NEW: I’m not sure people fully appreciate how dire the US life expectancy / mortality situation has got. My column: enterprise-sharing.ft.com/re… And some utterly damning charts. 1) at *every* point on the income distribution, Americans live shorter lives than the English.

Throughout Europe, school results for children of migrants are worse than natives Only in UK are they actually better In fact, UK 2nd-gen migrants do better than ANY Euro cohort: native or immigrant If demography is destiny, UK's doing ok. My column:- times-comment.com/britishkid…

So, to recap: - Productivity estimates were woefully off - Immigration estimates too - Inflation/GDP stats not great - Private rent data was wrong - Labour Force Survey is broken - Population estimates, employment statistics, business population statistics all broken too (1/2)

In the @ft today, I gave @SoumayaKeynes a sneak preview of the work I will be presenting at the Brookings Institution (still under embargo) next week, on how cheap knowledge reverses the long term trend towards specialization: "The Vanishing Advantage of Specialization: AI, Knowledge Utilization, and the Boundary of the firm." ft.com/content/9edd3f8c-c28b…

If you want to know what has everyone at the AI labs spooked, it’s this - the AI rewrote its own instructions. (Read the last line.)

Wild. As recently as Dec 2024, the typical AI researcher thought AI wouldn't solve a Millennium math problem until 2054.

I feel like I'm taking crazy pills with how many smart people I read and respect are saying stuff like this. The idea that the AI labs are suddenly, only now, just in September of 2026, advocating for AI safety, and that they just came around to this position bc of sudden financial precarity, is just completely, utterly, conclusively, 100% wrong. Here's Dario Amodei telling Ross Douthat in February that he agrees with the case for slowing down; that he's in favor of "collaborating" internationally to organize a slowdown; that "I would be all for" a global slowdown. This was 7 months ago, when Anthropic's annualized recurring revenue was rising faster than any company in modern history. He's been saying stuff like this for years, when Anthropic was worth millions of dollars and when Anthropic was projected to IPO for trillions.

Thanks for all the feedback on the post. One reaction we've been getting is "so you're saying nothing much will change." It's worth clarifying this because that is not at all the point of the post—it's actually quite the opposite. What it's saying is that the economy and society can change profoundly, to become almost unrecognizable, while still resulting in measured GDP growth of "only" 4–5% (which already implies doubling living standards in 15 years). We can get cures for major diseases, agentic transactions can explode, and we could have massive agent-to-agent marketplaces, but this will not show up as double-digit GDP growth precisely because of how GDP is calculated. This is the hallmark of our Assumption 2: the automation of agriculture had absolutely profound societal implications. Before this, people routinely died of starvation and malnutrition—a concern that largely vanished in the developed world. The increase in living standards and the precipitous drop in child mortality had huge welfare consequences, but because automated agricultural goods became so much cheaper, the sector actually shrank as a share of GDP. The same applies to agentic interactions: GDP only counts final output sold to human consumers, not intermediate transactions. Unless agents are counted as humans (final consumers) in the economy, massive agent-to-agent marketplaces net out as intermediate inputs—and even where agents serve humans directly, falling prices shrink their weight in GDP. This is all to say: I expect AI's impact on society to be nearly unprecedented, particularly with respect to human welfare. But you should not be looking at GDP growth numbers for evidence, precisely because some of the biggest welfare improvements will not show up there.

New essay on @alexolegimas's blog: Will AI Soon Deliver Double-Digit Growth? Probably not. Here is why. aleximas.substack.com/p/will… 1. We outline the economics behind oft-discussed predictions that AI will soon deliver double-digit GDP growth in advanced economies. We list the assumptions that need to all hold in order for double-digit growth to happen and explain why we think they won’t. 2. To be clear: we are extremely bullish on AI and think the capabilities explosion predicted by technologists is already happening (e.g. yesterday's Navier-Stokes news!). But predictions of GDP growth in the 2030s of 15%, 30% or even 100% per year are off the mark. What we take issue with is the timeline. To paraphrase Milton Friedman's dictum on monetary policy, AI will affect GDP growth with "long and variable lags." 3. Start with some growth rate arithmetic. It is often much more useful to first think in levels rather than growth rates. Ask yourself: how much richer will we be in, say, 15 years? If you think twice as rich, that implies 4.7% annual growth, which would already be massive. Ten times as rich requires 16.6% per year; it would also imply that we are 100 times as rich 30 years from now! Asked in levels, we bet that most people would come up with much lower growth rates. 4. It's important to be clear: there is absolutely nothing in standard growth theory that constrains growth rates to be in the single digits. In fact, it's pretty easy to write down theoretical models that deliver explosive double-digit growth. We show this by writing down a standard textbook growth model of the type we routinely teach our undergrads (a souped-up Solow model), plug in some seemingly innocuous parameter values, and get AI-driven double-digit growth by the mid-2030s benjaminmoll.com/task_based_…. The basic logic is that, by replacing labor with capital, automation alleviates / eliminates diminishing returns and removes labor as a bottleneck on growth. Fancier models, in which AI also automates R&D, deliver even wilder numbers. 5. But just because something is possible in theory doesn't mean it will happen in practice. The explosion rests on five assumptions, and each is unlikely to hold within the next 10-15 years. These assumptions are: Assumption 1: Fast, economy-wide automation, with machines doing two thirds of all tasks by 2035. Historically, automation has proceeded at about 2% of tasks per year. Most work is physical, not cognitive. And politics will slow things down. Assumption 2: People keep spending on whatever gets automated. They don't. As things get cheap, their share of spending falls, as it did for agriculture and manufacturing. Messy jobs, relational goods and scarce physical inputs like energy, chips and land become the new bottlenecks. Assumption 3: Someone buys the new output and firms invest to produce it. Automation shifts income from workers to capital owners, who spend a smaller share so demand may not keep up with supply. Assumption 4: No AI-driven cyber incidents destroying economic value. AI can also destroy output, and every incident slows deployment and investment. Assumption 5: Explosive technology growth because AI automates R&D. The wildest scenarios in which the economy doubles each year all rest on this feedback loop. There is no evidence for it so far. Why do many people who are closest to the technology (and who have been consistently right about the capabilities explosion) consistently predict double-digit growth? Our best guess: they extrapolate from their own sector to the rest of the economy. This reminds me of the 2022 German gas debate: industry insiders were right about their own firms and very wrong about the economy as a whole. Our bottom line: a much more likely outcome is a large increase in the level of GDP spread over a decade or two, which is what 4-5% growth is. If you remain unconvinced and still believe in double-digit growth, we are still looking for counterparties for our bet benjaminmoll.com/growth_bet/ 😃

The Economist article from the other day aligns closely with the take in the quoted tweet: in aggregate AI has been net positive for workers so far and created a ton of new jobs relative to a no-AI counterfactual. Interestingly, it also highlights how AI exposure is translating into different impacts on labor demand. The highly exposed jobs where employment appears to be falling are more likely to have lower "adaptive capacity" to transition to good new jobs (see figure below from our paper on adaptive capacity from earlier in the year). These are largely back-office administrative, clerical and customer service type roles. By contrast, the highly exposed roles where employment is growing are more concentrated in technical and high-level professional occupations: Lawyers, Financial Analysts, Data Scientists, etc. These workers are already quite well off, on average, and may be benefiting most from AI, while back-office workers who have lower incomes, lower savings, and more narrow skillsets to sell in a labor market, may be initially losing out. Overall the BLS expects office and administrative-support jobs to fall by ~ 750,000 by 2035. For policymakers concerned with managing AI's disruptive effects on workers, it seems like two segments to focus on right now are incumbents in these low-adaptive-capacity roles and recent college grads entering a low-hire market for many white collar professions.

I've just heard my MP @helenhayes_ has spoken at an event in the House of Commons blaming England's attendance problems on its "narrow knowledge-based curriculum". If a knowledge-based curriculum causes attendance problems, how come Scotland & Wales have a worse attendance problem than England when they have a skills-based curriculum?

I hear roughly 3 reasons people keep working at AI labs despite believing in ~10% extinction risk: 1) Techno-determinism: Someone will build ASI no matter what, and I can do it better & more safely than China/OpenAI/etc 2) Consequentialism: ASI might kill us, but it also might produce utopia/immortality/superabundance, so it's a +EV bet 3) Self-interest: I am personally having fun & getting rich working on cool tech with friends. I don't think about the macro stuff. Notably, none of this is "I'm hyping up the risk for marketing reasons." People believe what they say, while being capable of a lot of internal dissonance / compartmentalization / self-justification. (Personally, I think the public is better served by AI researchers talking about & creating consensus for the specific safety solutions — regulatory or otherwise — they want vs. vagueposting about extinction.)

OpenAI employees who agree with this analysis (which seems very reasonable to me) need to understand that their company’s lobbyists and the Leading The Future superpac their executives are funding is doing everything possible to make this coordination not happen.

Still more absolute 🔥 from Terence Tao: “We have now seen that even the rumor of someone working on a problem can trigger a massive amount of AI-powered effort to flatten it before the original research project has time to reach its full potential. The incentives may now be pointing in the direction of no longer sharing any promising research directions with the broader community, which would reverse centuries of traditions of open science and do serious long-term damage to the future of the field”

Wrote up my thoughts on the whole OpenAI Navier–Stokes Millennium Prize Problem story, and how it highlights the still confusing question of what using my data "to improve model performance" actually means simonwillison.net/2026/Sep/8…