“In the history of science and that of ideas, the thickness of time is not uniform.”
- John A. Goldsmith and Bernard Laks, Battle in the Mind Fields
“Part of the problem is the extraordinary place economics currently holds in the social sciences. In many ways it is treated as a kind of master discipline. Just about anyone who runs anything important in America is expected to have some training in economic theory, or at least to be familiar with its basic tenets. As a result, [its] tenets have come to be treated as received wisdom, as basically beyond question (one knows one is in the presence of received wisdom when, if one challenges it, the first reaction is to treat one as simply ignorant — ‘You obviously have never heard of the Laffer Curve’; 'Clearly you need a course in Economics 101’ — the theory is seen as so obviously true that no one who understands it could possibly disagree.)”
- David Graeber, Debt: The First 5,000 Years
“I read the news stories and never know quiet how much is hype and how much is reality. I’m highly uncertain about what the economic pay-off to AI is.”
- Paul Krugman, The Wolf-Krugman Exchange
As a student, I worked security at the Oxford Union Society, which mostly consisted of checking member cards at the door of the debate club and its bar. Occasionally, I had to eject an inebriated socialite or aristocrat from a wood paneled library. Most nights were slow, so I’d pick up The New York Times and Financial Times before my shift. I’d browse the headlines during lulls, but paid special attention to the economics columns. Back then, Martin Wolf and Paul Krugman were towering figures. Both were major advocates of globalization in the fin de siècle Anglosphere. Krugman’s New Trade Theory was received wisdom. He literally wrote the textbook I used in international economics. To disagree with him implied one didn’t understand economics, at least in my tutorials at Oxford. This was unfortunate for both Krugman and the field of economics. Interestingly, this was before he won the Nobel Prize, but obviously after the John Bates Clark Medal.
Wolf and Krugman recently had a few conversations and shared them as podcast episodes. The two towers provide some much-needed illumination in dark times. But their fourth exchange centered entirely around artificial intelligence, a topic on which neither is an expert. It was disappointing. I hope they’ll take up the topic again in the future, but prepare more first.
Their conversation disappointed me because there’s so much more substance Krugman could have offered, but didn’t. In particular, I’m surprised he didn’t say more about DeepSeek and US-China relations or the potential efficacy of US controls on the export of advanced graphics processing units from NVIDIA. He didn’t even touch on TSMC and Taiwan’s geopolitical importance or speculate about America’s prospects for reshoring semiconductor production.
Surely, there was plenty of good in the conversation. But it recalled a passage from Battle in the Mind Fields, Goldsmith and Lak’s delightful history of linguistics. “Isaiah Berlin, the most profound raconteur one would ever hope to meet, wrote about his life in philosophy, and he put his finger on an interesting phenomenon that is not at all uncommon, and by its very nature involves the group within which one works–and in part, but only in part, its size. He wrote about what happens when one chooses an artificially small and personal group of associates to serve as one’s intellectual cohort…
"One of the shortcomings of these meetings is something that seems to me to apply to Oxford philosophy in general, at least in those days. We were excessively self-centered. The only persons whom we wished to convince were our own admired colleagues. There was no pressure upon us to publish. Consequently, when we succeeded in gaining from one of our philosophical peers acceptance or even understanding of some point which we regarded as original and important, whether rightly or, as was more often the case at any rate with me, in a happy state of delusion, this satisfied us completely, too completely. We felt no need to publish our ideas, for the only audience which was worth satisfying was the handful of contemporaries who lived near us, and whom we met with agreeable regularity.”
- Isaiah Berlin, Personal Impressions
Philosophy of Science and AI were my focus at Oxford, but it was a great place to study economics, and I read the subject extensively while I was there. In my career, I’ve spent a lot of time thinking about economics and AI, both abstractly and in negotiations. In their conversations, Wolf sometimes comes off as an energetic Oxford apprentice eagerly thinking through multiple possibilities, and Krugman as a seasoned and respected don, nervous to be on the wrong side of history, uncertain yet about what to say. While both seem hung over from web3, Wolf is enthusiastically prepared to tackle many subjects within the scope of AI. His interlocutor, not so much. I understand that Krugman is not an expert on AI specifically and that economics is a dismally inexact science. Krugman even says so. But as much as he takes pains to explain how the manufacturing sector took 40 years to fully adopt electricity, it took economists a hundred years to figure out how to conceptualize the impact of technological innovation. (For his work on Endogenous Growth Theory, Paul Romer shared the 2018 Nobel Prize in Economics.)
Krugman is right to point out that the future is broadly uncertain. But clearly, AI is not crypto. He is right to point out that AI consumes vast amounts of capital and power, to admit that if they are “stochastic parrots, they do really useful stuff.” And to compare its potential impact on knowledge work to the impact that technological innovation had on the coal industry. And he’s right to point out the potential impact of Deep Research on finance-focused cities like New York and London. And he’s right to point out the curious and sudden collapse of the employment market for young graduates, including from top computer science programs and elite universities.
But Ethan Mollick et al’s collaborative research with BCG on the jagged edge of AI, plus the rapid pace of improvement of performance of newer models combined with the declining cost of inference, suggest that the shock could eventually spread and become secular. Krugman is a professor and an economist. Still, he has honed a polemical style through popular writing and there’re many ways for an economist to lend relevant insights on the subject. Instead, we got banal provocations like “Turing was wrong.” Meanwhile, they don’t use the phrase human-in-the-loop even once.
Back when I was a student, it was an essay on the philosophy of artificial intelligence that accompanied the application that won me admission. Specifically, I designed an alternative to the Turing Test. This is just to say: a) I appreciate Krugman’s intelligence and capacity to contribute to the discourse and b) I’ve also disagreed with Turing for a long time. But I think the evidence that this time is different has been building up, so he could do a lot better if he knows where to look. There’s plenty of precedent from other economists who have deployed the tools of their profession to better understand the new technology and its impact.
For instance, three economists working with the Federal Reserve Bank of St. Louis, shared research on the speed of adoption of generative AI in comparison to personal computing and the Internet. It took 3 years for PCs to penetrate to 20 percent of American adults. The Internet reached the 20% mark in 2 years, i.e. 33% faster. In less than 2 years, ChatGPT was adopted by 40% of American adults, i.e. double the penetration of the Internet in less time. What’s particularly funny about this recent working paper from the St. Louis Fed is that it perfectly functions as an update to Paul A. David’s 1990 paper, “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox”, which Krugman specifically cites, while ignoring Paul Romer’s paper that was published the same year.
Douglass North once said that, after winning the Nobel Prize, people seemed afraid to disagree with him. And that worried him. Before the prize, North couldn’t get folks to listen to him about the topics on which he was an expert. After winning, he was taken to be infallible, even in fields where he had little background knowledge. (Krugman won the Nobel in 2008.)
I wouldn’t call it epistemic trespassing, but one of the valid concerns we have about new technologies is that they encourage us to be lazy. At a moment when the world’s only super power has imposed sanctions on the world’s most valuable company specifically designed to slow down the progress in AI research of its chief national rival (i.e. NVIDIA had to invent the H20 chip just for China); when the world’s wealthiest companies and people are deploying massive shares of their capital on video game chips and city-scale power supply to engage in a race to superintelligence; where GPUs allow humans to convert electricity into intelligence, but where the quality of intelligence output is directly correlated with the talent of those who train and fine tune one’s models - in such a moment, we need our greatest minds to not be lazy.
Even more breathtaking than the rate of user growth has been the pace of improvement of LLMs. In late 2023, the most advanced models were achieving less than 10% proficiency on SWE-bench. In May 2025, OpenAI’s o3-high (their frontier reasoning model) achieved 85% success on SWE-bench tasks on eight attempts. This is junior programmer level intelligence.
Last year, the Palmyra-Fin model passed the CFA Level II exam and received a score of 73% on the Level III, better than most humans, on a 0-shot attempt. These are not trivial achievements. And the models keep getting better. We’re literally running out of tests. The benchmarks are almost all saturated.
As performance has improved, costs have come down substantially. Aspiring analysts in London, New York, and Cambridge would be right to fret.
The DeepSeek slump in early 2025 demonstrated both that the US does not have a monopoly on AI research breakthroughs and raises the question: Are American frontier labs wasting compute? With DeepSeek V3, we saw that one can achieve frontier level performance with FP8 rather than FP64 precision. R1 demonstrated that further jumps in performance could be achieved through throwing compute at the problem during test time, so the scaling law will not hit any walls any time soon. (Wolf drew the exact opposite conclusion from the evidence.)
As we move from an emphasis on pre-training to inference, to satisfy our ever-hungrier and ever-larger language models with synthetic data, increased memory and interconnects will become more important than raw processing capacity. The gated H20 chips that have been designed for the Chinese market are constrained on processing power, ie fewer FLOPs over time. But if the paradigm is shifting to reasoning, and DeepSeek has demonstrated that frontier level performance can be achieved with less precision, then the export restrictions are based on false premises. What will matter more in the future is memory. And the H20 chips actually have MORE memory than the export restricted H100s.
The world has changed much. And it will continue to change with increasing speed. If GPUs allow us to convert electrons into intelligence, then we still have to convert that intelligence into knowledge, and that knowledge into productivity. But that’s not the question. The question is how will we adapt to a world wherein one superhuman-in-the-loop can do the work of 1000 knowledge workers. To address such questions, we need more humans like Krugman in the loop.
