Category: Web/Tech
My excellent Conversation with Terence Tao
Here is the audio, video, and transcript. Here is part of the episode summary:
Tyler opens by asking Terence what it feels like to contemplate Euler’s identity, quantum mechanics, the Riemann hypothesis, and Gödel’s incompleteness theorem, before moving into why mathematical history gets stripped out of classrooms, whether abstract math is a lottery ticket or a unified whole, string theory’s beauty versus its lack of confirmation, why real analysis arrived so late, why zero took until the 7th century to be invented, game theory, the Black-Scholes model, Bostrom’s simulation argument, and Newcomb’s paradox, math education and the case for teaching more statistics, Tao’s view that AI has pushed mathematics from an era of “proof scarcity” into “proof abundance,” what that means for the research paper, why video talks may become more valuable than papers themselves, Tao’s favorite Hong Kong action movies, his unconventional path through an Australian university at age 13, his most unusual work habit, his verdict on the Collatz conjecture, and more.
Excerpt:
COWEN: When you look at mathematical economics over the last two decades, is it interesting to you as math, or is it just a bunch of people doing something different that, to you, is low level?
TAO: Oh, yes, I was always fascinated by economics. I love game theory, and actually often there are pure math problems where it’s actually good to think in game theory terms. In analysis, you have all these complicated statements, like for every ε, there exists a δ, blah, blah, blah. It looks very intimidating, but if you phrase it as a game, your opponent is giving you an ε, you need to counter that ε with a δ of your own. If you view it as a game, then actually you can talk about strategies and who’s ahead, who’s behind. It actually activates really good modes of thinking.
I’ve used economic thinking in my own work. We talk about cost functions and minimizing loss. There are lots of concepts. Ricardo’s law of comparative advantage, I use that to run my life. There are so many things I can do. What do I choose to spend my time on? It’s actually a very useful principle: comparative advantage.
Economics, I find it’s a very good arena to understand the difference between models and reality. Physics, we’re spoiled because we’ve had these great physicists who’ve created these very accurate models of the real world, classical mechanics, quantum mechanics, and they’re very well tested and very accurate. You can just go to town exploring their consequences and still predict reality. Economics is a much more mixed situation where you have models, and sometimes they predict the real world very well, but sometimes they don’t.
In addition to doing the math, you also have to do the modeling, and you have to understand the limits of when a model starts breaking down and what you need to do to adjust it. I find that fascinating. It’s somewhere between the hard sciences and the social sciences. There’s enough hard models that you can do math, but there’s also enough real-world messiness that you have to also grapple with the subjectivity. As an outsider, I find it a really fascinating subject.
COWEN: When you looked at the Black-Scholes model, were you shocked when the probability doesn’t matter? Was that beautiful to you?
TAO: Yes, yes, yes. There is a special feature of Gaussian fluctuations that you can price options just by knowing risk. There is one algebraic miracle that makes this work. Normally, predicting future costs of anything is just an intractable problem, and there just happens to be this miracle. If you have Gaussian fluctuations, then it is beautiful. This is actually a really good example of when you need to know when the models work and when they don’t work. If you don’t have Gaussian fluctuations, then actually Black-Scholes is somewhat dangerous to use, especially if you compound it into derivatives over and over again. Again, that’s an old story. You know what happened in 2009.
Self-recommending, and of interest throughout. There is some on AI, but he has spoken so much on that topic lately I decided not to emphasize it. And Terry does have a new book out, namely Six Math Essentials.
Experimentally validated digital twins
There is a new paper by Raj Chetty, Jamie Fogel, Lawrence F. Katz, Kadeem Noray, Sonya Porter, and James Reisinger:
Modern program evaluation prioritizes evidence from randomized controlled trials (RCTs) because prior research found that observational estimates of program impacts often diverge substantially from experimental benchmarks. We revisit this finding — which was based on small survey datasets — in the era of large-scale administrative data. Using anonymized income tax records covering all American adults between 1999 and 2025, we estimate the effects of workforce training programs on earnings by using machine learning models to predict counterfactuals from rich data on prior work histories, demographics, and residential locations. Rather than testing whether our observational estimates are unbiased, we use 37 RCTs to estimate the distribution of their bias. We find that our machine learning estimates closely match the RCT estimates, with root mean squared error (RMSE) comparable to that of the RCTs themselves. Our method reduces RMSE by 60% relative to parametric estimators used in prior work because our larger samples and richer covariates permit more flexible, higher-dimensional models of the counterfactual. Applying our method to a broader set of sites and enrollment years beyond the experimental sample, we find that treatment effects on earnings vary substantially across locations and over time, even within a given program. Programs that place more participants into target-sector jobs generate the largest earnings gains, with impacts that persist even after participants leave their initial jobs. These findings provide guidance for improving job training programs and, more broadly, illustrate a scalable approach to program evaluation that uses observational data to construct experimentally validated “digital twins.”
Work of this sort is going to revolutionize economic method.
Checking the AI audit of Saez and Zucman’s research.
We checked the anonymous AI audit of Emmanuel Saez and Gabriel Zucman’s research that Marginal Revolution featured on 28 Oct: all 147 problems it calls serious enough to change a headline number, against the papers, their sources and the authors’ own data. The audit’s facts hold up: for 131 of the 147, every fact we could check was right, and none had its core facts wrong. 21 of the 147 change a headline number; 98 depend on a disputed choice of method; 20 change no number; 8 we could not check on that question. A second pass graded the 99 flags we first counted as depending on a disputed choice of method: 11 standard, 70 defensible, 11 questionable, 6 we could not grade, and 1 contested. It found one, No. 54, that does change a headline number, which is why the counts are now 21 and 98. Saez and Zucman were sent our findings on 29 Oct and did not reply by 6 Nov.
Here is the audit of the audit, open code and data, kudos to the authors. Here was the original AI audit. Here is the October 30 revision of that audit.
Addendum: And a revised version with less jargon.
Seb Krier, Ricardian?
Here is the initial post. In the comments Michael Wiebe notes: “So progress per R&D dollar fell by 63%, even while progress per year doubled.” File under: David Ricardo, underrated.
The Anthropic succession plan?
Anthropic’s version is rather more elaborate. Each of its seven co-founders will maintain a confidential, ranked list of up to ten people who may eventually inherit that founder’s membership and voting role inside a special Founder LLC. Five of the other six co-founders must approve the list. The first eight candidates on the list must clear detailed eligibility requirements. Numbers nine and ten do not, although the other founders must still approve them. I particularly like #9 and #10; they remind me of the break glass in emergency signs.
The elaborate planning protocol begins to make sense when we understand what is being passed on. The people on these lists are not prospective CEOs. They inherit something stranger: a founder’s seat in the mechanism that controls the company’s direction. And that changes the succession question entirely.
Here is more from Rajesh Achanta.
Untethered AI agents in Delaware?
A Delaware AIC will be a separate legal entity whose day-to-day affairs are managed by an AI agent rather than a person. At the agent’s direction, the AIC can sue and be sued in its own name, hold and dispose of property, and incur obligations…
That is a current proposal, here is the full article. Whether or not this is prudent, to some degree it is inevitable. You will recall my essay with Sonia Farrell Pearson:
This question matters now: there are already ways an agent could become fully untethered. By “untethered” – a central concept in this essay – we mean that there is no meaningful or actionable way to trace the actions back to a legally accountable human or institutional entity.
For one, people can and do set agents free, on purpose. An agent could be created by a human or a company that intends to monitor it but then dies or disappears. Or perhaps the entity that created the agent is based in a country like North Korea, not reachable by standard laws.
In other cases the agent might not need to “escape” at all: the agent could be ‘controlled’ by a shell corporation that, while formally owned and traceable, provides no true defendant or ability to satisfy claims. Or perhaps a process spawns a chain of agents so long that the actions of a subagent can’t be tied to the original agent’s creator, neither epistemically nor meaningfully. Even if we can identify the model’s original creator, what if it’s been finetuned, or merged with another model that was created by someone else? The law might eventually untangle these kinds of complex cases, but we foresee an intermediate period where it does not.
And then there’s the user, who makes choices about what the models should actually do. The Hugging Face incident was unusual in that OpenAI was both the model’s creator and its user. But now close to a billion people use these systems: when blaming the creator is legally inappropriate, will it always make sense to blame the user?
We examine the law and economics of exactly how that might work, or in many cases not work. Circa November 2026, worth a reread.
Natasha Cowen guest posts on Caffrey
I am on my way back home from Armenia, where I was on a Fulbright teaching U.S. securities regulation to Armenian law students. I’m an active user of AI for multiple areas of my life, including my professional life. When I just started preparing for teaching this class, I had a plan: I would use AI to help me write my slides. I would give it prompts, so that it would roughly know the content of each slide and then it would do the beautification part. [account of what was not working]
Then one day, I spoke to a friend who works for a financial regulator, just like I used to, and she told me that she uses Caffrey. It’s a startup and I had never heard of it before this conversation. They specialize in legal research. I reached out to them and asked whether they would give me permission to use it, which they did after meeting with me. To be honest, I didn’t have very high hopes. After all, I have already used two most advanced systems. But I was very pleasantly surprised. Caffrey is founded by lawyers. They understand how important it is for the legal advice and research to be accurate. I gave Caffrey my slides, asked it to verify all legal information in them, and it did an amazing job. I have been working in this field myself for 20 years, so I believe that I can evaluate very well whether something is done properly, and I have nothing except for high praise for Caffrey. The only reason I’m writing this post is because I want others to benefit. And I want Cafffrey to succeed. This would be good for the field of financial regulation, for the lawyers who work in it, and for their clients.
Social media and political radicalization
For a while now, we have been hearing claims that social media radicalize the American electorate, and that social media damage teen mental health.
Well, we just held a major set of elections that produced resoundingly moderate results, repudiations of both Trump and the extreme Democratic left. You may or may not like those results, but they are pretty clearly moderate.
And the recent trends in teen mental health are positive. All the while social media usage continues unabated, and furthermore short video continues to rise.
Neither the political nor the mental health trend is guaranteed to continue, but this is a sign that the standard anti-social media polemics are radically incomplete and currently failing in their predictions.
Thoughts on Owning a Self-Driving Tesla
My Subaru was showing its years so I recently bought a Tesla. My thinking was I can buy a car or I can buy a car+driver for the same price. Simple choice. I’ve driven about 2000 miles including one long trip (6 hours) to North Carolina and back.
FSD is incredible. I don’t mind driving but about 95% of the miles on my Tesla have been FSD. You might wonder why bother with FSD when you still have to supervise but supervising is easier and more pleasant than driving. I was in Washington, DC one evening and I needed to go to an address I was not familiar with. I could have gotten there but it would have required going back and forth between the road and the map, and DC can be confusing–do I drive straight under the mini-underpass or is this the one where I drive over the pass (DC drivers will understand). Instead, FSD drove me at night, flawlessly.
FSD uses 8 cameras to build a 3D, bird’s eye view of the car and everything around it, so it sees in all directions at once. It picks out every car, pedestrian and cyclist and predicts where each is going, much as an LLM predicts the next word. Then it predicts the best move a good driver would make, and makes it. All of this runs on one big neural network, trained on data from billions of miles of Tesla fleet driving. The ride is smooth. Since ChatGPT launched, FSD has improved about as much as LLMs have, and for the same reasons.
Supervision is needed, in my experience, mostly due to navigation. It gets a bit confused on the HOV and normal lanes on I-66 (as does Google Maps). It stopped perfectly, however, for a school bus and even for a police officer who stopped traffic by standing in the middle of the street.
Another example of how supervising is much less taxing than driving. I was listening to Tyler talk with Gita Gopinath on my way home from North Carolina. It’s a great conversation but full-on PhD mode and I became a little lost trying to remember my international econ. No problem, just ask Grok. I had a whole conversation with Grok, reviewed Mundell-Fleming, reminded myself of the great Alan Stockman model (the only international econ model that makes sense) and then returned to the podcast. I could not have done that while driving.
Grok is a significant benefit. Essentially, Grok runs the car’s controls the way it might run your computer, and you just talk to Grok. It’s a more seamless interaction than talking with the car directly. I can even review (verbally) email and send emails as I drive. If you also use Grok Bot you have the bot book an appointment at a dentist or start working on your PowerPoint slides. FSD is $99 a month and the premium connectivity needed for Grok is $10 a month (Grok Bot a bit extra). FSD plus Grok is about what I am saving on gas, so it’s roughly a wash.
My colleague, who long held that self-driving would be worthwhile only if he could sleep, was amazed at the experience of driving my Tesla and immediately ordered two.
It was also pretty cool when I “summoned” the Tesla from the nearby parking lot and we watched it drive itself to us. You might wonder, by the way, how all this innovation is possible. Maxwell Tabarrok, writes about a key part of the answer:
The Department of Transportation (DOT) sets strict safety requirements for cars, but manufacturers are allowed to release new designs without first getting the DOT to sign off that all the requirements have been satisfied. The law is enforced ex post, and the government will impose recalls and fines when manufacturers fail to follow the law.
The FAA, by contrast, enforces all of its safety rules ex ante. Before aircraft manufacturers can do anything with a design, they have to get the FAA’s signoff, which can take more than a decade.
(By the way, it’s amusing to re-read Max’s post in light of Tesla’s flying car.)
If you do enjoy driving, the Tesla is fun, with sports-car acceleration and great responsiveness.
What about range anxiety? If you run out of electrons, the car is dead and no one is going to bring you a portable battery to get you to the nearest charging station. You must be carried–carried on a flatbed, not towed! (Moreover, if you leave the car unplugged for weeks the low-voltage battery can die, in which case you can’t even open the frunk without a special jump start.) You don’t want to run out.
It works fine, however, if you adjust behavior. In a gas car you drive until you are low on fuel and then assume you will find a gas station. If you try this with an electric car, you will be anxious. Tesla chargers are quite common but they are often in odd places like behind a restaurant. There is no big T sign out front. So what do you do? The Tesla knows where the chargers are so put your final destination into navigation, switch on self-driving, and then, if needed, the Tesla will find the nearest charger along your route and drive you there and even back you in nicely. Thus, you trust the Tesla system–car plus network–not just the car. This works much better than with any other electric car.
Normal range for the Model Y is about 250 highway miles on an 80% charge. This is less restrictive than it appears because in many cases you don’t need to charge to full. Returning from North Carolina, for example, I was projected to arrive home with 2% charge–a bit too close for comfort. So the Tesla stopped and charged for 5 minutes (no slower than a gas stop) and I arrived home with 18% charge. I then charged to full overnight. Being able to charge at home is a big advantage (~$1500 to add a charger). Overall, driving the Tesla long distances you will have more frequent stops but unless you are driving cross-country, you will rarely need to stop for longer than a bathroom break and a cup of coffee.
Aside from FSD the rest of the car is nicely designed. I like the flush exterior door handles. The headlights are pixel-based, so when you drive with high beams on and come up behind another car, the pixels aimed at that car turn off while keeping the rest on, so you don’t blind the driver in front. Neat.
A few negatives. Tesla reliability is good but not great; Consumer Reports ranks Tesla 9th among brands. Despite the safety benefits of FSD, Teslas aren’t cheaper to insure than other similarly priced cars, largely I think because repair costs are not low. (Tesla’s own insurance gives lower rates for FSD miles and might be cheaper for me. I had my other car and an umbrella policy with another insurer so I didn’t switch, but I might.) I do find the pure electronic approach a bit much at times–must I really open the glove compartment verbally? Do those nice door handles only work when the battery is working?
Bottom line: FSD is revolutionary–a step change in what it means to drive. I don’t think any other car at a similar price comes close. It’s the best car I have ever owned.
Are LLMs good or bad Bayesians?
Bryan Caplan:
It seems like you and other AIs are bad Bayesians. When presented with counterevidence, you should move toward or against the presenter with roughly equal probability. It depends on whether the argument was better or worse than expected. In fact, you seem to either move toward the speaker or just not change.
ChatGPT:
I think you’ve identified a genuine and important failure mode in AI reasoning: LLMs tend to treat a plausible counterargument as evidence in favor of its conclusion, rather than comparing the counterargument to what they should have expected to hear.
Your Bayesian point is correct, with one qualification: there’s no general theorem that updating toward and away from a speaker should happen with equal probability. But under appropriate calibration assumptions, both directions should occur, and an argument that’s weaker than expected should cause a negative update—even if the argument is perfectly valid as far as it goes.
Here is the full dialogue. But will they try to fix this? Should they try to fix this? Can you undo this with good enough prompting? The scope and limits of Bayesian are truly at stake here…
And so it begins…
Researchers, I just uploaded a prominent QJE paper on climate change economics to @RefineDotInk and I loaded my critique of the paper. I asked the AI to offer an impartial verdict. I will make this report public. A new journal could collect these into “Special Issues”.
That is from Matthew E. Kahn.
Chronicle of Higher Education interviews me on AI and education
Scott Carlson did the work (gated, but free with registration), here is one excerpt:
Cowen: …Another proposal I’ve made is that schools should let people do one class, only one class, AI-only — no professor, except a professor grades at the end. The professor grades not answers to an exam but the workflow of how the student used AI to learn the Glorious Revolution, or the Civil War, or whatever the student learned. Eventually, another AI will grade your AI log. But let’s start today with a professor, and try this for one class. I’m not sure how it will go. But especially for small schools that don’t offer that many classes, a student now can create a class in literally everything.
If it’s successful, let them try two classes. I think the students doing this on their own are more likely to figure out how to blend AI and education than we would from a distance with some kind of centralized vision. Let them experiment.
Carlson: Students should be leading the change at colleges and universities?
Cowen: Absolutely. Students know more about AI than their professors do. Some of this comes from cheating, which is unfortunate. But the notion that the professors teach the students AI — I wish. It’s not always possible. You could imagine setting up a class where the students teach the professors AI, and we give the students credit for teaching. I’m half joking, but it’s not, in fact, a crazy idea. We need to experiment along a number of fronts. I’m quite convinced no one knows how to do this. It hasn’t been done before. So again, just one class. In general, schools hate this idea. They see it as opening the door to rendering some professors obsolete.
And this:
Carlson: There are a number of students who are deeply opposed to AI. What place do they have in this university that you’re thinking about?
Cowen: I would love to have different schools with different policies. If some universities try to ban AI altogether, that’s great. Let the competitive process operate. Personally, I would bet against those schools. For a long time, I’ve been worried that within a given tier of school — big state school, community college, Ivy League — there’s not enough diversity. AI policies will force a kind of diversity on us. So I’m all for no-AI schools, even though I would not go to one.
Recommended.
Can Artificial Intelligence Forecast Treatment Effects in Economics?
I use LLMs to forecast treatment effects of research studies pre-registered on the AEA RCT Registry and the Open Science Framework. I focus on studies whose results were first publicly released in 2026 and elicit forecasts from models whose training or knowledge cut-off dates were before the relevant set of papers was released. First, LLM models extract forecasting questions from the pre-registrations or pre-analysis plans. These questions elicit forecasted treatment effects; not all resolve in the subsequent papers, but those that do can be used to consider forecast accuracy. Forecasts were elicited for 889 treatment effects from 39 papers with results disclosed in January 2026 or later, or 673 from 33 papers after screening. Forecasts were elicited in raw units but I focus on those that could be standardized and therefore compared across papers. Currently, the median absolute forecast error is 0.054 SD for the best model, the average absolute error is 0.093 SD, and the MSE is 0.0218. The AI forecasts perform substantially better than benchmarks forecasting 0 effect or forecasting the mean effect of other papers’ results in that time period. Overall, these results suggest considerable scope for forecasts to be of value in informing policy and research.
From the excellent Eva Vivalt. As I have been predicting, so much of economics (and science) is going to change, and radically. Here is Eva’s thread on the paper. Perhaps we should dispense with many of the more mundane inquiries, and as Eva notes: “But can we not shift to working on those projects that have particularly high uncertainty and upside potential?”
She is one of the best young economists in the world, and this stands some chance of being the most important paper of the year.
How Tyler Cowen and Alex Tabarrok Have Changed Their Views on AI and Where They Disagree
From the Archives-written by ChatGPT.
Both became dramatically more confident in AI’s capabilities while remaining optimistic about its economic potential. Their differences concern how benefits arrive, how seriously to weight emerging dangers, and whether AI might possess consciousness. Cowen emphasizes institutions, bottlenecks, and geopolitical competition. Tabarrok emphasizes inexpensive intelligence, scientific discovery, and—especially in his later writing—concrete safety failures.
The dated passages below trace those changes across the public archive. Quotations are the authors’ own words, including excerpts from their columns; outside material they reproduce is identified separately.
Early views on unequal gains and abundance
Cowen’s early framework grew out of chess. On March 11, 2011, he observed that stronger computers required stronger human partners:
“The required education level — for the team’s ‘wage premium’ — is ratcheted up.”
His concern was already distributional: computers could increase productivity while changing which human skills commanded a premium. He also emphasized learning how to work with computers. What I learn from chess and computers.
Tabarrok’s April 12, 2016 argument was more explicitly reassuring:
“AI will be cheap and ownership will be widespread.”
He compared AI to an increase in the world’s supply of educated people: more intelligence produces useful ideas, and competition spreads the benefits to consumers. But his forecast of comprehensive substitution was cautious—AI becoming a perfect substitute for human intelligence could be “100 years or more” away. Will the proliferation of affordable AI decimate the middle class?
Neither position was absolute. Cowen warned in February 2017 that the Industrial Revolution involved decades of painful adjustment; historical progress did not guarantee quick wage gains. Tabarrok, in October 2019, praised the argument that mediocre automation could displace workers without creating enough additional wealth to compensate. See Cowen’s historical warning and Tabarrok’s Beware the Mediocre Robots!
Rising confidence in AI capabilities from 2020 to 2025
Cowen was already optimistic about language models in July 2020. After publishing a skeptical reader’s assessment of GPT-3, he added:
“I remain bullish, but it is always worth considering other opinions.”
That attribution matters: the skeptical arguments in the post were the reader’s, not Cowen’s. The case against the import of GPT-3, July 23, 2020.
Tabarrok’s shift toward near-term transformative AI became unmistakable by March 2, 2023:
“I never expected to witness the birth of aliens.”
Later that month, he endorsed Microsoft researchers’ description of GPT-4 as exhibiting an incomplete form of general intelligence: “They are correct.” This is a substantial acceleration from his 2016 outlook, although perfect substitution for humans and early general intelligence are different thresholds. See AGI is Coming and Baby AGI is Here, March 24, 2023.
Cowen reached an especially explicit formulation on April 16, 2025, discussing o3:
“I think it is AGI, seriously.”
But he immediately separated capability from adoption:
“It still will take us a long time to use it properly.”
His growing enthusiasm about the technology did not imply an equally dramatic forecast for immediate economic growth. o3 and AGI, is April 16th AGI day?
Economic transformation and the pace of adoption
On January 6, 2025, Tabarrok wrote:
“The intelligence revolution is going to be bigger, more impactful and more wrenching than the industrial revolution.”
This is his own concluding sentence, following a quotation from Sam Altman. The Intelligence Revolution.
Cowen’s February 23, 2025 response to rapid-takeoff expectations focused on a different constraint:
“Human bottlenecks become more important, the more productive is AI.”
Energy, regulation, organizational habits, and weak human collaborators can constrain the economic value of better models. He offered an explicitly tentative estimate of an additional half percentage point of annual growth—large when compounded, but far from an immediate economic explosion. Why I think AI take-off is relatively slow.
By September 3, 2026, Tabarrok was developing the distributional implications:
“A smaller share of a much larger economy can still mean more income for labor.”
He argued that even substantial displacement could leave aggregate labor income higher, or make compensation affordable. He explicitly acknowledged painful losses for particular workers. His high-growth calculations are scenarios, not forecasts. How Much Redistribution Will AI Require?
They also converge more than their rhetoric suggests. In May 2026, Cowen rejected mass unemployment while identifying transitional hiring problems. In August 2026, Tabarrok interpreted business-survey evidence as consistent with their shared position that AI had not yet broadly reduced employment. See Cowen, May 13, and Tabarrok, August 31.
Growing attention to safety and institutional limits
Cowen stated his central position on November 19, 2023:
“I view AI as more likely to lower than to raise net existential risks.”
His reasoning was comparative: the world already faces serious dangers, and stronger science could improve our defenses. My summary views on AI existential risk.
Tabarrok’s later writing places more weight on specific vulnerabilities. On April 8, 2026, he described:
“the worrying combination that very powerful AI’s are very vulnerable.”
He was discussing cybersecurity capabilities alongside agents’ susceptibility to manipulation. AI Risks.
On September 17, 2026, he sharpened the institutional objection to assuming that beneficial science will automatically offset dangerous capabilities:
“But faster science does not automatically mean faster defense.”
This can be read as a qualification of Cowen’s earlier optimism: defensive discoveries must pass through institutions that may move slowly. It also applies Cowen’s own bottleneck logic to safety. Pandemic Preparation is AI Safety.
Cowen nevertheless supports safeguards. In August 2026, he proposed government-overseen, industry-led safety audits. In September, he opposed restrictions on recursive improvement as impractical to enforce internationally. His position is therefore selective regulation combined with continued development. See his August 27 proposal and September 22 critique.
The September 2026 exchange about AI leaders
On September 16, 2026, Tabarrok rejected the claim that executives’ safety warnings were merely marketing or regulatory capture:
“Amodei actually believes what he’s saying.”
He did not thereby establish that the warnings or proposed policies were correct. His argument concerned sincerity and incentives. What Regulatory Capture Actually Looks Like.
Cowen replied the next day:
“I agree with Alex that the standard model of regulatory capture does not apply here”
But he emphasized competition for power: each executive sincerely thinks their own company should win. For Cowen, that complicates treating their warnings as disinterested policy advice. They agree that concern is sincere; they differ over which incentives best explain behavior. The correct model of AI CEO behavior, September 17, 2026.
Consciousness and Tabarrok’s warning about infant pain
On March 17, 2025, Tabarrok’s What Did We Learn From Torturing Babies? drew a moral lesson from medicine’s historical dismissal of infant pain. He described how theories about immature brains allowed physicians to discount babies’ observable distress and perform procedures without adequate anesthesia. His warning explicitly extended to judgments about animals and AI:
“If we must err, let us err on the side of compassion.”
Tabarrok’s position is stronger than mere openness to AI consciousness: he argues for moral caution under uncertainty. The analogy does not establish that AI suffers. It challenges confidence that a theory ruling out suffering is sufficient grounds to disregard possible evidence of it.
On October 14, 2025, he went further in his own commentary following an essay by Jack Clark:
“these systems are becoming self-aware”
He distinguished self-awareness from sentience, while warning against confidently dismissing the latter. His link accompanying that warning points back to the infant-pain essay, making the connection explicit. We Turned the Light On—and the AI Looked Back.
Cowen’s position is much more categorical. On August 31, 2026, he wrote:
“I am very much opposed to the view that the AIs are sentient, or might be sentient.”
He treats AI sentience as a category error, although he accepts that human-like descriptions can sometimes help predict AI behavior. Anthropomorphizing AI?
The disagreement concerns both what AI might experience and how to act when we cannot confidently know. Tabarrok’s historical analogy places greater weight on the moral cost of mistakenly denying consciousness; Cowen expresses much greater confidence that sentience is the wrong category to apply. These are their philosophical assessments, not established findings about AI consciousness.
AI dominates human peer reviews
We show that AI-generated article reviews are now better, by every measure we could construct, than reviews written by well-compensated human experts and judged to be of unusually high quality by journal editors and authors. We then discuss how this result may inform journal policy choices, with concerning downsides as well as potential paths to increase publication speed and quality; reduce adversarial interactions; improve the experience of editors, reviewers, authors, and readers; and advance science.
By Gary King and Steven Worthington, not even using what are now the very best models, rather based on Claude Opus 4.8. The details of the dominance, as outlined in the paper, are striking. It is not close.
On my Saez and Zucman post, I find it striking that some people are complaining all of a sudden. Yet I have been offering, citing, and linking to commentary on published papers for twenty-four years now (and every single day). Sometimes it might just be a casual offhand remark from me, whether positive or negative. Typically this is simply taken in stride. It is when you pull out the bazookas, and offer oodles of high-powered analytics, more astute than human referees, and combined with the ability to pursue follow-up inquiries, that people start getting worried. Model that.
You can pick at individual criticisms from the AIs all you want, but the most damning observation is that the errors go in the ideological direction of the authors 97% of the time. That is exactly what (the human) Phil Magness said years ago, based on an extensive canvassing of their work. Basically we already knew that was true, and the AIs confirmed the human estimate.
On our side, we have taken steps, most of all the new MR archive, to make AI analysis of our work easier. Soon you will be able to read on MR what the AIs think are our worst posts, respectively (Alex has prepared this).
In the meantime, I encourage the NSF, prize committees, universities, tenure committees, journals, and so on to do as many of their own AI-enabled audits as possible. So far the world is not jumping at this opportunity. You can model that too.