Why was the peak oil theory so widely believed?

Why is rent control so popular?

Because we suffer from the sleepwalking illusion: we keep missing how people adapt to changing circumstances. Many failed to see that programmers would fix the Millennium Bug, or that oil companies would increase production through new extraction techniques. And rent control supporters overlook that developers respond by building less. We often see other people as more static than they really are. As I will show in a series of posts, this illusion distorts our entire worldview.

I’ll start by explaining the illusion’s psychological roots: the rationality gap between actors and observers. As actors, we usually respond fairly rationally to new situations. But as observers of other people, we often fail to see this. We don’t live up to the same standards.

One reason is that it’s hard to put yourself in the shoes of someone trying to solve a problem like the Millennium Bug. As the situation unfolds, they often discover new possibilities that outside observers can’t imagine. While actors think in concrete near mode – where the world is full of details – observers think in abstract far mode, where most of these details are lost.

The other reason is an asymmetry of attention. Actors typically focus on a single issue they really need to get right. The programmers who fixed the Millennium Bug knew they could be held personally responsible for failure. By contrast, observers typically track many events and pay a much smaller price if they get some of them wrong. The Millennium Bug doomsayers could move on with little consequence.

Subscribe now

Everyday rationality

The rationality gap is pervasive in ordinary life. Joining a fleeing crowd on a station platform is often viewed as a textbook case of human irrationality. But other people running is actually evidence that something is wrong. People who might be in danger grasp this logic, while outside observers often dismiss it as senseless panic.

Another example is advice columnists telling people to stop hiding stress at work. In fact, it’s often sensible to keep quiet, since many firms don’t want to promote people who risk getting overwhelmed. But while employees are rightly worried about this, advice columnists have much less at stake and often choose the warm glow of recommending courage.

Or consider the debate about credentials. People often scoff at the focus on prestigious degrees, but there is a rationale. In the absence of more reliable information, they’re an imperfect but useful signal of competence. Recruiters know this, and job applicants know that they know it. But many onlookers don’t bother to think it through, instead pushing the populist story of vain status-seeking.

Unfortunately, psychologists have lent support to this overly pessimistic view of human decision-making. Part of the problem is academic incentives, as journals have long favoured findings of bias and error. Many of these findings were overturned in the replication crisis, but public discourse has been slow to catch up. The idea that we’re largely irrational lives on.

Institutional rationality

If we underestimate other people, we underestimate our institutions even more. When we discuss society, we often fall into a cynical mood and don’t even try to describe it fairly. For instance, many people dismiss politicians as much less competent than ordinary people, contrary to what the evidence suggests. I think this dismissal arises from that cynical mood, not from a serious assessment of the facts.

The media compounds this tendency by focusing disproportionately on institutional failures. While disasters that actually occur dominate the news, those that institutions prevent are all but forgotten. Over time, this gives a badly skewed picture of how well institutions work.

Another factor is that institutions are typically harder to understand than individuals. Many of them are complex and opaque, handling technical questions beyond most observers’ grasp. Our pessimistic biases face little resistance, leaving many people with a level of cynicism that’s completely unmoored from reality.

This isn’t to say that institutions don’t make mistakes – of course they do. But on average, they’re not as incompetent as we intuitively think. If we zoom out and take a broader perspective, we see that Western societies work remarkably well. I’ll soon return to this theme.

I have some conservative instincts myself, but I’m not sure they favour long AI timelines anymore. That might have been the case ten or even five years ago, but things have changed.

• Bio Anchors

• Capability benchmarks

• Revenue growth

• Expert surveys

Subscribe now

Bio Anchors

It’s no accident that the AI timelines debate long lacked empirical grounding. While climate change has a natural metric – temperature – AI progress doesn’t. As a result, forecasts have often relied on intuition.

But in recent years, some researchers have tried to put timeline forecasting on a firmer empirical footing. One attempt that received plenty of attention was Ajeya Cotra’s Bio Anchors report (2020), which plotted compute projections against estimates of the human brain’s compute usage. The model produced multiple forecasts of when it would become feasible to train transformative AI, with a median date around 2052.

Bio Anchors was an impressive research effort, based on real empirical trends. But even so, it was far more theory-laden than climate-style projections. The analogy between brain compute and AI training compute is far from obvious. In addition, the model’s forecasts varied by several decades depending on parameter choices, such as whether AI training was compared to learning over a human lifetime, the evolutionary process that produced human intelligence, or other biological processes. It wasn’t solid evidence by the standards of epistemic conservatism.

Capability benchmarks

A more direct approach is to estimate AI performance on suites of relevant tasks. For instance, METR tracks the length of tasks AI can complete, as measured by the time they’d take a human expert. According to their most recent estimates, this time horizon is now doubling every three months. If this trend continues, AI could be doing tasks that take humans a month within a few years.

Most people would agree that METR’s method involves fewer contestable assumptions than Bio Anchors. Instead of looking at inputs and biological comparisons, METR focuses on outputs: what AI systems can actually do. This is much closer to what the epistemic conservative wants.

METR’s work has generally been well received, but it also has its limitations. To facilitate their evaluations, the problems they study are unusually well-defined. It’s not clear that the results generalise to the messier, more open-ended tasks of real-world jobs. Even some of METR’s own researchers acknowledge that this is an important issue.

Revenue growth

But I think there are even less theoretically loaded reasons to think AI timelines won’t be very long. People are paying more and more money for AI. Plausible extrapolations of this revenue growth provide arguably the most direct empirical case that AI will become a large share of the economy within the next decade.

Some sceptics suggest that this growth merely reflects hype – that AI isn’t as valuable as the numbers suggest. But that is unconvincing. Claims about hype and bubbles carry more force when directed against valuations and investments than against actual consumption. It’s not particularly conservative to assume that people who use AI on a day-to-day basis are mistaken about its value.

While we cannot rule out that growth will taper off, the current momentum does seem incredibly strong. And it converges with other evidence like the rapid benchmark improvements. Very long timelines would require revenue growth to slow dramatically. I think those who claim that have the burden of proof.

Expert surveys

But there’s another kind of evidence that’s important for epistemic conservatives: expert surveys. Some of them suggest that a transformative economic impact is still many decades away. In a recent survey by the Forecasting Research Institute, most respondents thought that AI would only increase economic growth fairly modestly. The economists surveyed projected just 3.5% annual growth by 2050, even assuming rapid AI progress to 2030 – AI outperforming top humans in research, coding, and leadership, producing award-winning creative work, and handling nearly all physical tasks. Likewise, they expected the labour force participation rate to be 55%, down only slightly from today’s 61%. And while the survey’s AI experts predicted a greater economic impact, they also believed that most people would still be working by 2050 even under this rapid scenario.

I’m generally a fan of expert surveys, but there are some reasons to interpret these results carefully. As I’ve previously discussed, respondents may not have fully internalised the rapid progress scenario when answering questions about its economic impacts. Relatedly, I think they simply haven’t thought very much about the impact of AI on economic growth. It’s not clear they’re experts in the same sense as climate scientists who are asked about future warming. Therefore, I think epistemic conservatives should put less weight on FRI’s results than on extrapolations from benchmark and revenue trends.

*

This doesn’t mean that long timelines can be ruled out. Besides new technical obstacles, political intervention could delay AI progress. I take this possibility seriously, and plan to return to it. But the point isn’t that the trend couldn’t break – it’s that this is hardly a conservative position. The AI sceptics can no longer dismiss short or medium timelines as speculation.

Subscribe now

• Women don’t think 80 percent of men are below average

• The share of children with a secret dad is far below 10 percent

• Why so few bootmakers lost their jobs during the Industrial Revolution

• Europeans spend more time on household chores than Americans

• Why Britain stopped building nuclear power

• In brief: the history of wargaming, how to fix drug development, and more

Women don’t think 80 percent of men are below average

You might have seen this chart by Christian Rudder, founder of the dating site OkCupid.

While men’s ratings of women roughly follow a normal distribution, women’s ratings of men are decidedly skewed. More than 80 percent of men are rated below average.

This chart tends to go viral whenever someone shares it on social media. It fits a stereotype many people find both funny and true.

But let’s look at other studies. In the social science literature, men do tend to rate women higher than women rate men, but the difference is usually much smaller than the OkCupid chart suggests. Take this 2003 study of American participants, where women’s ratings of men sit almost exactly at the midpoint:

The US General Social Survey doesn’t distinguish by rater sex, but the overall ratings rule out women finding 80 percent of men unattractive.

So do the ratings in an equivalent German study of 20–30-year-olds:

It’s not clear why the OkCupid chart looks so different, but the overall body of evidence shows that it hugely exaggerates sex differences in attractiveness ratings.

Subscribe now

The share of children with a secret dad is far below 10 percent

Since people enjoy juicy stories, they often spread far beyond what the evidence supports. Another example is the widely believed claim that one in ten children has a different father than they think. This case is slightly different because some scientific studies have actually reported such numbers, or even higher. But many of those studies were based on biased samples, such as people who already had doubts about paternity. In rich countries, more representative studies find much lower rates of misattributed paternity. And they have likely fallen over time, partly thanks to better contraception. A Swedish study of nearly two million families based on blood group data found that the rate of misattributed paternity fell from around three percent in the 1930s to about one percent by 2010.

Why so few bootmakers lost their jobs during the Industrial Revolution

The Industrial Revolution increased productivity, reducing the need for many types of workers. But in a paper on the British bootmaking industry, Hillary Vipond shows that this didn’t necessarily result in many individuals losing their jobs. While much artisanal bootmaking disappeared, the decline was slow enough that most existing bootmakers could continue working in their trade. Instead, the main driver of the shift was young people going into other jobs.

Some think this is how AI will play out, but I think we should be careful with this analogy. One issue is the speed of AI progress: if an industry automates quickly enough, existing workers are bound to be affected. Another is that there might not be other jobs to move into. There’s no reason AI couldn’t eventually do all the jobs humans can do.

Europeans spend more time on household chores than Americans

A common explanation for why Americans are so much richer than Europeans is that they simply work more. But it depends on how you count. If you consider household labor, the gap in total working time is only two percent, according to a Dallas Fed paper. Higher taxes and transfers incentivize Europeans to do less paid work and more unpaid work at home.

Why Britain stopped building nuclear power

Britain was an early adopter of nuclear power, building reactors faster than countries like the US and France. But construction slowed sharply in the 1970s, and no new reactor has been completed since 1995. In Works in Progress, Alex Chalmers explains what happened. Contrary to what I’d thought, it wasn’t primarily disasters like Three Mile Island and Chernobyl that brought British nuclear to a halt. While they did contribute at the margin, the main cause was a series of dubious decisions in Britain itself: dumping nuclear waste at sea, awarding contracts to failing firms, and hiding the true costs of nuclear power. These mistakes produced political resistance and onerous regulations that made further nuclear construction prohibitively expensive.

Subscribe now

In brief

• In Asterisk, Clara Collier and Angela Chen interview historian Jon Peterson about the history of wargaming and how it gave rise to role-playing games

• On the Works in Progress podcast, Ben Southwood and Saloni Dattani talk to Ruxandra Teslo about why drug development keeps getting more expensive and how to fix it

• Since 1979, American postdoc positions have quadrupled while graduate student enrollment has only doubled

• Danish paper finds that top CEOs’ handling of trade shocks generates value worth over a hundred times their pay

• Soldiers below the previous IQ cutoff mostly performed adequately in the Vietnam War, though not on par with their peers

• Study finds that paying participants real money in standard economic experiments like the prisoner’s dilemma usually doesn’t change their behavior

• Paper argues that estimates of the economic costs of climate change are far more fragile than commonly acknowledged

• AI systems match or exceed human economists at a standard causal inference task

• California’s $20 fast food minimum wage raised prices in the restaurant sector as a whole by 3.5 percent

• American study finds that a ten percent higher minimum wage increases robot adoption by eight percent

• Paper argues that crabs’ sideways walk evolved only once

• Scientists use airborne DNA to monitor biodiversity and invasive species

• Will there be relational jobs in the weird world of advanced AI?

• Why AI propaganda hasn’t been as effective as feared

• Experts don’t expect a robot expansion to transform the economy

• In brief: possible deal between Pentagon and Anthropic, new book on messy jobs, and more

Will there be relational jobs in the weird world of advanced AI?

Economists keep devoting more and more attention to the long-run impact of AI. A few weeks ago, I covered the debate over the Forecasting Research Institute’s survey on AI-fueled growth. Now, one of the contributors to that debate, Alex Imas, has published an essay on how the labor market could evolve as AI becomes more advanced. It’s a model of science communication, using the conceptual machinery of economics in a way that’s accessible to a general audience.

When farming was mechanized, employment shifted to manufacturing. When manufacturing became more automated, it moved on to service jobs. But where will it go if AI can do office work and many other services more efficiently than humans?

Alex’s answer is the relational sector – jobs like teaching, therapy, hospitality, and art. He argues that people are attracted by scarcity, and that even in a world of abundance, one thing remains scarce: genuine human involvement.

He supports his argument with experimental evidence. In one study, Alex and Graelin Mandel found that people’s willingness to pay for a product roughly doubled when others couldn’t buy it. Moreover, this exclusivity premium was cut in half when the product was produced by AI. Alex argues that AI products feel inherently reproducible, making them less valued.

Alex’s essay has received an overwhelmingly positive reception, but there has been some pushback. Philip Trammell argues that Alex ignores a crucial factor: that historically, most of our additional wealth has been spent on entirely new categories of products. Despite centuries of sustained growth, spending on ‘human performers’, as he puts it, remains close to zero. He argues that this pattern will continue in the age of AI, with new goods and services diverting spending away from ‘human-intrinsic goods’.

Alex responds that Philip’s focus is too narrow (quotations have been lightly edited for clarity):

I do not think performers are the right proxy for the category I am describing; I think the example of performers in general here is a red herring. My argument is not that the future economy is mostly about art, performance, or authenticity goods in the narrow sense . . . the relevant category is much broader.

Here’s the full list, from the essay itself:

The durable jobs will be in the relational sector, where the human element is the product itself.

Some already exist and are growing: nurses, therapists, teachers, boutique fitness instructors, personal chefs, bespoke tailors, craft brewers, live performers, spiritual guides, childcare workers, and many varieties of hospitality and care work. Others are emerging: experience designers, human-AI collaboration artists, provenance certifiers, community curators. Many haven’t been invented yet, just as six out of ten jobs people hold today didn’t exist in 1940.

Trevor Levin quote-tweets a screenshot of this list, arguing that many of the jobs aren’t as relational as they seem:

Extremely unclear to me that ‘the human element is the product itself’ for almost any of these examples.

Trevor’s point is that we’re more interested in quality for its own sake than Alex suggests. We wouldn’t stick with human-made products if robot beer and chatbot therapy were better and cheaper.

In the comments, Alex clarifies that he doesn’t see all these jobs as purely relational:

In many of these sectors, a large share of what is currently bundled together as ‘services’ consists of tasks that are not themselves relational, and those tasks may well be automated away. But that does not imply the disappearance of the job category. My argument is that the remaining relational component will become the main value proposition of those jobs, and these services will now provide a higher-quality product since people can focus on the relational aspect of the job (see the O-ring model for why).

This is an interesting argument, but I’m not sure it shows that there will be many relational jobs. Let’s look at Jason Abaluck’s succinct yet thorough critique:

I liked @alexolegimas’s essay too, but I think the argument smuggles in a lot of likely false assumptions about the stability of preferences in a wildly different environment.

‘Chefs, tailors, craft brewers, experience designers’ are all jobs which are valued principally for their outputs. These jobs won’t exist at large scale in a future of full automation where machines can make any food essentially free.

People in the status quo value art more if a person made it, but that will change in a world where machines have all the capabilities of people and more. It is an equilibrium outcome, not a permanent law of nature (LLM today = your desktop, not a brilliant companion).

Things like live performance and sports would still exist, but these are niche. The world of the essay might persist for a few years during a transition period but I think it’s not a stable outcome.

And this is just about how humans would change organically when the equilibrium changed – once we are in a world where most jobs are in the ‘relational sector’, we are also likely in a world with machine gods who can alter human biology.

With full automation, people won’t have jobs and that’s okay – they won’t need jobs, and they’ll find other sources of value in their lives. Maybe they’ll link up with the machine intelligences or use far-future drugs to achieve previously inconceivable experiences.

Essentially, Jason gives three arguments:

1. Many of the jobs Alex lists have a smaller relational component than he suggests (see Trevor).

2. A small fraction are indeed purely relational, but since superhuman AI would dramatically shift consumption patterns, demand for those jobs could still collapse (see Philip).

3. Superhuman AI could reshape the human brain, drastically reducing our interest in relational services.

I agree with all three, as well as Jason’s broader point: Alex describes a transition period, not where we’re eventually heading. This is a common issue in the debate on advanced AI. Many participants underestimate how deeply it would change the world. Jason’s argument about altered preferences might seem outlandish, but I think it’s entirely plausible. Witness the immense popularity of GLP-1 drugs, which change our preferences for eating. In my view, they’re a pale forerunner of what’s coming.

Subscribe now

Why AI propaganda hasn’t been as effective as feared

In 2021, Daniel Kokotajlo wrote ‘What 2026 looks like’: an AI scenario predicting rapid progress and growing societal effects. Since it’s now 2026, we’re in a position to evaluate his predictions. Many people think he got a lot of things right, especially about technical milestones. But in an interview with Asterisk’s Clara Collier, he concedes that he overestimated the effectiveness of AI propaganda.

Clara:

You have this story that unfolds through 2024 through 2026 where there’s a vast increase in the production of AI propaganda, more sophisticated use of highly partisan AIs to create online filter bubbles, and much more AI-enabled censorship.

. . .

AI has been a hugely influential technology, but it has influenced the information environment a lot less than many people – not just you – anticipated. Of the people who were making bold predictions about AI in 2021, I think a lot of that was about misinformation and censorship. And it’s interesting that this angle, which even relative AI skeptics were concerned about, seems to have had less of an impact than many of us anticipated. Why do you think that is?

Daniel:

Normally when people ask me this question, I say that governments, corporations, political parties and campaigns, and other powers that be have been slower to adapt and push this technology for evil purposes than I feared they would be. But actually, I think that might not be true. I’m kind of confused and want to think about it more.

On a technical level, I think all the things I talked about – like the use of AI for A/B testing – are totally possible. I have heard anecdotally that some of them are being done. Someone at OpenAI told me that in their opinion, the product teams are basically training ChatGPT to maximize user retention and probability of upgrading subscriptions.

. . .

Maybe what I got wrong is that in 2021, it felt like a lot of tech companies were on the left explicitly and were putting their thumbs on the scales to influence discourse through their moderation decisions. It felt like if that trend continued, they’ll be using AI to do it in five years, and it’ll be more effective. And then I predicted there would be this counter-backlash because there are people who don’t like the left who want their own space, and so they would create their own spaces.

I agree with Clara that many people worried prematurely about the threat of AI propaganda and misinformation. In 2019, OpenAI cited these concerns as a reason for delaying the release of GPT-2.

People would have been wise to listen to psychologists like Hugo Mercier, who has long argued that humans aren’t nearly as gullible as the misinformation discourse suggests. And there’s the problem of getting people to listen to your propaganda, as he points out on the 80,000 Hours podcast:

There’s already an essentially infinite amount of information on the internet. So the bottleneck is not how many articles there are on any given topic, because there is already way more than anybody will ever read. The bottleneck is people’s individual attention – and that bottleneck is largely controlled by, to some extent peers and colleagues and social networks, but otherwise mostly by the big actors in the field: by cable news, by big newspapers. And there’s no reason to believe these things are going to change dramatically. So having another thousand articles on a given issue, just no one is going to read them.

Some of the predictions about AI propaganda have been psychologically and sociologically naive. Folk theories are often far too cynical about human capabilities and motivations.

This cynicism also leads many people to underestimate our epistemic norms. AI companies that expose people to mass persuasion would face a major backlash, meaning it’s in their interest to develop models that are broadly truth-seeking. Consequently, I share Dan Williams’ optimism about the epistemic impact of AI.

Experts don’t expect a robot expansion to transform the economy

The Forecasting Research Institute has released a new set of surveys on the future of robotics, following up on its recent paper on the effects of AI on economic growth. On average, experts expect autonomous robots to pass the Coffee Test – reliably making coffee in an unfamiliar kitchen – by 2034, and to remove an appendix without human assistance by around 2040. They predict that the global industrial robot stock will more than triple by 2040, but don’t think this will have a transformative economic impact. Workers’ share of economic output is forecast to decline, but no faster than the historical trend.

Subscribe now

In brief

• Donald Trump says a deal between the Pentagon and Anthropic is still possible

• Luis Garicano has a forthcoming book, Messy Jobs

  • See my previous note on his messiness heuristic for keeping your job in the age of AI

• Anthropic introduces identity verification for certain use cases

• Oliver Habryka on ten non-boring ways he has used AI in the last month

• Democrat Alex Bores launches plan to incentivize hiring humans and pay workers displaced by AI

• Greg Brockman says OpenAI fell behind Anthropic on coding by failing to ground its models in messy real-world codebases

• Benjamin Todd argues that AI agents are unlikely to be aligned by default

  • Séb Krier argues several of Ben’s arguments are out of date

• OpenAI’s roon argues that an AI pause would be squandered since we ‘deal with problems as they come up’

  • Eliezer Yudkowsky responds that roon’s strategy will produce catastrophic errors, leading to human extinction

• Facial recognition to be rolled out across Britain

• Unauthorized users have accessed Anthropic’s restricted Mythos model

• Sentinel forecasters put the chance of a cyberattack causing a 27 percent S&P 500 fall within a year at 1.6 percent

• Eliezer Yudkowsky and Nate Soares’ If Anyone Builds It, Everyone Dies (September 2025) has sold 150,000 copies in English alone

  • It might overtake Nick Bostrom’s Superintelligence (2014), which had sold 250,000 copies by 2020

Will AI drive humans off social media?

On networks such as X (Twitter), an increasing share of content is written by AI. Replies to big accounts are often dominated by bots. Ethan Mollick argues it could spell the end of traditional social media (quotations have been slightly edited for clarity):

Human interaction is going to shift to discords and group chats, invite-only. The open web and social media are going to be left for the agents lurking amongst the ruins. Everything public will be Moltbook.

I suppose social media companies can figure out a way forward to ensure an element of human interaction, but they haven’t seemed to be able to do that, yet.

Maybe we will all go back to bowling leagues and masonic lodges.

Email is also affected. Arvind Narayanan:

AI agents seem to be killing the ability to do cold outreach. Look at it from my perspective – someone who gets dozens of unsolicited invitations/requests in my email from strangers every day. In the past there used to be a sharp distinction between unpersonalized mass emails and those that were meant specifically for me. My spam filter could mostly automate this classification, and even when it didn’t, it would take me only a few seconds to tell if an email was worth reading/responding to. Now the distinction is gone because it is trivial to use AI to send out mass emails that appear deeply knowledgeable about each recipient’s work and expertise. I guess we have to go back to the bad old days of needing to be ‘introduced’ to someone before communicating with them.

Alex Imas comments:

I’ve been beating this drum: by making traditional signals of effort/information obsolete (e.g., writing a thoughtful email), AI will *increase* the value of social capital and networks.

This is not a good outcome: information technology was supposed to ‘flatten’ interactions. But I don’t see a solution on the horizon.

Abi Olvera:

The ease of applying to jobs with AI is also changing the nature of hiring.

I’ve seen more places do hiring rounds privately, only by word of mouth. This was always happening, especially for jobs that want people within specific networks, but now it’s happening more.

Employers are reacting to the lower friction cost of applying to jobs.

I agree that AI will reshape social and professional interactions in a number of ways, but I think its impact will vary across cases. It’s useful to distinguish between two types. First, there’s evaluation – for instance, of job applicants or students. Since AI makes it easier to fake competence, I think it will definitely change our evaluation processes. As I’ve previously reported, some professors are already returning to in-person exams in response to students using AI on essay assignments.

The second issue is that people are inundated with AI-written content in some contexts. Mollick’s problem isn’t that he can’t distinguish between good and bad accounts – it’s that the flood of AI-generated replies is annoying in itself. But while it’s true that these replies are a nuisance, I’m not sure that they will kill social media like X. Most people with big accounts are using it to promote their own ideas, not to engage in debates with small accounts. While more advanced AI could have a big impact on social media, I don’t think this type of AI bot will.

Subscribe now

In Scandinavia, more intelligent men have more children

What is the relationship between intelligence and fertility? Earlier American studies reported a negative correlation, but demographers have raised concerns that they were based on unrepresentative samples. Recent Scandinavian studies based on military conscription records and national birth registers have found a strikingly different pattern. In a study of virtually all Swedish men born between 1951 and 1967, men with the lowest scores on military cognitive tests had around 0.5 fewer children by age 45 than median-scoring men.

Fertility by age 45 for Swedish men born between 1951 and 1967, by IQ category and eventual parity (total number of children per man). Source: Kolk and Barclay, 2019.

Less intelligent Norwegian men also have fewer children, primarily because they are more likely to remain childless.

Most ChatGPT users are now women

In the early days of ChatGPT, almost all users were men. But the share of women has gradually grown, and recently, they probably became the majority. However, surveys suggest men still use AI more overall.

OECD expects AI to increase productivity the most in the US and the UK

In recent years, productivity has grown significantly faster in the US than in other G7 countries. A new OECD report predicts that AI will enable the US to extend this lead, thanks in part to greater adoption rates. The UK is a close runner-up, while Italy and Japan are expected to lag behind.

Offshore processing wasn’t what stopped the crossings to Australia

Each year, many migrants cross the Mediterranean in small boats, leading to thousands of drownings. To stop the crossings, many European governments want to copy an old Australian policy: processing asylum seekers in other countries. But in a new Works in Progress article, Amelia Wood argues that the decisive policy was turnbacks: the Australian navy escorting migrant boats back to where they came from. Both arrivals and drownings fell quickly after this policy was introduced in 2013. Since then, offshore processing has been quietly wound down, and yet the crossings haven’t resumed.

Subscribe now

In brief

• Data centers don’t produce harmful infrasound

• English landlords are quitting the rental market ahead of the Renters’ Rights Act

• Fish fail at the Monty Hall problem

• Dan Williams and Henry Shevlin interview Robert Long on the philosophy and politics of AI welfare

• Abi Olvera draws on her experience in the US Foreign Service to explain why it’s more apolitical than outsiders realize

• Steve Newman argues that many underestimate the value of AI because they don’t see what iteration and clever setups can achieve

• The Centre for British Progress and Labour Together argue that street votes should be the British Labour Party’s next big housebuilding policy

• The case that some drug trials should be funded only retrospectively, if they succeed

• Coefficient Giving closes its Forecasting Fund

• Swedish study finds that selective conscription hurts the most capable recruits’ labor market outcomes through age 34

• Pentagon asks automakers to steer capacity toward arms production, as they did during the Second World War

• Kelsey Piper argues that education research should bring in economists to fix its lack of rigor

Nature study finds fast human evolution over the past 10,000 years

Are we the same as our hunter-gatherer ancestors, only more technologically advanced? Many think so, but a new study in Nature shows we’re not. DNA analysis of modern and ancient samples finds that natural selection has rapidly reshaped populations across Europe and the Middle East.

Some of the changes are unsurprising, like lighter skin, which aids vitamin D synthesis at higher latitudes. But how come genes associated with smoking became rarer long before tobacco reached the Old World? They probably correlate with other traits that reduced reproductive success, such as impulsivity.

Many changes look like improvements. Genes associated with schizophrenia and bipolar disorder have grown rarer, while those associated with intelligence, years of schooling, and income have become more common. For several traits, the shift over the past 10,000 years is around one standard deviation, equivalent to the difference between an average person today and someone at the 84th percentile. The genetic gap between us and our ancestors is substantial.

Subscribe now

French electricity is five times cleaner than the EU average, thanks to nuclear

France has long stood apart from its European neighbors for its sustained commitment to nuclear power. While Germany began phasing out its reactors after the 2011 Fukushima accident, this was never on the agenda in France. It still generates more than two thirds of its electricity from nuclear power.

This has given France an electricity system with remarkably low carbon emissions – 20 percent of the EU average.

AI use at work rises sharply with income

A recent survey of American workers finds that the use of AI is strongly correlated with income. Two thirds of workers earning over $200,000 have used AI at work in the previous year, compared with only 16 percent of workers earning under $50,000. Economist Alex Imas comments: ‘AI cannot reduce inequality if this is what the adoption margin looks like.’

Returns to university degrees vary enormously by subject

Does it pay to go to university? It very much depends on what you study. In Britain, medicine graduates at the 10th percentile earn higher net lifetime returns than creative arts graduates at the 90th percentile. Because their earnings are so low, creative arts graduates are expected to repay only a quarter of their student loans.

ChatGPT still has the most AI users, but Gemini is growing

ChatGPT’s share of generative AI website traffic keeps shrinking, according to Similarweb. While OpenAI’s chatbot had 77 percent of traffic a year ago, the latest figure is 57 percent. Its closest competitor is Google’s Gemini, whose share has risen from 6 percent to 25 percent over the same period. Meanwhile, Anthropic’s Claude remains a distant third at 6 percent, even though it has gained ground in recent months.

But some users pay much more than others, and AI companies that attract them earn more than overall user numbers might suggest. Thanks to strong enterprise sales, Anthropic’s annualized revenue is now roughly on par with OpenAI’s. On Ventuals, a small market for private company valuations, Anthropic just passed $1 trillion, ahead of OpenAI at $860 billion. (Google’s market cap is far larger – roughly $4 trillion – but its AI business can’t be valued separately from the rest of the company.)

Subscribe now

In brief

• Niko McCarty and Saloni Dattani round up recent advances in biology and medicine

• Our World in Data is hiring a writer

• In Development launches with an article on the surprising findings of cash transfer trials

• Dylan Matthews reflects earnestly on his 2015 dismissal of effective altruists who prioritize AI risk

• Matt Yglesias on his most right-wing views: that we should punish rule-breakers, that humans aren’t blank slates, and that it’s fine to be rich

• American college enrollment is falling

• In a US study, rainy or hot campus tours made students 8–10 percent less likely to apply

• The case that acetaminophen (paracetamol) is safer than ibuprofen for most people

• Study at a US public university finds that high school grades predict college success better than standardized test scores

  • An earlier study found the opposite at elite private colleges in the US

• The Economist ranks Musk and Bezos as less powerful than Ford and Rockefeller were

• Norway approves driverless bus on regular public route

Growth won’t keep creating new jobs forever – and that’s good, actually

In 2016, Geoffrey Hinton famously argued that we should stop training radiologists. But as radiology employment has since grown, this has become the go-to example for those who think that worries about job displacement are exaggerated. Marc Andreessen argues that Hinton’s mistake is rooted in the view that the number of jobs is fixed:

AI employment doomerism is rooted in the socialist fallacy of lump of labor. It is wrong now for the same reason it’s always been wrong.

Elon Musk responds:

Working will be optional in the future.

Who is right? It depends on what stage of AI development we’re talking about. Here’s Andreessen’s case in more detail:

When a technology increases the productivity of labor or replaces labor entirely in a given task, it lowers the cost of producing whatever that task was part of. Lower production costs mean either:

• Lower prices for consumers (real purchasing power rises), or

• Higher profits for producers (which get reinvested, distributed as dividends, or spent as wages for other workers), or

• Both.

Either way, aggregate real income in the economy rises. That additional real income does not evaporate. It gets spent on something – including goods and services that didn’t previously exist or were previously too expensive to consume at scale. That spending creates demand. That demand creates jobs.

Andreessen is right that people tend to miss these dynamics, and that this makes them worry prematurely about job displacement. But unless we stop AI development, we will eventually reach the stage Musk is talking about, where AI systems are better than humans at all tasks. At that stage, the last step in Andreessen’s argument no longer applies. More demand for goods and services won’t translate into human jobs.

Is talking about job displacement a poor communication strategy?

Musk isn’t the only CEO of an AI company who is talking about job displacement. In response to an interview with Sam Altman, Noah Smith suggests this is a poor communication strategy:

The heads of the big AI labs continue to insist that their products are going to take all your jobs, and also pose various catastrophic risks.

I take a different view. Sufficiently advanced AI does have these implications, and telling the world about them ahead of time is clearly the right thing to do. And from a purely self-interested perspective, trying to downplay them would probably backfire. More and more people are figuring out where AI is heading, and an AI leader who tries to hide it would look disingenuous.

But the deeper problem with this debate is that most participants view job displacement as a bad thing. It’s true that a transition to a world without work would require extensive political reform. But automatic abundance should fundamentally be welcomed. That’s another thing AI leaders should communicate.

Subscribe now

The case for egg freezing

Women are having children later than ever, leaving many struggling to conceive. But there is a solution: egg freezing. In Works in Progress, Luzia Bruckamp and Ruxandra Teslo write that it works far better than widely believed. They recommend women freeze around 20 eggs in their late twenties or early thirties, arguing it may give a cumulative 85 percent chance of a live birth.

The share of British students who get a First has quadrupled since the 1990s

Grades at British universities have risen sharply over the past three decades, even though there’s little evidence that students are actually learning more. To reverse the trend, some people have suggested limiting the share of Firsts that universities can award. In a new paper for the Higher Education Policy Institute, Tom Richmond proposes a 15 percent cap. A similar cap is under consideration at Harvard.

Leapfrogging landlines for mobile phones

When countries develop, they don’t always need to repeat every step the technological leaders took. Sometimes they can skip a step: what’s called ‘leapfrogging’. Landline networks require expensive physical infrastructure, and many countries in sub-Saharan Africa never built them. By contrast, they did build cheaper mobile networks once the technology arrived. As a result, phone access has surged since the turn of the millennium.

Are changing relative prices driving the loneliness crisis?

In recent years, people have started to spend more and more time by themselves. Why is that? Alex Mayyasi argues that a major driver is changing relative prices. Technology has made solitary digital entertainment cheaper compared with social experiences like restaurant visits and concerts.

I think his argument has some appeal. While many accounts of the loneliness crisis treat people as helpless or compulsive, Mayyasi shows that they’re simply reacting to changing incentives. But my explanation would be a bit different. First, an important factor Mayyasi doesn’t discuss is convenience. Many people who can afford to go out don’t do so because it requires more effort than staying home scrolling Instagram or X. Second, I think that the key change is that digital entertainment is simply more fun than TV and other predecessors. While Mayyasi does mention this factor, I would put it front and center.

Climate change is increasingly polarizing in the US

Though climate change receives less attention than it used to, Americans haven’t grown less concerned. The share of people who worry a great deal about it is 44 percent, close to the peak.

But this trend hides a growing partisan divide. Currently, 72 percent of Democrats worry a great deal about climate change, whereas only 6 percent of Republicans do.

Subscribe now

In brief

• In a Michigan study, baby bonuses induced an additional birth for every $107,000

• Recessions have grown rarer, but pundits overpredict them

• Study finds that home staging adds ten percent to sale prices

• Former allies Trump and Meloni fall out over the Iran war and the Pope

• More Russians are moving to Belarus

• Study finds that AI is rated as better than humans at fact-checking on X

• Maine becomes the first US state to pass a data center moratorium

• Political scientists increasingly cite journals over books

• Endpoint Arena, a prediction market for clinical trials

• Tibor Rutar argues that strong kinship structures hold back economic development

• Emergent Ventures grants, for writers, creatives, and intellectuals

• Reported agreement in principle to extend the ceasefire between the US and Iran

American college costs are falling fastest for the poorest families

Many people believe that the cost of college is spiraling in the US, but as I’ve previously reported, this isn’t true. Net prices after financial aid are falling.

But there is a further reason the conventional wisdom is off the mark. The costs have fallen the most for those who are least able to afford them.

At the richest private colleges, families earning $250,000 now pay almost twenty times as much as families earning $45,000.

These numbers run counter to much of the discourse about American colleges, not least here in Europe.

Subscribe now

Hungary sees a democratic transition

Many worried that Viktor Orbán’s hold over Hungary was so complete that he would be impossible to unseat as prime minister. But in Sunday’s election, Péter Magyar won in a landslide and Orbán quickly conceded. Democracy is not as fragile as the alarmists claim.

Switzerland’s constitutional referendums have a populist slant

While referendums initiated by citizens have long been a feature of Swiss democracy, they’ve become much more frequent in recent decades. One reason is that campaigns to collect the required signatures have become more professionalized. It also helps that the Swiss population has grown by almost 50 percent since the 1970s, while the thresholds to force referendums have remained fixed.

Several types of referendum can be initiated by citizens – including the vetoing of new legislation – but the most consequential are popular initiatives to amend the constitution. While a large majority of these initiatives are rejected, the pass rate has risen. Since 2000, 14 popular initiatives have been adopted. Most of them are decidedly populist. (Translations from Wikipedia, slightly edited.)

• For Switzerland’s accession to the United Nations

• Life imprisonment for untreatable, extremely dangerous sexual and violent offenders

• For GMO-free food

• No statute of limitations for pornographic crimes against children

• Against the construction of minarets

• For the deportation of criminal foreigners

• End the unchecked construction of second homes!

• Against rip-offs (on executive pay)

• Against mass immigration

• Pedophiles should no longer be allowed to work with children

• Yes to a ban on full facial coverings

• The Nursing Initiative (more support for nurses)

• Yes to protecting children and youth from tobacco advertising

• For a better life in old age (a 13th monthly state pension payment)

On the plus side, cantons with lower thresholds for triggering referendums have higher trust in their governments, according to a recent study.

Banning AI for writing is a bad idea

In the past few weeks, there’s been an intense debate over the use of AI for writing. When Megan McArdle said she uses AI to do research, transcribe interviews, and suggest edits, some reactions were scathing. Philosopher Ben Burgis:

In a healthier media culture, an admission like this would at the very least get her fired.

Journalist J Dale Shoemaker:

I think it should be deeply embarrassing for any journalist employed by a news organization to admit to this level of AI use.

In academia, hostility to the use of AI is widespread. Sociologist Nathan Kalman-Lamb:

I’m sorry, but it is disgraceful to be an academic who uses this technology to conduct research. It should be prohibited in all of our scholarly institutions, including universities and journals.

Science communicator Jey McCreight:

Scientists who get caught using AI to write scientific papers should be banned from ever publishing in any journals

Critics are right that AI makes it easy to produce low-quality content that superficially seems credible. But AI is a general tool that can be used in a diverse set of ways. More diligent people are using it to improve their writing through detailed feedback and fact-checks. Prohibiting them from doing so is not a good idea.

Many people assume that you must either be for or against AI across the board: that it’s hypocritical to use it in your own research if you don’t allow it for students. But as political scientist Alexander Kustov points out, this just isn’t true. Every case needs to be judged on its own merits.

We also need to consider how fast AI is improving. I don’t think the association with low-quality slop is going to last. Instead, I think people who use AI will increasingly produce better writing than those who don’t. Calls for blanket bans won’t age well.

Applications are open for Invisible College 2026

Works in Progress is hosting its third Invisible College, a week-long seminar for 18–22-year-olds in Cambridge, UK. The program covers three topics: the origins of modern abundance, obstacles to scientific progress, and the political economy of cities. It’s free to attend and takes place 17th–22nd August. Several past attendees have won Emergent Ventures grants or found jobs in progress circles. Apply by 8th May.

Subscribe now

In brief

• Works in Progress is hiring a social media lead

• OpenAI CEO Sam Altman’s home attacked with a Molotov cocktail

• Saloni Dattani on how public discourse often seeks to explain trends that aren’t actually real

• The Works in Progress podcast on egg freezing, Australian migration policy, ASML, and more

• The American data center boom is delayed by a shortage of electrical equipment

• Patrick Collison argues the YIMBY movement should give more weight to aesthetic concerns about new housing

  • Noah Smith surveys the debate this triggered

• Anthropic overtakes OpenAI on Ventuals, a small market for private company valuations

• American consumer sentiment is at its lowest since records began in 1952

• Predictions of a compute bubble now look mistaken

• Workshop for university faculty on how to teach progress studies in Washington, DC

• In a trial, a new pancreatic cancer drug doubled median survival time

Driverless car travel more than doubles in California

Driverless cars are rapidly becoming more widely used in the US. Waymo has expanded to 11 cities, and in existing markets, travel keeps rising. In California, monthly passenger kilometers more than doubled between December 2024 and December 2025.

Subscribe now

The political salience of AI will only grow with time

As I’ve previously reported, Americans are increasingly concerned about the impact of AI on society. On the Odd Lots podcast, Byrne Hobart said that this makes him worry about misguided regulation, but he also gave a reason this might not happen after all (lightly edited for clarity).

The thing that makes me more optimistic is that the salience of these different issues fluctuates over time. Maybe AI deployment is so fast that it becomes part of the background noise, like the internal combustion engine or electricity or even the internet, which is not a campaign issue right now. It was sort of an issue in 2000, there were internet-y issues in 2016 and 2020, but it’s becoming less and less salient. We just don’t think about it in terms of ‘are you pro or anti the internet?’

This kind of habituation is no doubt very common, but I don’t think it will happen with AI. The reason the internet has faded into the background is that it hasn’t changed much in recent years. By contrast, AI promises to improve rapidly. As it grows more pervasive, so will its political salience.

Electric vehicle sales are surging in Southeast Asia

Southeast Asian countries are highly dependent on fossil fuel imports, which has proved costly during the recent supply disruptions. But their rapid adoption of electric vehicles could reduce this dependence. In Singapore and Vietnam, EVs already have a larger market share than in Europe.

Why Japan’s railways are so successful

More than a quarter of all passenger travel in Japan is by rail. In Works in Progress, Matthew Bornholt and Benedict Springbett explain why. One reason is that driving isn’t as subsidized as in other countries: motorways are tolled and parking is expensive. But the railways also have a smart ownership structure. They’re mostly run by private companies that own not just the trains and tracks but also part of the surrounding land. This gives them an extra reason to grow their ridership, as it raises the value of that land. Conversely, they can increase their ridership by developing the land. A large share of the railway companies’ revenue comes from these developments: everything from retirement homes to amusement parks. It’s a model that other countries could learn from.

A one-person billion-dollar company?

In 2024, Sam Altman predicted that AI would make it possible for a single person to run a company worth one billion dollars. On a recent podcast, he suggested this has now happened. The company is Medvi, an online seller of GLP-1 weight-loss drugs that’s on track for $1.8 billion in revenue this year. However, Medvi actually has two employees and nine contractors. While AI does enable startup founders to do more with less, they’re more reliant on human labor than it might seem.

Subscribe now

In brief

• Anthropic withholds new Mythos model from public release, citing cyberattack capabilities

• Developed democracies virtually never break down

• Major works of literature are written by much older authors today than a century ago

• Hollywood jobs have fallen sharply since 2022

• Large Nature project finds that many studies in the social and behavioral sciences don’t replicate

  • Paul Novosad argues these results aren’t as bad as many say

• Study finds that light skin color arose earlier in East Asia than in Europe

• Gene therapy improves hearing for some people born deaf

• After Anthropic’s clash with the Pentagon, the British government is trying to get the company to expand in the UK

• New OpenAI paper on how to manage economic disruption from AI

• OpenAI puts Stargate UK data center on hold, citing energy costs and regulations

• METR-like study finds that AI is rapidly improving at performing complex tasks in a wide range of domains

  • Another paper applies METR’s methods to cyberattack capabilities and finds fast progress

AI and economic growth, continued

The debate about the economic impact of AI continued over Easter. Quotations have been lightly edited for clarity.

Tom Cunningham:

As far as I can tell, every coauthor on the FRI econ-impacts-survey who’s tweeted about it thinks the respondents are underestimating the impacts of rapid capabilities growth.

Itai Sher gave a critical response:

So now we do surveys and the methodology is if the authors (actually just the ones who tweeted about it) think the survey respondents underestimated something, we treat that as evidence that they did?

Tom:

I’m not dismissing the survey results, but the observation from these authors seemed notable. I personally thought most of the respondents are misinterpreting the questions, or are mistaken, and then it was notable to me that four authors of the paper publicly stated they felt the same way.

As I mentioned on Thursday, I also wonder whether respondents fully understand and accept the scenario with rapid AI progress. It’s hard to survey people on hypothetical scenarios with so many moving parts.

There has also been further debate on the object-level question: how fast growth will be. Andrey Fradkin:

I am more optimistic than my fellow economist forecasters about annual GDP growth by 2050. That said, even my forecast is not as high as the >10 percent forecasts that some are predicting. A few reasons why:

1. The spread of AI capabilities is greatly increasing variance (which includes substantial weight on the negative scenarios including war and revolution).

2. It’s likely that AI will cause a greater decoupling of apparent progress and GDP, so that focusing on GDP will miss something.

3. A lot of economic bottlenecks in the US are regulatory, and in many states of the world, it will take more than 20 years for governance to adjust to let AI subsume these.

4. Healthcare is a large share of GDP, and even though AI will drastically increase medical progress, we may still need RCTs before medicines are widely used and affect GDP.

Nonetheless, the possibility of extremely fast growth cannot be dismissed, and I place about 10 percent weight on these scenarios.

Andrey clarified that his forecast doesn’t assume rapid AI progress by 2030:

Yes, this forecast is unconditional. I believe the capabilities in the ‘rapid’ scenario will almost certainly be achieved within 20 years, just not necessarily by 2030. In the survey, I didn’t take enough time to consider the conditionals (was lazy), but my weight on the > 10 percent annual GDP growth scenarios is much higher in that case (probably 30 percent probability). I still think that my points above apply. Will AI capabilities result in more housing in SF?

Tom Cunningham:

Just to push on this because I feel this is a critical ambiguity in all this discussion: if you *literally* believe that AI can do all intellectual work, including robotics, then things get crazy, right? Like outputs of goods and services will double every year? Housing in the Bay Area seems totally irrelevant somehow. Maybe I need to write something more concrete about this.

Andrey:

Does GDP make sense in that world? Housing is a major component. I still think humans will want to live in SF. But would love to read anything you write on this topic.

Ajeya Cotra:

Rather than talking about GDP we could consider, e.g., energy production or production of robots in that world (though I think GDP would still likely be meaningful and grow enormously like it did during the Industrial Revolution).

Jason Abaluck also had a thread about how transformative AI could make GDP a less useful metric of economic progress:

In worlds where you get full automation and a production explosion, GDP ceases to be a decent proxy for welfare. We should really be forecasting other metrics in those worlds.

. . .

If automation drives the price of many or most goods to zero, then GDP growth will be a very bad proxy for welfare growth.

If robots make a free pill that makes you live forever, GDP does not increase. If robots send whatever food you want to your doorstep for free, GDP does not increase.

So really, forecasts about GDP in a world with complete automation are irrelevant. What matters is forecasts of welfare. Makes me think that next time I coauthor a paper asking partly about such futures, we should use alternative measures of growth!

This measurement issue is interesting, but I don’t think it explains why respondents predicted relatively muted economic growth. The fact that they thought that labor force participation rates would remain fairly high even in the scenario with rapid progress suggests they simply don’t expect a radically transformed economy by 2050.

Subscribe now

Why does the US take AGI more seriously than Europe?

As a European, I’m struck by how much this debate is dominated by Americans. Almost all contributions I found came from people working in the US.

The same pattern holds in broader public discourse, as far as I can tell. In the US, the prospect of transformative AI is taken very seriously. Bernie Sanders recently said that ‘we are at the beginning of the most profound technological revolution in world history’. I don’t hear that kind of language from European politicians. While there obviously is a debate about AI here, too, it isn’t as ‘AGI-pilled’. Why is that?

1. AI adoption

One possibility is that the difference is driven by Americans’ greater use of AI. However, I doubt it’s the main explanation, since adoption rates are fairly high in some European countries, too.

2. AI leadership

Since all leading AI companies are from the US, Americans may believe they can shape how AI develops. By contrast, few hold this belief in my own country, Sweden.

3. Cultural openness to speculation

Another possibility is that Americans are simply more willing to entertain speculative questions. I think this may well contribute.

4. Safety nets

It’s also been suggested that a thinner safety net may make Americans more worried about AI leading to job losses. But my sense is that this factor is somewhat exaggerated, since unemployment hurts in Europe, too.

5. Scale of public discourse

Lastly, the US is such a large country that there’s room to debate a range of issues at the same time. Since European countries are much smaller – and since there’s no pan-European public discourse – a topic like transformative AI can more easily be sidelined.

There may be further factors I’ve not considered. Please comment if you have a view.

American birth rates are falling much faster in cities

The recent fall in American birth rates is disproportionately an urban story.

The relationship between fertility and density was weak as recently as 2005, but since then it has grown much stronger.

In some cities, the trend toward fewer children has accelerated since Covid. Manhattan had 19 percent fewer 0–4-year-olds in July 2024 than it had in April 2020.

Most Y Combinator founders are now under 25

Startup founders in the elite incubator Y Combinator are much younger than just a few years ago, according to Rebel Fund estimates. This may be related to the rising share of AI startups.

American electricity consumption is expected to rise

Several forecasts predict that American electricity consumption will increase again, after having remained flat for many years.

You might think it’s all about AI and data centers, but actually, a broad range of sectors are expected to contribute.

Subscribe now

In brief

• Saloni Dattani on why she’s skeptical of accidental discoveries. (See also the good discussion in the replies.)

• Polymarket gives Viktor Orbán’s Fidesz a 31 percent chance of winning Sunday’s Hungarian election

  • Sentinel gives him a 44 percent chance

• The Burn Selection Hypothesis: humans may have adapted to fire injuries

• Younger Democrats always disapproved more of Biden’s foreign policy, but the Israel–Gaza war widened the gap

• Daniel Kokotajlo and Eli Lifland shorten their (already short) timelines to advanced AI

  • As does Ryan Greenblatt

• Helen Toner argues that the term ‘artificial general intelligence’ has become almost useless

• Two METR researchers present a model of how AI could accelerate its own development

• Contamination from laboratory gloves may have inflated some estimates of microplastic prevalence

• Paper finds that censuses disproportionately undercount poor people, both within and across countries

  • Population estimates are often wrong for other reasons too

How much will AI increase economic growth?

Elon Musk and Dario Amodei predict that AI will transform the economy within the next few years. But a new paper from the Forecasting Research Institute finds that most experts disagree. By 2030, they expect AI progress to be slow or moderate.

And while they predict that growth will be higher if AI progress turns out to be more rapid, they don’t think it will take off dramatically. The economists are the most conservative:

This paper triggered a lively debate on X. Quotations have been lightly edited for clarity.

METR’s Tom Cunningham:

This is a great paper but contains a puzzle: forecasters expect even if we automate most labor and wait 20 years, GDP will only increase by 45 percent.

I would love to hear how people are thinking about this.

. . .

Specifics: the median prediction for 2025–2050 GDP growth with ‘rapid’ progress is 1.5 percent above the slow scenario, meaning that after 20 years of diffusion and AI progress GDP will increase above trend by 45 percent (=1.015²⁵).

Some comparisons for a 45 percent increase in GDP per capita:

• This is roughly the difference between US and Canada in GDP per capita

• This is the growth the US would typically get in 15 years

• This is the growth that China had in 5 years (over 1992–2019)

. . .

Some explanations:

1. Diffusion takes centuries, not decades.

2. Forecasters expect that there will still be thousands of critical tasks which AIs cannot do, even in 2050, but then it seems there’s an ambiguity in the scenario.

3. Powerful AI causes counterbalancing effects pushing GDP down – e.g. war, or disengagement from the market (but then would expect confidence intervals to explode further).

Basil Halperin, who coauthored the paper, also finds it puzzling that respondents don’t expect rapid AI progress to lead to higher growth. He tries to steelman them, but finds both of his attempts unconvincing:

1. Yes bottlenecks will slow progress, this is extremely important. But the rapid scenario under consideration has already chewed through most bottlenecks.

2. Political backlash + regulation could prevent rapid capabilities progress from accelerating GDP.
   But again, in the rapid scenario, it’s plausibly too late. Also, to my understanding, this wasn’t a common justification (underrated IMO).

Arvind Narayanan thinks that the backlash will be powerful:

Many people seem surprised by the prediction in this report that even rapid AI progress won’t lead to explosive GDP growth. Couldn’t AI be different from past patterns?

In my view the fundamental barriers aren’t technological or even regulatory but social. Explosive economic growth implies a pace of social change that I predict the public in most countries will vehemently reject.

Take self-driving cars. Assuming continued rapid AI progress, there will soon come a time when it would be great for growth if you could get human drivers off the roads. Freed from the need to accommodate the limitations of human drivers, roads could be dramatically redesigned and land use patterns transformed.

But my head hurts even trying to imagine the backlash. I think banning human driving will remain completely outside the Overton window no matter the daily death toll. The actual debate we’re having now is whether to allow self-driving cars at all.

AI boosters are right that regulation is not a fixed constraint, and that if AI is powerful enough, it will be a force for change. But they get the direction wrong! It is true that regulation is largely a reflection of social preferences, but those preferences tend to be even more strongly opposed to rapid change than is reflected in policy at any given time. As AI-driven social change kicks in, we will see regulation start to change – in the direction of erecting more barriers, not tearing them down.

If your mental model of AI impacts starts and ends with capability growth, you’ll miss the layer that actually matters. Far from AI being a wild west, the diffusion layer is already highly constrained, and it’s quite possible that those constraints will tighten.

Economist Paul Novosad:

Economists and AI experts have very similar forecasts on what AI will be able to do in 20 years.

But the AI experts think it will have a much bigger effect on the economy than the economists do.

Why? Because economists study this stuff.

Self-driving cars have been considerably safer than human drivers for about 3–4 years. How many people have access to self-driving cars? Even 20 percent penetration is years and years away.

Adoption is slow, society doesn’t transform instantaneously, sometimes it doesn’t transform at all.

We still don’t have high-speed trains in the US and we are banning windmills. Technology is a constraint but it’s far from the only one.

Fellow economist Alex Imas responds:

My worry is that economists’ forecasts of AI’s impact on growth is colored too much by historical precedent. Historical precedent is important, but we should be humble about the possibility of AI being more transformative than prior technologies.

For example, the speed of transformation is at least partly determined by adoption and diffusion through the economy. If one’s model is based on AI being adopted within existing organizational structures, then diffusion will be quite slow – this is what we’re seeing now.

But consider the possibility of a ‘Coasean Singularity’ – a scenario where AI drives the transaction and coordination costs that traditionally dictate firm size to near zero. This could lead to the (potentially fast) emergence of smaller, more nimble AI-first firms, new types of organizations that are outside of our current models, that don’t have historical precedent. These firms will not have the sort of bottlenecks of traditional firm structure, and the transformation and resulting impact on economic growth would be much closer to the technological frontier.

I know that many economists are already thinking through these transformative scenarios. My guess is that as these ideas are developed further, forecasts will change as well.

I think this debate is excellent. What should we make of it? Why don’t people expect higher growth under the rapid AI progress scenario?

I agree with Basil that belief in a political backlash is unlikely to be a major driver. People often fail to fully account for this factor. Instead, I think they simply believe that integrating AI into existing workflows will be painful, as it is today. But it’s not obvious how to square this with the powerful capabilities the rapid progress scenario assumes. As Basil points out, many of the current bottlenecks should no longer apply given this scenario. You don’t need to worry as much about workflow integration when AI can autonomously do complex jobs end-to-end. Here’s part of the paper’s description of the rapid progress scenario:

A single AI agent can generate a Pulitzer (or Booker Prize-) caliber novel according to current (2025) standards, adapt the book into an engaging two-hour movie, negotiate the resulting book and movie contracts, and launch the marketing campaigns for both while its sibling agents manage the book publishing company and movie studio at the level of highly competent CEOs.

It could be interesting to study a limiting case: even more rapid progress, where AI quickly reaches superintelligence. Is there any scenario where people would expect ten percent growth or more? If not, one might suspect they simply don’t take these kinds of scenarios fully on board.

Subscribe now

Europe replaces American aid to Ukraine

American aid to Ukraine has fallen to almost nothing, but as Europe has increased its support, the net decrease has been muted.

Source: Kiel Institute

Financial and humanitarian aid mostly comes from the EU budget, but military aid is provided by individual countries. Some of them give much more than others. The Nordic and Baltic countries account for just 8 percent of the GDP of the tracked European donors, but provided 33 percent of military support in 2025. Conversely, Southern Europe has 19 percent of GDP but its share of military support was only 3 percent.

Source: Kiel Institute

Philosophy has grown more empirical

Philosophy is often viewed as an armchair discipline that barely engages with empirical science. But in a new paper, Michael Prinzing shows that this view is increasingly outdated. In leading analytic philosophy journals, citations of empirical sources have risen substantially since the 1970s.

I think too many philosophers still ignore relevant empirical research, but this is a welcome change. Lagging curricula should be updated.

Poorer countries have very little mental healthcare

While richer countries have better healthcare across the board, the difference is especially striking in mental healthcare. High-income countries have roughly 70 times as many psychiatrists per million people as low-income countries. In terms of spending, the gap is even wider.

Climate change has different drivers over the short and long term

In public discourse, climate change is primarily associated with carbon dioxide (CO₂) emissions. But many other emissions also affect the climate, such as methane from agriculture and fossil fuel extraction.

Why then the focus on CO₂? Because most emissions break down within a few decades, whereas CO₂ persists. If you’re pessimistic about long-term climate prospects, it makes sense to prioritize CO₂ emissions.

Source: IPCC. Legend abbreviated.

Subscribe now

In brief

• Benjamin Stubbing and Oscar Sykes on why it proved so hard to make good instant coffee

• Iran increases repression to forestall protests

• Paper finds that noncompete agreements reduce American wage growth

• Bioarchaeology links violence to weak state institutions

• Study suggests that reducing top income taxes would increase tax revenue in Norway

• Ethan Mollick argues that the chatbot interface doesn’t realize AI’s potential

• Luis Garicano and Jesús Saa-Requejo on how Europe could take the lead in AI adoption

• Sam Harris interviews Will MacAskill on the ups and downs of effective altruism

• David Oks on the rise of scientific citation metrics and the problems they cause

• Robert Long argues that LLMs are neither mere parrots nor similar to humans

• American literacy isn’t as poor as recent discourse has suggested

• March newsletters from Coefficient Giving and the Institute for Progress

How smart can AI become?

There’s little reason to think that AI couldn’t become smarter than humans. But how much smarter? On Sunday, François Chollet initiated a debate on that question (quotations have been lightly edited for clarity):

One of the biggest misconceptions people have about intelligence is seeing it as some kind of unbounded scalar stat, like height. ‘Future AI will have 10,000 IQ’, that sort of thing. Intelligence is a conversion ratio, with an optimality bound. Increasing intelligence is not so much like ‘making the tower taller’, it’s more like ‘making the ball rounder’. At some point it’s already pretty damn spherical and any improvement is marginal.

Now of course smart humans aren’t quite at the optimal bound yet on an individual level, and machines will have many advantages besides intelligence – mostly the removal of biological bottlenecks: greater processing speed, unlimited working memory, unlimited memory with perfect recall… but these are mostly things humans can also access through externalized cognitive tools.

I do believe that a large collective of the smartest humans, aided by external tools, sits very close to the optimality bound – i.e. humans should be able to solve any solvable problem (where the required information is available) if they pay enough attention to it.

This received plenty of responses. alz:

This is basically my view: infinite free intelligence will not hyperaccelerate science (though it may well put all the humans out of a job!) because science in 2026 is not actually intelligence-bottlenecked. We have basically enough intelligence already.

However, I don’t think this follows from François’s tweets. Here’s Toby Ord’s response to him:

So it sounds like you are using a definition of intelligence where time is factored out (so you wouldn’t count a society that could discover calculus in 5 minutes as more intelligent than one that took 300,000 years).

. . .

And it would make the consequences of your claim weaker than people would assume. i.e. it is compatible with this that humans are 99 percent (or 100 percent) of perfect intelligence, yet future AIs could run rings around us, inventing the next thousand years of technology within months, etc.

So François’s claims are consistent with AI hyperaccelerating science. alz misinterpreted him in exactly the way Toby warned about, suggesting he wasn’t clear enough.

Ryan Greenblatt makes a point similar to Toby’s:

I probably agree with a claim like ‘human civilization with cognitive tools and unlimited time and attempts could eventually reproduce any specific cognitive feat AIs could do’, but the quantitative factor may be extremely large!

We are nowhere near the limits on how much cognitive firepower is possible.

Notice that Ryan replaces the word ‘intelligence’ with ‘cognitive firepower’, which carries fewer associations. Many debates get stuck over definitions. When that happens, it’s often a good idea to reformulate your claims in other terms. ‘Cognitive firepower’ is well-chosen, since it makes his claim hard to dispute. Clearly, we’re far from maximum cognitive firepower.

Ryan also gives another argument, referencing human individual differences:

AIs could have way higher data efficiency/learning speed/fluid intelligence than humans. Humans vary wildly in these attributes and there isn’t a clear cliff/sigmoid in the human population!

Máté Bedő makes the same argument:

I find it strange that smart people look at how smart the average human is and then consider how smart von Neumann was with macroscopically indistinguishable thinking parts (running on 20W) and they conclude there is minimal headway for improvement for terawatt minds.

However we define the word ‘intelligence’, sufficiently advanced AI would have a tremendous real-world impact. There’s an argument that such AI is still far away, but I don’t think there’s an argument that its impact would be moderate.

Subscribe now

The US leads Europe in AI adoption

A new paper from the St Louis Fed finds that a larger share of workers use AI in the US than in Europe.

Within Europe, people in richer countries use AI more:

There are differences across industries, but perhaps not the ones you’d have expected. Many have argued that AI would disproportionately affect white-collar industries, but this study doesn’t seem to bear that out.

Note also that there’s no clear evidence that AI adoption is affecting employment at the industry level.

Electric vehicle batteries have become 99 percent cheaper

In China, some of the best-selling electric vehicles now cost as little as $10,000. How have they become so cheap? The key reason is the falling cost of batteries: 99 percent in the last three decades.

How microbubbles can carry drugs to the right place in the body

Standard methods often fail to deliver drugs to their targets, especially the brain. In Works in Progress, Ambika Grover writes about a possible solution: microbubbles, tiny gas-filled spheres injected into the bloodstream. Doctors direct ultrasound at the target – such as a tumor or clot – bursting the bubbles to release the drugs where they’re needed. This spectacular technology could improve the treatment of everything from stroke to brain cancer.

Subscribe now

In brief

• Iran is now earning nearly twice as much from oil sales as it did before the war

• US troops in the Middle East relocate from bases to hotels and offices to avoid Iranian attacks

• Tech philanthropy isn’t nearly big enough to replace government science cuts in the US

• Speed reading shortcuts don’t work, writes Tom Stafford

• How the quest to amplify electromagnetic signals led to four major inventions

• YouGov retracts poll on surging church attendance among young Brits over fraudulent responses

• GiveWell uses AI to red team its research with mixed results

• Data centers aren’t raising the temperatures of the surrounding land

• Stuart Ritchie and Tom Chivers on studies about power posing that failed to replicate

• People believe we’re less cooperative than we used to be, but the opposite is true

• Paper claiming liberal reforms harmed the Global South receives plenty of criticism

• How Demis Hassabis tested Mark Zuckerberg and went with Google instead

How to disagree about AI risk

There have been some public disputes in the AI safety community recently:

I think it exemplifies a pattern we’ve seen many times before: people being overly inclined to pick fights with others working for the same cause. But these disputes aren’t just a strategic concern: they also raise an epistemological question. If other smart and knowledgeable people disagree with you, to what extent should you trust your own judgment, and to what extent should you defer to theirs? Under the name epistemic modesty, this has received plenty of attention in the effective altruism and rationalist communities.

In this debate, I come down on the side of modesty. We instinctively trust ourselves too much and others too little, and need to adjust for that. Besides making our beliefs more accurate, I think more modesty could lead to less conflict in the AI safety community. If you’re less sure that you are right and others are wrong, you’ll be more inclined to cooperate.

How to be modest without killing the debate

But there is a further wrinkle. On Wednesday, Ian Leslie tweeted about Anthropic CEO Dario Amodei:

This is a modesty argument: Ian is saying that economists have relevant expertise and that Dario should defer to them. I agree in part: Dario seems too confident, given how much disagreement there is on this question. And yet, I don’t think he should simply fall in line with the economists. Dario is probably one of the people with the most knowledge about AI progress, and we need such people to share their personal views.

How can we be modest without killing the debate? Greg Lewis suggests an elegant solution: ‘distinguishing “credence by my lights” versus “credence all things considered”’.

So one can say ‘Well, by my lights the credence of P is 0.8, but my actual credence is 0.6, once I account for the views of my epistemic peers.’

I think this is a good norm, and encourage people to follow it whenever they can. But I doubt it will fully resolve the problem. It’s not just that public discourse doesn’t reward this kind of careful distinction. It’s that the distinction itself is hard to draw. Our private views are tacitly infused with deference to others.

But while I wouldn’t require Dario to state his credence by his lights and his credence all things considered, he could acknowledge that relevant experts disagree with him and indicate if it moves him. I think that if people did this more, the debate would be both better informed and less heated. Most of us are more likely to engage with someone who signals that they take our views seriously.

Subscribe now

Self-driving cars could cause gridlock unless we tax them

How will self-driving cars change the way we use roads? In Works in Progress, Ben Southwood argues they could cause widespread congestion. When you can sleep or work in the car, people will use it far more. To prevent gridlock, Ben proposes road pricing for self-driving cars. And he suggests we do it now, as people will resist it far more once self-driving cars are common.

The US drops in NATO’s spending ranking

In 2014, the US spent more of its GDP on defense than any other NATO member. But as the threat from Russia has led European countries to increase their spending, the US has dropped to seventh place.

Airlines could cut the climate impact from contrails by two thirds

Two percent of humanity’s climate impact comes from contrails, the clouds that form when planes fly through cold and humid air. A recent trial found that rerouting planes to slightly different altitudes can cut this impact by nearly two thirds. But as adoption was voluntary, airline planners chose the alternative route for only 15 percent of flights – and 40 percent of them weren’t flown as planned. The technology is here: now we need airlines to use it properly.

Source: Hannah Ritchie

Rising protests in Cuba

Over the last five years, Cuba’s economic crisis has led almost a quarter of the population to emigrate. Now the US has tightened its blockade further, threatening tariffs on countries that export oil to Cuba. Blackouts and fuel shortages have sparked a wave of protests, including an attack on a Communist Party office. Polymarket gives a 64 percent chance that Cuba’s leader Miguel Díaz-Canel is out by the end of the year.

Subscribe now

In brief

• Judge pauses the blacklisting of Anthropic

• Fewer American newborn boys are being circumcised

• Britain expected to suffer the most from the Iran War out of all large economies

• Drones intrude on Louisiana air base, forcing it to shut down

• Iran says it will charge fees for shipping in the Strait of Hormuz after the war

• How a TB vaccine became a treatment for bladder cancer

• A remembrance of the evolutionary theorist Robert Trivers

• AI models are much worse when queried in languages other than English

• How to train science funders

• Scotland’s financial system drove its success during the Industrial Revolution

• David Oks argues that the spreadsheet made American companies more shortsighted

• Asimov Press is pausing

Right-wing parties avoid Trump effect in Danish election

Many outside commentators expected Donald Trump’s demands for Greenland to help the Danish left, but this wasn’t borne out in yesterday’s election. The right-wing bloc actually gained ground against the left compared with 2022.

It was specifically the right-wing parties most opposed to immigration that grew, contradicting the narrative that the Social Democrats’ stricter stance has undercut them.

As neither bloc won a majority, most experts expect a new broad coalition, but with different parties from the previous one. Social Democrat Mette Frederiksen is predicted to remain prime minister.

Subscribe now

First-time homebuying in the US keeps getting delayed

In 1975, first-time American homebuyers had a median age of only 28, but today, they are almost eight years older.

The growing ideological divide between young men and women

The Financial Times has previously reported that the political views of young men and women are drifting apart in several countries:

Now they report on the same thing happening in Spain:

One driver may be that questions young men and women often disagree on – such as gender and immigration – have gotten more attention.

Cutting the knot of AI understanding

Do AIs possess real understanding, or are they actually little more than stochastic parrots? Matthew Adelstein argues it matters less than it might seem.

We know that AI can code, discover novel math proofs, provide original analysis of philosophical arguments, and more. We know that AIs can complete programming tasks that take humans a number of hours, and this has been doubling consistently, only to speed up recently. If you ask an AI to analyze a political issue, you’ll get a much more sensible answer than you’d get from 99 percent of the population – and not just in terms of knowing more facts but also in terms of reasoning better. We similarly know that AI is being designed to complete increasingly long tasks and function as an artificial agent.

The case for AI being a big deal, and potentially a serious threat to the world, simply depends on extrapolating out these trends. If AI is getting smarter, more ingenious, and more agentic by the day, then in the limit, we should expect AIs that are very smart and agent-like. That poses a number of quite severe risks. It also has the potential to radically upend the global economy by replacing most human workers.

You don’t have to sort out any of the interesting philosophical questions to evaluate this case. Whether the AI really understands what it’s doing . . . just isn’t relevant to this case.

I think this is mostly right. Too much of the current AI debate consists of people with different intuitions about the meanings of words refusing to see each other’s point of view. Sometimes the best approach to philosophical knots is not to undo them but to cut them.

That said, some underlying disagreements are substantive, and not merely verbal. Extrapolating current AI trends is less straightforward than this passage might suggest. As I’ve reported previously, smart and well-informed people disagree on how to extrapolate from METR’s time horizon studies of AI capabilities. And a major reason they disagree is that they hold different views on how AI systems work.

Bluesky is gradually declining

After Elon Musk acquired Twitter (now X) in 2022, some users left for Bluesky, a platform with similar functionality. The number of posters surged in late 2024 in response to Musk’s campaigning for Donald Trump. But since then, posting has gradually declined. Though activity has fallen at X, too, it is estimated to be around 10–30 times larger than Bluesky, depending on the metric.

Running AI models now costs more than training them

During the first half of this decade, most money invested in AI went into training better models. But now, the bigger expense is inference – actually running the models. This could widen the gap between free and premium models, as it becomes too costly for AI companies to give access to good models for free.

Coal is all but over in Britain

Coal has a special place in British history, as it helped launch the Industrial Revolution. It was the dominant source of electricity as late as the 1980s, but since then, the decline has been steep. In 2025, only 0.1 percent of British electricity came from coal.

Subscribe now

In brief

• Many carmakers are delaying their shift to electric vehicles

• Venezuela opposition leader María Corina Machado says US pressure is leading to reforms

• Jobs that are more exposed to AI aren’t necessarily more likely to be displaced

  • Another paper finds that a key factor is how tightly job tasks are bundled together

  • A third one claims moderate AI exposure benefits workers but high AI exposure harms them

• Automation may have increasing returns

• Ben Todd publishes a book on how to do good with your career, with a special focus on AI

• Rob Wiblin interviews Samuel Charap on the new dangers that may follow a ceasefire between Russia and Ukraine

• Fermented foods may have contributed to the evolution of bigger brains

• Kenyan field experiment finds dramatic effects of cheap GMO biofertilizers

• Social science papers lean left, and increasingly so since 1990

• Stuart Ritchie and Tom Chivers on the different theories of Tourette’s syndrome

Trump approval rating hits a second-term low

Silver Bulletin reports that Donald Trump’s net approval rating has fallen to -15.9, a second-term record low. A key driver is inflation, which is expected to rise as a result of the Iran War and could drive Trump’s ratings lower still.

Subscribe now

Why the standard redistribution playbook may not work with AI

AI may eventually lead to widespread job automation and wealth concentration. How should governments respond? Matt Bruenig writes in The Argument:

There is a well-established playbook for achieving economic security and equality that can be applied to an AI-infused economy just like any other advanced economy. Countries can levy taxes to fund universal welfare states in order to hold down income inequality and keep people’s material lives stable over time. They also can increase social ownership and impose targeted corporate taxes to put a lid on wealth inequality and excessive accumulation in specific industries.

The reason why the solutions to economic inequality and instability are always the same is because the problems are fundamentally the same. Capitalist economies distribute income by paying people for their labor and paying returns to people who own capital (wealth). Due to certain feedback loops, capital factors – or wealth – end up overwhelmingly concentrated in the hands of a small percentage of people.

I agree that there are similarities to previous economic paradigms, but a crucial difference is that a single country may come to dominate the AI economy: the United States. Since redistribution has so far mostly taken place within countries, that could make Matt’s well-established playbook insufficient. Deena Mousa (lightly edited for clarity):

One (big) problem: these solutions assume the country redistributing will have captured some gains from AI advancement. This is true for the US, but not for most of the world.

. . .

This isn’t an issue with Matt’s piece specifically; it’s intentionally US-centric. But when we talk about labor market impacts, we usually mean ‘for US white collar workers’, and I am much more concerned about people in countries that won’t capture as much value from AI.

Deena’s focus is on low- and middle-income countries, but this could also be a problem for rich countries that lack competitive AI companies.

Canadian brain drain

In the last decade, Canada has seen poor per capita growth, both compared with the US and the OECD as a whole.

Part of Canada’s problem is that a large fraction of the most skilled workers emigrate to the US. The top one percent of earners in Canada make only forty percent of what the top one percent make in the US. Consequently, at least forty percent of this group of Canadians emigrate, according to an estimate from a Bank of Canada working paper. This share almost doubled between 2005 and 2019, likely reducing Canadian growth. Net migration has also risen among Canadians in the top 2–10 percentiles, whereas it has remained low for those who earn less.

The EU introduces duties on cheap Chinese goods

In the last few years, a huge influx of cheap Chinese goods to the EU has overwhelmed customs authorities. At the Liège airport, officers can check less than one percent of packages – and of those that they do check, almost a third are unsafe, counterfeit, or misdeclared. The EU now plans to introduce duties on goods worth less than €150, which have so far been exempt.

Source: European Commission. LVC: packages worth less than €150.

Chinese girls are now taller than American girls

Thanks to better nutrition, average height has increased significantly in China. By 2019, Chinese 19-year-old girls were taller than their American peers, and boys were catching up as well. But they remain much shorter than the Dutch, one of the tallest peoples in the world.

What does it actually mean to be unemployed?

Economic statistics are often harder to interpret than it might seem. A good example is youth unemployment. According to Eurostat, Sweden has the second-highest unemployment rate among people under 30 in the EU.

But if we instead look at young people neither in employment nor in education or training, Sweden ranks second-best. One reason is that Sweden has many students seeking part-time work. Perhaps surprisingly, they count as unemployed – but since they’re obviously in education, they don’t show up in the second category.

Adapted from Eurostat

Subscribe now

In brief

• Electric vehicle batteries that charge in less than ten minutes

• We don’t have AGI yet by the most established definitions

• Liberal Robert Golob is expected to remain prime minister of Slovenia after Sunday’s election

• Why AI struggles with revolutionary science

• Moderation is only a winning election strategy if applied to the right issues

• Why GLP-1 weight-loss drugs need a far higher dose as an oral pill than as an injection

• Peter Singer and Kasia de Lazari-Radek interview Will MacAskill about effective altruism and the possibility of an intelligence explosion

• Prime Minister Giorgia Meloni loses referendum to overhaul Italy’s justice system

• Sentinel forecasters believe:

  • There’s a 37 percent chance that at least 1,000 US troops will enter Iran before July

  • The chance of two consecutive quarters of negative GDP growth (often used as a definition of recession) in 2026 is 29 percent in the US and 39 percent in the EU

• Many countries generate more power from coal to compensate for lost oil and gas supply

• Stock markets rise as Trump suggests a diplomatic solution to the Iran War may be near

Corn farming alone uses over 100 times more water than AI

Many people worry that AI uses too much water, but this is a hugely exaggerated concern. Andy Masley has put together an interactive graphic of American water consumption that lets you compare AI with other industries, such as agriculture. The difference is striking:

Subscribe now

Canada’s population is shrinking for the first time on record

A few years ago, Canada’s population grew at more than three percent per year. But since then, stricter rules have led to a collapse in temporary migration. As a result, Canada’s population declined in 2025 for the first time since records began in the 1940s.

The Anglosphere is falling in life satisfaction rankings

The US, Britain, Canada, and Australia are falling in the rankings in the World Happiness Report (which despite its name asks participants to evaluate their lives, not rate how happy they are).

American elites have become much more Democratic

Republicans used to have higher incomes and more education than Democrats, but over the last few decades, this has changed. A new paper studies this shift among white Americans and finds that it’s strongest in elite groups, who have quickly become more Democratic.

Anthropic could overtake OpenAI

Recently, Epoch AI projected that Anthropic could be on track to pull ahead of OpenAI in revenue. According to the latest estimates, OpenAI’s annualized revenue is $25 billion and Anthropic’s $19 billion.

While OpenAI still dominates the consumer market, higher-paying enterprises have increasingly turned to Anthropic. The Wall Street Journal reports that OpenAI plans to respond by focusing on its coding and business tools.

Americans find their compatriots immoral

Americans have become very negative about their society in recent years. This negativity extends to moral judgments about other Americans. In a recent Pew survey, American respondents were more likely to find their compatriots immoral than respondents from other countries.

The Argument has put together an interesting comparison between these moral judgments and measures of interpersonal trust. In all countries, participants are less likely to say that people are trustworthy than that they are moral. One possibility is that the bar for trusting others is higher, since even a relatively low chance of being cheated may be viewed as unacceptable. It’s also possible that people are more generous when judging their compatriots’ character in the abstract than when deciding whether to actually rely on them. The chart suggests this may be especially true in some countries.

Southern Europeans live longer than richer northerners

Eurostat has released life expectancy data from 2024. While richer regions generally have higher life expectancy, there’s an exception: people live longer in poorer southern Europe than in the richer north. Meanwhile, the formerly communist countries are lagging, with Bulgaria being last.

The US economy will suffer from the Iran War despite energy independence

The US was long dependent on foreign oil, but that has changed in recent years, largely thanks to fracking. This attenuates the economic impact of the Iran War, but Goldman Sachs still estimates that it will lower American GDP growth by 0.3 percentage points this year.

Americans think AI companies should pay for the jobs they eliminate

American voters find AI much more important than they did last year, according to a survey from Blue Rose Research.

How should people who lose their jobs to AI be supported? The survey found that Americans want the government to create good-paying jobs, not just hand out money.

They also think that AI companies should be held financially responsible for the jobs they eliminate.

Across these questions, the differences between Democrats and Republicans are relatively small. As I’ve noted previously, AI is not yet a particularly polarizing topic among American voters.

Surging sales of electric trucks

The share of new trucks that are battery-powered is rising steeply in China.

In Europe, the shift is more modest, but there are exceptions.

Subscribe now

In brief

• Alex Chalmers writes for Works in Progress on how Britain turned its fragmented electricity systems into a national grid

• Denmark and Slovenia both vote on Sunday

  • Social Democrat Mette Frederiksen is forecast to remain prime minister of Denmark

  • Right-wing hardliner Janez Janša is the slight favorite to become prime minister of Slovenia

• The Icelandic government proposes a referendum on resuming EU membership negotiations, with the aim of joining in 2028

  • A narrow majority of Icelanders want to join the EU

• Austin rents are falling fast thanks to a housebuilding boom

• Patients quit weight-loss drug trials after figuring out they got a placebo

• Regulatory reforms that could help old, sick, and poor Americans

• Freelance workers are worse off relative to employees than they think

• The Economist explains Iran’s drones and how they can be fought

• Iran destroys 17 percent of Qatar’s liquefied natural gas capacity (Qatar has roughly a fifth of the global market)

• The US may remove sanctions on Iranian oil tankers currently at sea to slow the rise in oil prices

Society won’t ignore AI risk

In recent years, many people have warned that advanced AI could cause large-scale harm or even lead to human extinction. But how will society respond? That factor is crucial, yet often overlooked.

Pessimists believe that the response will be too little and too late. This rests on two claims: that we won’t get clear signs of the risk in advance, and that even if we do, the response will remain feeble.

In a new post, Coefficient Giving’s Nick Gabrieli takes issue with the first of these claims. There will soon be what he calls ‘weak agents’: AI systems that can ‘complete tasks that would take humans days or weeks, but not months’. He argues that if there is indeed a risk that we will lose control of advanced AI, then weak agents will give us early signs. The mere experience of using agents that independently execute tasks can make the risk feel real. Moreover, there will be concrete evidence of misalignment (first two links mine):

Weak agents will probably act misaligned reasonably often . . . [They] will probably:

• Be hijackable by prompt injection attacks

• Be subtly sycophantic in their text responses

• Tell ‘white lies’ to users, e.g. failing to mention mistakes they have made while completing a task

• Actively deceive users, e.g. comment out errors in code or spoofing results

• Engage in weird/concerning if not actively harmful behavior in some contexts, e.g. scheming against humans on moltbook

I agree that weak agents will likely display concerning behavior. Likewise, human misuse of AI is set to grow more salient as the systems improve.

And if we do get these warning signs, I expect society to react forcefully. A politician as influential as Bernie Sanders has already called for urgent action on AI risk. In my view, the AI safety community has long underestimated society’s likely response. People don’t tend to sleepwalk into disaster.

Subscribe now

Flash mob shoplifting is falling in the US

A few years ago, videos of shoplifting teenagers swarming American stores went viral on social media.

But as Jeff Asher reports, these incidents are now becoming less common.

AI will make smartphones and PCs more expensive

On the Dwarkesh Podcast, SemiAnalysis founder Dylan Patel argued that the price of smartphones and PCs will increase drastically as semiconductor production shifts toward data centers for training and running AI models. He estimates that this will add $250 to the price of the next iPhone. The impact on demand could be dramatic, especially toward the lower end of the smartphone market:

It used to be that 1.4 billion smartphones were sold a year. Now we’re at about 1.1 billion. Our projections are that we might drop to 800 million this year, and down to 500 or 600 million next year.

Dylan argues that the impact on PCs will be similar:

This probably leads to people hating AI even more. Today, you already see all the memes on PC subreddits and gaming PC Twitter. It’s cat dancing videos saying, ‘This is why memory prices have doubled and you can’t get a new gaming GPU or desktop.’ It’s going to be even worse when memory prices double again.

Though note that other forecasters suggest the effects on prices and demand will be somewhat less severe.

Mosquitoes and snakes are the world’s deadliest animals

There are many animals that can kill humans, but some do so much more frequently than others. A new Our World in Data chart shows that mosquitoes are by far the deadliest animal, mostly through spreading malaria. Number two is snakes, which kill around 100,000 people per year (see Mathias Kirk Bonde in Works in Progress on how to reduce that number).

The British public says a near-term world war is likely

More than half of Britons now think another world war within the next 5–10 years is fairly or very likely.

But forecasters on Metaculus estimate that there is only a 25 percent chance of a world war before 2050. The public is often overly pessimistic in surveys of this kind.

Subscribe now

In brief

• Study finds that AI lowers skill requirements and reduces wage inequalities

• High school teachers who grade students too leniently reduce their test scores and future earnings

• The case that the world is still unipolar

  • Previous note of mine on the same theme

• The government research institute that nurtured Taiwan’s chip industry

• How to fund biotechnology projects that academics and venture capitalists don’t pursue

• Many hiring managers admit to ‘AI-washing’ of job cuts

• Stuart Ritchie and Tom Chivers on the mystery of dark oxygen at the bottom of the ocean

• Africa is predicted to remain energy poor through 2050

• Land readjustment: pooling plots to let landowners share the gains from housebuilding

• Student loan defaults are rising in the US

• Agricultural bioterrorism may be a more likely threat than bioweapons targeting humans

Democrats now favorites to win both the House and the Senate

Forecasters have long expected a split result in the November midterm elections, with Democrats flipping the House and Republicans holding the Senate. But in the last month, Democrats’ chances of taking the Senate have risen steeply, on the back of falling approval ratings for Donald Trump.

Subscribe now

Only half of Ukrainian refugees wish to return

Since Russia’s invasion, Ukraine’s population is estimated to have declined from 41 to roughly 31 million, as people have fled or died and fertility has fallen. How much could it recover if the war came to an end? Modestly, if a recent survey of refugees (excluding those in Russia and Belarus) is correct. Only 43 percent say they wish to return to Ukraine after the war, whereas 36 percent say they would prefer not to.

Lower-ranked American colleges have lost nearly half their students

Since 2010, the number of students at American universities has fallen.

As the American Enterprise Institute’s Preston Cooper shows, the decline has been uneven. While enrollment has held up at the top 40 percent of institutions – as measured by completion rates and economic outcomes – it has plummeted at middle- and low-ranking institutions.

Source: Preston Cooper, American Enterprise Institute

American companies are diverging, but consumer spending and incomes aren’t

In recent years, many have described the American economy as K-shaped, with one group pulling ahead and another falling behind, like the two diverging arms of the letter K. But is this true? When it comes to companies, yes. As Stripe’s chief economist Ernie Tedeschi shows, the most profitable firms have increased their share of the stock market.

Household wealth has also become more concentrated at the top, likely driven by the strong stock market.

But spending growth has been relatively equal across the income distribution.

And incomes have actually converged: the lowest third has done the best, and the highest the worst. As Ernie argues, lower earners have likely benefited disproportionately from the tight labor market.

Shipping in the Strait of Hormuz has collapsed

Despite Donald Trump’s call for ship captains to ‘show some guts’, few of them dare to risk an Iranian attack in the Strait of Hormuz.

Since a large share of vital commodities such as oil and fertilizer is transported through the Strait of Hormuz, the world economy could be disrupted if it remained closed.

In particular, many low- and lower-middle-income countries stand to suffer.

Markets expect a short war in Iran

But in fact, oil futures suggest markets don’t think the war will last very long. While a barrel of Brent crude costs more than $100 if delivered in May, contracts for delivery later this year are much cheaper.

Adapted from oilprice.com.

In line with this, inflation expectations haven’t followed the oil price the way they usually do.

Richer countries generate much less plastic pollution

The degrowth movement argues that economic growth is bad for the environment and that we therefore need to shrink our economies. But this isn’t right. Richer countries are in many ways better at protecting the environment. For instance, they generate far less plastic pollution even though they use more plastics.

China is aging fast

In 1965, China had a very young population with a median age of just 18. But since then, it has aged rapidly due to falling birth rates and increasing life expectancy. China’s median age has caught up with the UK’s and is projected to be ten years higher in 2050.

The EU employment rate keeps climbing

In the EU, the employment rate of people aged 20–64 hit a new record of 76.3 percent in the last quarter of 2025, thanks to increases in southern Europe. Likewise, unemployment and underemployment are at record lows.

Subscribe now

In brief

• Cuba struck by national blackout

  • Protesters attack Communist Party building over food and energy crisis

• Andy Masley has put together a calculator for your personal carbon emissions

• Astera Institute commits $500 million to research infrastructure for the life sciences

  • It also launches an essay competition on systemic bottlenecks in science

  • Via Coefficient Giving’s Abundance and Growth Team’s latest reading list

• American growth revised to 0.7 percent in Q4, sharply down from Q3

• How can we make AI-produced science legible to humans?

• Ethan Mollick: ‘Comments on my posts are no longer worth reading due to AI bots. That was not the case a few months ago.’

• AI lab founders hold smaller shares of their companies than the Google and Facebook founders did

• Sentinel forecasters think major US attacks on Iran will cease on 19th April

• Dan Williams argues that wishful thinking is a myth

Background

From 2024, Anthropic had a contract with the Pentagon that included restrictions on mass surveillance and the use of fully autonomous weapons. The Trump administration demanded that those restrictions be removed, and when Anthropic refused, the Pentagon designated Anthropic a supply chain risk. Initially, some thought this could force Amazon and Google to cut ties with Anthropic and even cause the company to collapse. The Pentagon’s decision was widely criticized, including by the lead writer of the White House’s AI Action Plan, Dean W. Ball:

The US government just essentially announced its intention to impose Iran-level sanctions, or China-level entity listing, on an American company. This is by a profoundly wide margin the most damaging policy move I have ever seen the US government try to take (it probably will not succeed).

Democratic control

But Anthropic also has its critics. Stratechery founder Ben Thompson argues that Anthropic assumes a power it’s not supposed to have:

Who decides when and in what way American military capabilities are used? That is the responsibility of the Department of War, which ultimately answers to the President, who is elected. However, Anthropic’s position is that an unaccountable Amodei can unilaterally restrict what its models are used for.

Anduril CEO Palmer Luckey takes a similar view:

Do you believe in democracy? Should our military be regulated by our elected leaders, or corporate executives?

Palantir CEO Alex Karp is characteristically blunt:

If Silicon Valley believes we’re going to take everyone’s white collar jobs AND screw the military . . . If you don’t think that’s going to lead to the nationalization of our technology – you’re retarded.

This debate has to an extent inverted standard patterns: left-wingers side with a private company, and right-wingers talk about democratic control. Noah Smith, quoting Marc Andreessen:

Some might say this shows that people are being hypocritical, but I don’t think it does. Consider Rutger Bregman, who has defended Anthropic forcefully:

This is the most important thing happening in the world right now.

The administration wants killer drones + mass surveillance of Americans.

Anthropic refuses to build it. While most tech companies fall in line, they are prepared to pay the price for their principles.

This is entirely consistent with thinking that elected governments should generally have more power than private companies. People don’t make judgments based on their general stance alone, but also consider the specifics of the question at hand. That’s not hypocritical but common sense.

Monopoly on violence

Anthropic CEO Dario Amodei has compared selling advanced chips to China with selling nuclear weapons to North Korea. Thompson now turns this analogy against him:

Why would the US government want to kneecap one of its AI champions?

In fact, Amodei already answered the question: if nuclear weapons were developed by a private company, and that private company sought to dictate terms to the US military, the US would absolutely be incentivized to destroy that company.

. . .

There are some categories of capabilities – like nuclear weapons – that are sufficiently powerful to fundamentally affect the US’s freedom of action; we can bomb Iran, but we can’t North Korea.

To the extent that AI is on the level of nuclear weapons – or beyond – is the extent that Amodei and Anthropic are building a power base that potentially rivals the US military.

. . .

It simply isn’t tolerable for the US to allow for the development of an independent power structure – which is exactly what AI has the potential to undergird – that is expressly seeking to assert independence from US control.

Noah Smith writes in support of Thompson’s argument.

This isn’t a question of law or norms or private property. It’s a question of the nation-state’s monopoly on the use of force.

To exist and carry out its basic functions, a nation-state must have a monopoly on the use of force. If a private militia can defeat the nation-state militarily, the nation-state is no longer physically able to make laws, provide for the common defense, ensure public safety, or execute the will of the people.

I think these arguments go too far. Whatever you think of Anthropic’s refusal to the Pentagon’s request, it’s not remotely akin to building a private militia. Setting conditions on how the military uses your products is categorically different from building military power of your own.

Subscribe now

Regulation skepticism

Anthropic has also received criticism from the opposite direction. The company has long called for more regulation of AI to ensure that it is safely deployed. In some critics’ view, this is now coming back to bite them. Dwarkesh Patel:

The AI safety community has been naive about its advocacy of regulation in order to stem the risks of AI. And honestly, Anthropic specifically has been naive here in urging regulation, and, for example, in opposing moratoriums on state AI regulation. Which is quite ironic, because I think what they’re advocating for would give the government even more power to apply more of this kind of thuggish political pressure on AI companies.

Rob Wiblin responds:

Dwarkesh argues that AI safety regulation is risky because it hands wannabe authoritarians easier options for seizing control of AGI and permanently consolidating their power (among other reasons).

But he also appears to oppose state-level AI regulation, and indeed supports a moratorium on it.

If you’re worried about losing democracy then you should naturally be much more worried about federal-level regulation where the military and intelligence services are located and which no external actor can meaningfully check.

California or Texas or Florida creating their own safety office for frontier models actually distributes influence over AI more widely than it stands currently, and doesn’t easily lead to dictatorships that can’t be challenged or escaped.

So this actually seems to be an argument in favour of abandoning federal regulation and aiming to do things at the state level long-term.

I agree with that. Too much of the debate focuses on regulation in the abstract. This is a good example of how much the specifics matter.

Rob continues:

You might also say that avoiding federal regulation now wouldn’t stop a bad actor cynically imposing and then exploiting them in future. But that’s a general problem with all arguments that say we shouldn’t do good things now because they’ll be turned to bad ends later.

If you don’t do good things now, the bad things can and often will be done later anyway.

I broadly agree with this, too, though I think there are some nuances. Since bad actors typically don’t have unlimited power, what legislation they can refer to can make a difference. The Pentagon might have been able to act even more forcefully against Anthropic if other legislation had been in place. Dwarkesh warns that ‘a purpose-built regulatory apparatus on AI’ could have this effect, though he doesn’t specify which particular regulations are liable to misuse.

Like nukes or the Industrial Revolution?

Dwarkesh also engages with the comparison between AI and nuclear weapons.

Ben Thompson had a post last Monday where he made the point that people like Dario have compared the technology they’re developing to nuclear weapons – specifically in the context of the catastrophic risk it poses, and why we need to export control it from China. But then you oughta think about what that logic implies: ‘if nuclear weapons were developed by a private company, and that private company sought to dictate terms to the US military, the US would absolutely be incentivized to destroy that company.’ And honestly, safety-aligned people have actually made similar arguments. Leopold Aschenbrenner, who is a former guest and a good friend, wrote in his 2024 Situational Awareness memo, ‘I find it an insane proposition that the US government will let a random SF startup develop superintelligence. Imagine if we had developed atomic bombs by letting Uber just improvise.’

In Dwarkesh’s view, this analogy is misconceived:

AI is not some self-contained pure weapon. A nuclear bomb does one thing. AI is closer to the process of industrialization itself – a general-purpose transformation of the economy with thousands of applications across every sector. If you applied Thompson’s or Aschenbrenner’s logic to the industrial revolution – which was also, by any measure, world-historically important – it would imply the government had the right to requisition any factory, dictate terms to any manufacturer, and destroy any business that refused to comply. That’s not how free societies handled industrialization, and it shouldn’t be how they handle AI.

People will say, ‘Well, AI will develop unprecedentedly powerful weapons – superhuman hackers, superhuman bioweapons researchers, fully autonomous robot armies, etc – and we can’t have private companies developing that kind of tech.’ But the Industrial Revolution also enabled new weaponry that was far beyond the understanding and capacity of, say, 17th-century Europe – we got aerial bombardment, and chemical weapons, not to mention nukes themselves. The way we’ve accommodated these dangerous new consequences of modernity is not by giving the government absolute control over the whole industrial revolution (that is, over modern civilization itself), but rather by coming up with bans and regulations on those specific weaponizable use cases. And we should regulate AI in a similar way – that is, ban specific destructive end uses (which would also be unacceptable if performed by a human – for example, launching cyber attacks).

But while AI has some similarities with the Industrial Revolution, there are also important differences. The Industrial Revolution only led to aerial bombardment and chemical weapons after a massively distributed process spanning several centuries. That gave governments ample time to act. By contrast, many people think individual AI companies can develop extremely powerful capabilities in a matter of years. If so, government intervention may have to be more proactive.

Subscribe now

Tenuous connections

Many of the arguments in this debate are very broad, and the link to the dispute between Anthropic and the Pentagon is sometimes tenuous. The state’s monopoly on violence and safety regulation of AI are important questions, but this episode sheds less light on them than it might appear. I think the debate would gain if the arguments became more specific.