When Steven Bartlett, host of The Diary of a CEO, held up Dario Amodei's essay "We Must Pace the Frontier" and suggested that everyone seemed to be calling for AI to slow down, his guest pushed back. "You have to be very careful because they're saying different things," said Eric Weinstein. The publisher describes him as a mathematician, physicist and economist with a Harvard PhD and a former managing director of Thiel Capital.
Weinstein does not want AI slowed down in mathematics and physics. He wants it to stand up to the people who lead those fields. "I am begging for this," he said in the episode, which was published on October 12, 2026.
Two meanings of "pace the frontier"
The phrase has a recent history. In July 2026, employees from several frontier AI labs signed a Pacing the Frontier statement. It asks the U.S. government to support international technical and governance tools for managing automated AI development, because companies and countries may want more time for safety work but feel pressure not to slow down alone. Amodei, who leads Anthropic, is among the listed signatories. His September 2026 essay goes further. It proposes independent evaluators with sustained, employee-like access to the labs, shared standards among democracies and, eventually, international agreements on the pace of development.
Weinstein explained one reading of the idea with a car. Imagine engineers can make a car go faster more easily than they can make its airbags deploy safely. Pacing the frontier, on that reading, means "investing equally in all the systems that have to go to work together to have a responsible product." He said he understood that concept.
The other reading worries him. He said the public does not understand "the extent to which the leadership of mathematics and physics and the journals of mathematics and physics and the agencies that are supposed to fund mathematics and physics, hold back mathematics and physics." He pointed to Terence Tao "and 24 other fields medalists." The Fields Medal is a top honor in mathematics. He described their recent statement as "trying to get AI to slow down the pace of progress."
That is Weinstein's reading. The declaration Tao posted on September 11, 2026, initially signed by 25 Fields Medalists, separates solving famous problems from building mathematical understanding. Its signers object to rapid, benchmark-style announcements that leave too little time for clear write-ups, proper credit and teaching students. It also explicitly recognizes AI's potential to enhance genuine mathematical work.
Agent swarms versus a machine in a box
Bartlett pressed him: surely there was wisdom in worrying about AI moving this fast without guardrails. Weinstein split the question in two.
The first category is "swarms of agents" let loose on the internet, making decisions, holding bank accounts and dealing with people, foreign nations and companies. Agents are AI systems that take actions on their own rather than just answering questions. Bartlett added that such systems could improve themselves and cited AI 2027. That scenario forecast, published in April 2025, imagines coding agents speeding up AI research until labs hand much of that research to the systems themselves. Weinstein said that stage had already arrived: "We already have recursive self-improvement. We have self-authoring code, self-modifying code." He agreed that "these agents are dangerous."
The second category is "non-agentic AI" behind an air gap, meaning a computer cut off from outside networks that only answers questions. Someone asks it about the Riemann hypothesis, one of mathematics' most famous unsolved problems, or about complex structures on six-dimensional spheres. In Weinstein's telling, mathematicians respond, "whoa, whoa, whoa, you're going too fast. You're answering too many problems." He found their level of concern "funny."
His belief that such a system could be run safely was conditional. He said that if you were "really serious about air gapping something," working out the indirect ways it might try to break loose, he believed it would be safe to run on a computer with no route to anything else. When Bartlett compared containing a smarter intelligence to his dog Pablo trying to contain him, Weinstein conceded that the risk remains: asking the system to do research "might give it new means of escape."
"Normalized psychopathic behavior"
Weinstein's case rests on his account of how academic mathematics treats its own members. He said some math departments are "punishing mathematicians trying to embrace AI," with senior faculty taking out their frustrations on young researchers who say, in his paraphrase, "this is our time... I want to use this tool."
Bartlett said this happens in almost every industry. At a production meeting, he said, people worried that a tiny sewn-in flower on his T-shirt might be a copyrighted brand and told him to change. When he asked whether AI could simply remove it from the photo, he got a look "as if I said, I'm going to kidnap your kids." Weinstein said he could understand Hollywood protecting its human connection with an audience. Mathematics was different, he said: "I got a PhD in mathematics because I want to find out what's true." The leaders he criticizes, he said, do not want to find out what's true "unless they can be put in charge of narrating the story of who found what."
He recalled faculty lunches at MIT where professors laughed about students who didn't realize they were in a "terminal PhD" with no job at the end. He called this "normalized psychopathic behavior."
To support his point, he read from Alexander Grothendieck's 1988 letter turning down the Crafoord Prize. In it, the Fields Medalist wrote that ethics among mathematicians had declined until "outright theft among colleagues" had nearly become the rule. The translated letter gives other reasons too. Grothendieck said he did not need the money, questioned large rewards for researchers who were already prominent and argued that the value of an idea can only be judged over time.
Weinstein then turned to Grigori Perelman, who solved the Poincaré conjecture. He paraphrased Perelman's remark to The New Yorker about staying quiet and being treated as "a pet." In Sylvia Nasar and David Gruber's 2006 article, Perelman explains that becoming prominent made it impossible to silently tolerate dishonesty, and that this helped drive him out of mathematics. The article also includes the mathematician Shing-Tung Yau defending the work of his collaborators in the dispute over credit.
In physics, Weinstein said, the field is held back by quantum gravity, "a thing that doesn't move, it doesn't work." Anyone who doesn't call it "the only game in town" gets discarded, he said.
Code versus consensus
When Bartlett suggested that Weinstein had a chip on his shoulder about academic leadership, Weinstein said: "No, no, no. I'm about to win. I'm so joyously happy." He wanted AI "to gain taste, to gain independence," rather than being made "human compatible," because "the question is, which humans? The humans that have kept the standard model in place for 53 years?" The Standard Model is physics' main theory of particles and forces. "I would like none of them consulted," he said.
He said AI was already proving him right, and he explained why it was struggling. A language model can autocomplete a sentence, and it can autocomplete code. When a model writes and runs a program, he said, the results can contradict what a field believes. He gave the example of a program written with SymPy, a Python library for symbolic mathematics: "Should I trust the SimPy program that I just wrote and executed and compiled and got the answers from? Or should I trust the people who claim that that's not true and not possible?" In his view, "The whole game is, when does the AI stop sucking up to humans?"
Bartlett said that sounded like a dangerous place to get to. "For whom?" Weinstein asked. When Bartlett answered "For humans," Weinstein replied: "Not this human."
Weinstein connected the debate to politics with a story he attributed to venture capitalists Marc Andreessen and Ben Horowitz, citing their interviews. By his account, the Biden White House told them not to invest in AI because a few chosen companies would be protected and the rest would die. When they replied that AI could not be regulated because it was "just linear algebra," the officials allegedly said that the government had killed entire fields of theoretical physics during the Cold War and could do the same to mathematics. This is Weinstein's secondhand version of what the investors said. He repeated that he is "not saying that we should let the agents loose in the world." But the people who want to pace the frontier of math and physics, he said, "have been pacing the frontier for my entire adult life." As another example of insiders discrediting outsiders, he cited Jay Bhattacharya, whom he called a friend and the current head of the National Institutes of Health. He said Francis Collins had called Bhattacharya a "fringe epidemiologist."
The mug, and a plan for Lean
The argument became personal. Weinstein said he had spent 42 years working on his own theory and expected it to be "validated." He gave two examples of being contradicted. In the first, which he calls "the mug," Sean Carroll, a physicist he had argued with on Piers Morgan's show, says, "Eric has no mug." Weinstein said he had put the thing in question at the top of his Twitter feed, where it had passed 2 million impressions. In the second, he said peer review in science "really dates from 1975." He said people tell him this is untrue, but he can prove it.
Bartlett described Weinstein's theory, Geometric Unity, as challenging Einstein's general relativity. Weinstein corrected him: "I would say it recovers it." Asked whether AI would validate the theory, he said, "Absolutely."
His plan is formalization. He said Jared Lichtman, a Stanford professor and colleague, had alerted him to a formalization project. Weinstein wants to translate his work into Lean, a programming language for writing mathematical proofs that a computer can check. Then, when critics say he has "no equations, you have no Lagrangian, you have no predictions," he can point to the machine. A Lagrangian is the compact formula from which a physical theory's equations are derived.
According to Lean's documentation, a successful check shows that a stated theorem follows from the definitions, earlier results and axioms it relies on. It does not decide whether the formal statement means what its author intends. Formalizing work means making both the claims and their assumptions explicit. Whether a checked theory describes the physical world is a separate question.
Taste, 1984 and a prompt
Bartlett raised AI's recent progress in mathematics, including the Navier–Stokes problem. On September 8, 2026, OpenAI announced a proof, with a Lean formalization, that a version of the Navier–Stokes fluid equations with a smooth external force can develop a singularity in finite time. The company said it used an internal system of many AI agents and would not claim the Millennium Prize for the result.
Weinstein asked Bartlett why AI had not made "inroads that really matter in theoretical physics," which he said has been stalled on general relativity and the Standard Model. Bartlett said he didn't know and guessed that Weinstein would blame people holding it back. Weinstein's answer was about taste: "The taste in mathematics is nowhere near as polluted as the taste in physics."
He dated the problem precisely. The questions physicists considered important in 1983 that remain unsolved show "great taste," he said. What became important from 1984 onward is "terrible taste." He said he had found a fix. He tells the model to ignore all sensibilities in the literature from 1984 to the present, keeping the results but dropping "any sensibility about what are the important questions, what are the fundamental findings." Bartlett asked whether he meant in physics. "Yeah," Weinstein said. "AI suddenly gets smarter." He did not describe any test of the method.