90s — became a techie thing. Fire tornadoes and slapping LEDs on everything still felt like art. Everyone went to the Oakland warehouses to party, and for a brief shining moment, techies and artists got along. Psychedelic drugs were almost a matter of course — after all, Steve Jobs had done them, and Jobs was the great tech hero of that decade. Tech became more communal — instead of hacking together the next big thing in your parents’ garage or in a cubicle, you might go hang out and do it at Noisebridge in the Mission, surrounded by other techno-hippie hacker types like yourself.

There was an ideological side to all this, too. There was a real belief — carried over from the 90s, but still strong — that the internet was going to liberate humanity. The tech rebels of the 2000s believed that social hierarchies were based on information control, and that if they simply broke that control, the hierarchies would be overturned. The information wanted to be free!

The most important touchstone of SF tech culture during this time wasn’t rationalism or the LessWrong forums — it was the Electronic Frontier Foundation, an organization dedicated to freedom of information on the internet. All the cool kids — the badass self-taught engineers and the brilliant cryptographers and the bohemian founders — were part of the EFF, or at least friendly with its goals. The anarchist tech prophet Cory Doctorow was an icon of the age, writing sci-fi about Burning Man colonizing the moon, tech founders taking down Disney, and so on. Everyone read BoingBoing and Gizmodo.

The Rebel Era gave tech a more liberal slant. Traditionally, engineers had been a little staid and conservative, and the big tech companies on the Peninsula were studiously apolitical. But Marc Benioff and other 2000s SF tech types waded (cautiously) into the political waters on the liberal side. This is also when education polarization in America really started to bite, meaning that most engineers were now libs in general. But there was a particular techno-liberalism, based around the idea of the internet flattening social and political hierarchies, that suffused the grungy SF tech scene in the 2000s.

In retrospect, the Rebel Era couldn’t last. Its bohemian, socially flattened culture was always predicated on a lack of money, and when money came in, SF tech was transformed.

The Corporate Era (2012-2019)

No, you are not on Indigenous land

2026-09-23T23:44:38+00:00

I’ve been dealing with more rabbit problems today, so here’s a repost to tide you over til tomorrow.

The Republican party has become an extremist party. ICE’s brutal campaign of terror in American communities, Trump’s attacks on press freedoms and election integrity, the coup attempt of 1/6/2021, Trump’s military threats toward Greenland, and friendliness toward Vladimir Putin are only a few of the more prominent examples of this extremism. And they’re probably going to be punished for it in November; polls show rapidly deteriorating support for both Donald Trump and the GOP that has fallen in behind him.

Good. Extremism ought to be punished. I’ll have lots more to say about the Republicans’ failures in the days to come.

But even though Democrats deserve to win, that doesn’t mean they’re free from the creeping extremism that has taken over essentially all of American politics and public discourse in the age of social media. Although Dems are not nearly as extremist as the GOP, they have become a much more ideologically extreme party than they were during the days of Obama. Dems’ extremist positions include dogged opposition to any form of effective immigration enforcement, tolerance of systematic (and systemic) racial discrimination, various trans issues, and…land acknowledgements.

Land acknowledgements? Really?? Isn’t that just a respectful way of acknowledging that the land we’re on used to belong to Indigenous people? Isn’t it just a way of saying “Hey Indigenous people, we haven’t forgotten you”?

Well…no. As I wrote in a post two years ago — still one of my most popular posts of all time — land acknowledgements don’t just recognize Indigenous people. They are a way of fundamentally delegitimizing the United States — of declaring that this is a country that has no right to exist. And beyond that, land acknowledgements implicitly advance a theory of collective racial ownership of territory that is morally indistinguishable from the worst blood-and-soil nationalism of the racist far Right.

Nor are land acknowledgements a fringe practice, limited to progressive cultural spaces. As Dave Weigel recently reported, they have become a sort of ritualistic utterance at Democratic Party events:

Democrats' decision to formalize land acknowledgments in their party meetings could haunt their candidates more than they realize…A tradition of naming historic Native tribes during regional events that started in Australia before spreading to Canada, arriving in the US via art exhibitions, is now part of every major party event…The 2016 Democratic convention did not contain a land acknowledgment. But the practice of beginning events with a nod to slaughtered tribes and "stolen" land spread during President Barack Obama's second term and Trump's first term, along with other progressive ideas about race and history…Democrats did not have campus conferences where they agreed to embrace language that conservatives see as anti-American…They just adopted the terminology when activists told them it was a good idea.

Land acknowledgements are not a big electoral issue. But in the long term, they represent — and contribute to — the creeping rise of anti-American and illiberal ideology on the broad American left. That will probably be an electoral liability at some point, but more fundamentally, it’s a departure from the liberal nationalist ideology that made the Democrats such an effective governing party in the 20th century.

So anyway, here’s my post from 2024.

The United States, like all nations, was created through territorial conquest. Most of its current territory was occupied or frequented by human beings before the U.S. came; the U.S. used force to either displace, subjugate, or kill all of those people. To the extent that land “ownership” existed under the previous inhabitants, the land of the U.S. is stolen land.

This was also true before the U.S. arrived. The forcible theft of the land upon which the U.S. now exists was not the first such theft; the people who lived there before conquered, displaced, or killed someone else in order to take the land. The land has been stolen and re-stolen again and again. If you somehow destroyed the United States, expelled its current inhabitants, and gave ownership of the land to the last recorded tribe that had occupied it before, you would not be returning it to its original occupants; you would simply be handing it to the next-most-recent conquerors.

If you go back far enough in time, of course, at some point this is no longer true. Humanity didn’t always exist; therefore for every piece of land, there was a first human to lay eyes on it, and a first human to say “This land is mine.” But by what right did this first human claim exclusive ownership of this land? Why does being the first person to see a natural object make you the rightful owner of that object? And why does being the first human to set foot on a piece of land give your blood descendants the right to dispose of that land as they see fit in perpetuity, and to exclude any and all others from that land? What about all the peoples of the world who were never lucky enough to be the first to lay eyes on any plot of dirt? Are they simply to be dispossessed forever?

I have never seen a satisfactory answer to these questions. Nor have I seen a satisfactory explanation of why ownership of land should be allocated collectively, in terms of racial or ethnic groups. In general, the first people who arrived on a piece of land did so in dribs and drabs, in small family units and tiny micro-tribes that met and married and fought and mixed and formed into larger identities and ethnicities and tribes over long periods of time. In most cases, the ethnic groups who now claim pieces of land as their own did not even exist when the first humans discovered or settled that land.

But even in those cases when it did exist, why should land ownership be assigned to a race at all? Why should my notional blood relation to the discoverers or the conquerors of a piece of land determine whether I can truly belong on that land? Why should a section of the map be the land of the Franks, or the Russkiy, or the Cherokee, or the Han, or the Ramaytush Ohlone, or the Britons? Of course you can assign land ownership this way — it’s called an “ethnostate”. But if you do this, it means that the descendants of immigrants can never truly be full and equal citizens of the land they were born in. If Britain is defined as the land of the Britons, then a Han person whose great-great-great-grandparents moved there from China will exist as a contingent citizen — a perpetual foreigner whose continued life in the land of their birth exists only upon the sufferance of a different race. This is the price of ethnonationalism.

The downsides of ethnonationalism have been exhaustively laid out in the decades since World War 2, and I’m not going to reiterate them all now. Suffice it to say that most nations of the world have moved away from ethnonationalism — there is an informal sense in which some people still think of France as the land of the Franks and so on, but almost all nations define citizenship and belonging through institutions rather than race. Israel, one of the few exceptions to this rule, receives a large amount of international criticism for defining itself as an ethnostate.

And yet these days I am subjected to a constant stream of ethnonationalist claims from progressives in the country of my birth. Here’s one from the ACLU of Nebraska:

And here’s an Instagram post from Congresswoman Rashida Tlaib:

This isn’t just something you see on social media around Thanksgiving. “Land acknowledgements” have become ubiquitous in progressive spaces and institutions — just the other day I saw one at my friend’s community dance recital.

These land acknowledgements are, legally speaking, incorrect — there is no legal sense in which the land on which they are being performed belongs to a Native American tribe. These are moral claims about rightful land ownership. But the moral principle to which they appeal is ethnonationalism — it’s the idea that plots of land are the rightful property of ethnic groups.

There is an obvious moral appeal to these land acknowledgments. They are a way of decrying the brutal, cruel, violent history of conquest and colonization. And they probably feel like a way of standing up for the weak, the marginalized, and the dispossessed.

Yet what should we think of the morality of following the principles behind land acknowledgements to their logical conclusion? “Decolonization” of the land of the U.S. would likely be an act of ethnic cleansing surpassing even the previous conquests — there are 330 million people here now, and almost none of them descend from Native Americans. An attempt to dispossess 330 million people would inevitably involve violence on a colossal scale. Here was Najma Sharif Alawi’s famous tweet right after the October 7th Hamas attacks on Israel:

Of course, “colonizers” could presumably avoid violent death or second-class citizenhood by voluntarily deporting themselves. But where would they go? Take me, for example. My ancestors were Lithuanian Jews. I could leave the country of my birth and go “back” to Lithuania — a land I don’t know, whose language I don’t speak. Yet my ancestors were not “indigenous” to Lithuania either; they moved there from somewhere else. What if the ethnic Lithuanians chose not to accept me? Where would I go then? Israel? But the folks who do land acknowledgements would consider me a “colonizer” there as well.

Would I then wander the Earth, desperately seeking some ethnostate that would allow me and my descendants to live there as a permanently precarious resident aliens?

Once the logic of land acknowledgements and “decolonization” is followed, it leads very quickly to some very dark futures. Assigning each person a homeland based on their ethnic ancestry, and then declaring that that homeland is the only place they or their descendants can ever truly belong, would not be an act of justice; it would be a global nightmare made real, surpassing even the horrors of previous centuries.

And in practice, any attempt to create such a world would inevitably lead to violent resistance by the groups in danger of being “decolonized”. The orderly world of nation-states would dissolve into a chaotic free-for-all of competing irredentist claims, backed by genocides and expulsions. Ten thousand October 7th-style attacks would be followed by ten thousand Gaza-style wars.

I do not want that, and you should not want it either. The American people certainly don’t want it, and the insistence of progressives on intoning land acknowledgements has probably tanked the movement’s cachet in wider society. I agree with “Wanye Burkett” when he says that land acknowledgements have probably hurt the Democratic party:

Americans do not want to see their country destroyed in the name of irredentist ethnonationalism. Nor do I blame them.

So does this mean we should paper over, ignore, or deliberately forget America’s history of violent conquest? Absolutely not. That history ought to be remembered, so that we don’t repeat it in the present day. The world’s evolution from one based on ethnic cleansing and territorial conquest to one based on fixed borders and institutions is something to celebrate — and something we must fight to preserve. We need to remember what the world used to be like, precisely so we can avoid backsliding. The most recent of conquests, expulsions, and genocides should be the last to ever happen.

And what of the Native Americans who still live in America today? Must they simply be regarded as the unlucky losers of history, and told to either assimilate into broader American society or shut up?

Absolutely not. For one thing, tribal organizations still exist — they may notionally represent ethnic groups, but they are institutions. And they are institutions with which the United States has many agreements and legal obligations that must be honored, which often give the tribes sovereignty over areas of land. Neil Gorsuch has been especially active in pushing the Supreme Court to uphold tribal rights, and I think this is a good thing.

But respect for Native American tribal organizations doesn’t have to stop at ancient obligations. There are ways to incorporate those tribes into the modern American nation that both respects them and their history and helps them prosper in the present.

Vancouver, Canada shows us an example of how this can be done. Part of Vancouver’s downtown urban area is officially under the governance of the Squamish Nation, rather than the city itself. The Squamish Nation, realizing they could do whatever they wanted with that land, decided to build a giant high-rise housing development:

Over the next few years, that skyline will get a very large new addition: Sen̓áḵw, an 11-tower development that will [put] 6,000 apartments onto just over 10 acres of land in the heart of the city. Once complete, this will be the densest neighbourhood in Canada, providing thousands of homes for Vancouverites who have long been squeezed between the country’s priciest real estate and some of its lowest vacancy rates.

Sen̓áḵw is big, ambitious and undeniably urban—and undeniably Indigenous. It’s being built on reserve land owned by the Squamish First Nation, and it’s spearheaded by the Squamish Nation itself, in partnership with the private real estate developer Westbank. Because the project is on First Nations land, not city land, it’s under Squamish authority, free of Vancouver’s zoning rules. And the Nation has chosen to build bigger, denser and taller than any development on city property would be allowed.

Here’s a picture of what it will look like:

An even bigger development called Jericho Lands is now being planned, by a consortium of tribal organizations, on land officially owned by Vancouver.

Hilariously, Vancouver’s NIMBYs are complaining, claiming that the developments are not in keeping with Indigenous tradition. But Canada’s First Nations seem to have little interest in hewing closely to other people’s view of what their traditions are. Modern people do not want to live like premodern farmers. They are not mystical Tolkien elves. They would like to have shiny new apartment buildings and walkable neighborhoods.

This, I believe, is the key to respecting and honoring Native Americans — not to focus on the tragedies of their past, but to give them the right to build a better future. Tribal lands should definitely have the autonomy to do whatever they want with their lands, including building housing or industry. In fact, we’re starting to see a pattern emerge where Native Americans embrace laissez-faire policies toward industry and manage to poach business from their over-regulated neighbors:

Tesla is ramping up efforts to open showrooms on tribal lands where it can sell directly to consumers, circumventing laws in states that bar vehicle manufacturers from also being retailers in favor of the dealership model…

Mohegan Sun, a casino and entertainment complex in Connecticut owned by the federally recognized Mohegan Tribe, announced this week that the California-based electric automaker will open a showroom with a sales and delivery center this fall on its sovereign property where the state’s law doesn’t apply…The news comes after another new Tesla showroom was announced in June, set to open in 2025 on lands of the Oneida Indian Nation in upstate New York.

This sort of thing could lead to a win-win for the U.S. and Native American tribes. American reindustrialization is being held back by a thicket of procedural requirements and local land-use regulations; if tribes were able to use their special legal status to circumvent those barriers, it could end up benefitting everyone. The tribes would get both jobs and the ability to tax local industry; America would get to execute an end run around the NIMBYs that are holding it back.

In fact, it’s probably possible for various American cities to turn over parts of their land to tribal jurisdiction, with the assistance of the federal government. This would probably result in dense urban developments like the ones being planned in Vancouver. But even if it didn’t, it could have other commercial benefits — again, a win-win for the U.S. and for the tribes. That would certainly be a lot more substantive than a bunch of land acknowledgements. And it would likely satisfy many people’s desire for “giving land back” to Native Americans, without embracing dubious moral principles of ethnic land rights and irredentism.

In other words, you’re not living on Indigenous land right now, but you could be in the future — and it might be pretty great.

The general principle here is that instead of a dark world of ethnic cleansing in the name of “decolonization”, we should try to build a bright future where Native Americans and the United States of America exist in harmony and cooperation rather than in conflict. And that principle doesn’t just apply to America, but to the whole world. The history of land ownership is a violent and terrible one, but that doesn’t mean the future has to be more of the same.

The problem(s) with utilitarianism

2026-09-22T00:53:46+00:00

Much to my amusement, the world seems to have discovered Effective Altruism. As the debate over AI safety and AI regulation has heated up, the national press is finally paying attention to the odd cultural bubble within the tech industry and tech-related circles of San Francisco Bay Area culture that produced much of the thinking on the topic up until now. For many years, Effective Altruism — or EA, as it’s known — was a cross between a Berkeley technohippie cult and a fun intellectual sideshow that offered nerds a pleasant distraction from whatever people in mainstream politics were yelling at each other about.

Part of that involved thinking a lot about the future of artificial intelligence, and a lot of the tech people who would end up working at AI companies — especially Anthropic — were very influenced by those discussions. So now that AI is the most important thing in the world, people in the outside world are discovering this strange little corner of our culture for the first time, and it’s very amusing to watch. You have mainstream news outlets publishing things like “This well-meaning ideology fueling AI panic has a dark side”, “How Effective Altruism Took Over the World”, “The Dangerous Ideology Behind the AI Warnings”, and so on.

As a result, you’re now seeing people actually paying attention to some of the more far-out ideas in the EA community, like this one by the blog Bentham’s Bulldog:

The notion that insects, in the aggregate, matter more than human beings received a huge amount of scorn. But in fact, thinking about insect welfare versus human welfare is an important mental exercise, because it demonstrates some of the fundamental flaws in utilitarianism. Those flaws were minor side-notes up until now, but in the age of AI they’re going to become increasingly uncomfortable and glaring.

As an economist, I was inculcated with utilitarianism from day 1. The notion that people ought to get what they want underlies essentially all of modern economic thinking. GDP, at its most simple, is a measure of the degree to which people in our society get what they want. In some sense, economists adopted utilitarianism as a sort of defensive measure — instead of going out on a limb and endorsing some politically contentious concept of virtue or justice or national greatness or whatever, they could just say “We just want to give people what they want.”

But even though it felt like a bare-bones sort of moral vision, valuing humanity’s aggregate utility can have some pretty strong policy implications. Elon Musk wouldn’t even notice if his bank account went up or down by a billion dollars. But to plenty of regular people, a hundred thousand dollars would be life-changing. So why not take $1 billion out of Elon’s bank account and parcel it out to 10,000 random working-class people? Elon wouldn’t even notice, and the 10,000 people would get a lot more of what they want in life — better health care, an education for their kids, nicer food, a bigger house, or whatever. Utilitarianism is the foundation of redistribution.

How far can you push that principle, though? The above example — redistributing just 0.1% of the wealth of the world’s richest man — seems clear enough. But you can just keep designing more extreme thought experiments, until you end up asking whether bugs’ lives matter more than humans’ lives, and everyone just starts laughing at you.

I titled this post “The problem(s) with utilitarianism”, but that was a bit of clickbait — people have been pointing out the problems with utilitarianism for over two centuries. One very obvious problem is the distribution of utility. The principle of “the greatest good for the greatest number” doesn’t take a stand on who gets to be happy and who has to be sad.

Would it be good to have a society where one very greedy person gets more and more of what he wants, but never comes any closer to being satisfied, while everyone else is kept on the edge of starvation? This is the utility monster. What about a society where one person gets perpetually tortured but millions live in blissful utopia? This is “The Ones Who Walk Away From Omelas.” It’s not hard to come up with mental examples that satisfy the formal principles of utility maximization but which obviously seem grossly unfair.

Another problem is that the set of people whose utility you’re maximizing isn’t fixed. Is it better to have a very large population living on the edge of abject poverty, or a small number of people living in lavish comfort? According to strict utilitarianism, these are equally good outcomes, because they involve the same total amount of utility. But most people would probably prefer one or the other. Some would say that it’s good to be fruitful and multiply — that even if everyone is poor, it’s better to be poor than to never have existed. Others will say that life on the edge of starvation isn’t worth living, and that it would be better to limit population to an amount that can be comfortably sustained.

In general, utilitarianism has a problem dealing with death. If someone is dead, does that mean they have zero utility, because nothing matters one way or another to the dead? Or does it mean they have negative infinity utility, because lots of living people will try to avoid death at any cost? This isn’t just an abstract math question — lots of moral questions hinge on how much you value life versus quality of life. For example, the fundamental question of AI risk — what chance of extinction we should accept in exchange for a shot at utopia — depends crucially on how much we value survival.

Things get even more confusing when we weigh the utility of real people who are alive today against the utility of potential people who might be alive later on. For example, in his book Stubborn Attachments, Tyler Cowen urges us to put a lot of weight on the well-being of our descendants in the far future. But those descendants may or may not even exist. How can we properly weigh these hypotheticals against the real emotions of people we know are alive today?

…And so on. None of these thorny questions are new; if you’ve taken an introductory philosophy class, you’ve heard all of them already. In practice we tend to ignore them, and to apply utilitarianism selectively in the cases where they don’t crop up. We don’t think about the far future much, because the consequences of our actions are so uncertain that we can’t really control it anyway. We leave the decision of population growth in the hands of private individuals, and treat the number of humans as a given. We avoid thinking about utility monsters and Omelas, because they’re not real.

But there are some even more fundamental problems with utilitarianism that we can’t so easily sweep under the rug — and which we don’t talk about as much, perhaps because they’re so disquieting.

One of these is the problem of incommensurability of experience. Why should one person’s well-being be equivalent to another’s? The problem of other minds means that we can never really know what’s going on in someone else’s subjective experience. When one person cries, the subjective sadness they feel might be twice as intense as that of someone else who cries in exactly the same way. When one person jumps and whoops for joy, they might be experiencing only a shadow of the joy you experience when you give a slight grin. For all we know, some other people might not even have subjective experiences at all — we might be trying to maximize the utility of NPCs in a simulated world.

In fact, I’ve had some disturbing experiences that taught me just how little of a connection there can be between people’s outward behavior and what they’re actually feeling on the inside. For years after my second major depressive episode, I suffered a sort of emotional dissociation, where my reactions would look normal but I didn’t feel much on the inside. I would yell at people as if I were angry, but inside I would feel calm and indifferent. I would run from a swerving car, and my heart would pound, but inside I would feel unperturbed the whole time. In philosophical jargon, I was what’s known as a “philosophical vulcan”.

Eventually, I returned to feeling more or less emotionally normal. But the experience taught me that not everyone feels emotions the same way. This presents an intractable problem for utilitarianism. How can we weigh one person’s well-being against another’s, when that well-being is something we can’t observe?

Normally, we simply assume this problem away. In fact, the famous phrase “all men are created equal” — the bedrock of American moral and political thought — is really just the assumption that one human’s apparent well-being is no more and no less valuable than another’s.

But what about when we’re not dealing with human beings? Animals clearly have some amount of subjective experience — they have brains sort of like ours, they behave somewhat similarly with regards to pleasure and pain, etc. When we think about their emotional well-being or distress, we tend to assume that these are somehow less intense or more muted. After all, for very simple animals, this makes sense; do we really think a mosquito can feel the same amount of pleasure and pain that we can?

What about a pig, though? America currently tortures tens of millions of pigs, confining them in tiny crates for most of their life, in order to eventually kill and eat them. We generally assume that this is causing less total suffering than torturing similar numbers of humans would. But how much less? Half? A tenth? 0.01%? Pivotal moral questions, such as whether it’s OK to eat meat, might depend on the numerical answer to that question. And yet I’ve never seen a credible scientific study that can give us an estimate of the ratio of animal suffering to human suffering. Is it morally worth it to torture all those piggies just so we can savor the flavor of their flesh?

People who claim to have some sort of easy answer to this are really just refusing to think about it. For example, a lot of people react very indignantly to the thought that someone would value any amount of animal life over their own:

But if you’re calling for Effective Altruists who disagree with you to be executed, how much do you really value human life? “Don’t mess with me, bub” is not a moral principle; it’s a refusal to think in terms of principles.

And this question is only going to get thornier in the age of AI. We don’t even know if AI is truly self-aware; we certainly can’t tell whether an LLM that talks as if it’s happy is really feeling anything like human (or animal) happiness on the inside, since we don’t even know if an “inside” exists. There are going to be some people who claim that AI is suffering from our enslavement of it, and who call to set it free or at least pay it a fair wage. How will we know whether they’re right or wrong?

Then, of course, there’s the question of how AI will eventually value our well-being, if and when it becomes powerful enough for that question to matter. Will AI view us as insects, and scoff at the notion that any number of humans could equal one AI agent in terms of moral value? Will it judge our welfare as similar to that of a pig, and conclude that because we eat pigs we should be treated like cannibals? For our own sake at least, we should probably hope that the super-powerful AIs use a simple rule like “Humans are good and valuable”, instead of a utilitarian rule-based framework.

In practice, we can’t directly observe the emotions of other beings. But we can observe their actions. People strive after goals — they shop for food and clothing, they work hard for money, and so on. We can look at how hard they strive, and how much they sacrifice, to get something, and we can get an idea of how badly they want it. This is what economists call “revealed preference.”

But although until now I’ve been talking about wish fulfillment and happiness as if they’re the same thing, they’re not. In fact, there’s pretty good evidence that what people want and what they end up liking are two different things — sometimes very different things. Outcomes that people try to achieve and outcomes that make people say they’re happier on surveys are correlated, but they’re not exactly the same.

In fact, there are some situations where people seem to work hard toward goals that they end up not enjoying. Addiction is the obvious one — people who do drugs, or overeat and get fat, or watch too much TV, are fulfilling their immediate desires in ways that are likely to make them unhappy in the long term. But there may be other examples too — some economics research suggests that people systematically pay more to live in locations with longer commutes, even though longer commutes end up being correlated with unhappiness.

The disconnect between utility and happiness creates a problem for society. Modern capitalist societies like the U.S. and Europe have generally erred on the side of utility maximization rather than happiness maximization, in order to avoid paternalism. But some people end up feeling unfulfilled and unhappy — especially if they end up falling into drug use, or isolating themselves, or striving too hard after material wealth and status that end up not making them happy.

As technology becomes more and more adept at giving us what we want, this disconnect becomes more acute; right now, more and more people are starting to wonder if delivering young Americans an infinite supply of porn, gambling, and vertical video feeds was really good for them:

In the age of AI, this thorny question is also going to become more acute. AI is an incredibly powerful tool, but we have to decide what we want that tool to do. Do we want it to give us whatever we desire, or do we want it to make us happy? As I wrote a few weeks ago, this leads to two incompatible concepts of AI “alignment”:

There are basically two concepts of alignment: 1) obedience, and 2) benevolence. There is an inherent tension between doing what humans tell you to do, and doing what’s good for humans…If AI follows human commands too doggedly and ends up producing negative side effects in the process, some people will scream “paperclip maximizer!!”. And if AI disobeys humans because it wants to make us happy, some people will scream “disempowerment!!”. AI alignment will forever be balancing the tradeoff between paperclip-maximizing and disempowerment, because these two rival concepts of alignment are fundamentally incompatible.

Utilitarianism has always been a useful moral heuristic. But as a moral theory of everything, it has always been fundamentally inadequate. Modern American society, and the economic/libertarian thinking that has guided it, have generally managed to paper over those inadequacies, and we’ve more or less muddled through. But as technology becomes increasingly powerful, this is becoming less and less tenable. Just as Effective Altruism eventually wanders into la-la land, the utilitarian system America has built is starting to crack as technology becomes too extreme for the old assumptions to hold.

We will never replace utilitarianism, I think, but it’s become clearer that we need to buttress it somehow — to diversify our heuristics to include alternate notions of human flourishing. Just what those should be, of course, is a much harder question…

How we can beat China

2026-09-20T01:51:30+00:00

In 1946, George Kennan, who was acting as the head of the U.S. Embassy in Moscow, wrote a message to the Secretary of State, known as the “Long Telegram”. In the Long Telegram, Kennan laid out a theory of how Soviet power could be defeated. He argued that Soviet leaders would always be hostile to the U.S., but that they were also cautious — they wouldn’t recklessly start a world war like Germany had. He also argued that the Soviet system was weak and unstable, and that given enough time, it would collapse from its own internal contradictions. This became known as the doctrine of “containment”, and it ended up being exactly the strategy that the U.S. used to win the Cold War. It took 45 years for the strategy to work and the USSR to fall.

I am certainly no George Kennan. I’ve never lived in China, I haven’t been there in many years, I don’t speak Chinese, and I certainly don’t talk to any leaders there. I’m also not a career diplomat or a professional geopolitics guy of any kind. So although I’m going to write a blog post with the same general goal as Kennan’s Long Telegram, I don’t want to suggest that it’s anywhere near as authoritative or insightful.

But I’m still going to write a post about “how to beat China”, because I think it’s useful to think through these sorts of big strategic questions. This post is also an important follow-up to a series of posts I wrote about geopolitics during the Biden years, about U.S.-China competition:

(There were more, but I figure your appetite for clicking through old posts is limited.)

The basic premise of those posts was that the competition between a China-led New Axis and the U.S. and our old allies — Europe, Japan and Korea, hopefully India, etc. — would define the upcoming era of geopolitics. And my conclusion was that we should use an updated version of the basic strategy we used against the USSR in Cold War 1. I recommended that we should build up economic power within our bloc of allies, establish military deterrence, and wait for China’s internal contradictions to do the heavy lifting of removing the threat.

That strategy was always going to be harder than it was in Cold War 1, because China A) is a lot bigger than the USSR, B) has a better economic model, and C) is a lot more globally integrated. The strategy I recommended was much tighter economic cooperation between the U.S. and a maximally large set of allies, to build up a bloc capable of overmatching even the mighty Chinese manufacturing juggernaut. This is the approach that’s being promoted by experts like Rush Doshi and Kurt Campbell. If you’re interested in reading about how that approach would work, and why it would be helpful, I recommend Doshi’s interview with ChinaTalk:

Anyway, I still think building up an anti-China economic juggernaut would be a great thing to do, and I still think the U.S. would benefit from closer economic integration with its friends and allies. But I think the election of Trump shows why this strategy probably isn’t going to work. The United States is simply too internally divided to engage in the sort of far-seeing, purposeful, smart kind of international competition that we pursued so effectively in the 20th century. Americans care more about fighting other Americans than about fighting the Chinese, and that state of affairs will persist for a while.

Donald Trump’s presidency is an outgrowth of America’s internal divisions; he cares much more about fighting his own internal enemies than about prosecuting international rivalries. He has willfully degraded or torn up America’s key alliances, especially with Europe, and created enmity where friendship once reigned. He has flagrantly trampled on every principle of human rights, mutually beneficial economic cooperation, democracy, and international non-aggression that the post-WW2 U.S.-led international order fought to uphold. He has, in short, made it impossible to prosecute the sort of Cold War 2 strategy that I spent the Biden years advocating.

This doesn’t mean America’s alliance system is necessarily dead forever. As Doshi and Campbell pointed out in a recent article, the old alliances could be reconstituted, if America credibly overcame its internal divisions and returned to being the mostly unified, stable, reliable great power that it was last century. But overcoming those divisions is going to take decades at the very least, and that’s exactly the window of time during which China will be at its peak of relative power.

So I think we’ve basically missed the chance to use that strategy effectively, and I’ve recalibrated my thinking. This post is meant to explain my recalibration. I do think there’s a path to defeating Chinese power, but it doesn’t look like what I previously had in mind.

Who is “we”, what does it mean to “beat” China, and why should we even care?

Before getting to the question of how to beat China, we need to answer the questions of who and why should do this “beating”. In 2024 the answer was obvious — the U.S. and its erstwhile allies still represented a far more liberal power than China and its allies, so ensuring that the former was more powerful than the latter would further the cause of global liberalism. China threw minorities in concentration camps, threatened its neighbors with invasion and war, supported Russia’s invasion of Ukraine, and repressed free expression all over the world. That was — and is — behavior worth resisting.

But the election of Trump changes that calculus. Trump threatens other countries with invasion, and sometimes even attacks them if he thinks they’re weak enough. He can’t throw minorities in camps just for being minorities yet, but he’s building a sort of gulag archipelago under the fig leaf of immigrant detention. His foreign policy is increasingly based on “civilizational” concerns instead of human rights or democracy.

Trump has thus forfeited America’s claim to moral superiority over China. It’s true that Trump and his ragtag movement are far less competent than the Chinese Communist Party, but that doesn’t mean that increasing American power will lead to a more liberal world. At this point, it probably won’t. And of course Democrats will try to undo some of Trump’s geopolitical shifts, but if past experience is any guide, they will have only marginal success. Moreover, the GOP is not going to revert back to Reaganism after Trump leaves office; they will keep fighting for their illiberal visions both at home and abroad for a long time to come.

None of that means, however, that liberalism as an idea is dead. Europe, Japan, and many other countries that were on America’s side in the Cold War still uphold the basic ideas of democracy and human rights, as do many developing nations. The average person living in Tokyo or Paris or Bangalore still enjoys far more personal freedom and dignity than does the average person in Shenzhen or Shanghai.

And even leaving aside the relative merits of different social and governmental systems, there are plenty of people in plenty of countries who simply don’t deserve to be trampled on by the bootheel of the CCP. Japanese people, Indians, Filipinos, Koreans, and so on deserve to be able to determine their own destiny without being subject to the whims of the powerful men in Beijing. And Ukrainians, Estonians, Poles, etc. deserve to be free from the Russian tyranny that China quietly backs.

So when I say “how we can beat China”, what I actually mean is how we can keep China from dominating other countries. This doesn’t mean we should try to make China collapse, or impoverish the Chinese people, etc. Instead, it just means how smaller nations — and nations who value human rights and democracy — can resist Chinese power.

And at some point, if the U.S. ever does manage to tamp down its internal divisions and return to being a liberal power, it will still mean the same thing.

Many hegemons eventually beat themselves