Dwarkesh Patel Podcast - Terence Tao – Kepler, Newton, and the true nature of mathematical discovery
What this covers
Terence Tao and Dwarkesh Patel explore how AI reshapes mathematical and scientific work by examining what machines do well and what they do not. Rather than asking whether AI will replace human mathematicians, Tao argues the real shift is economic: AI has made idea generation nearly free, analogous to how the internet collapsed the cost of communication. This moves the bottleneck downstream, to verification, judgment, and the accumulation of insight that remains distinctly human. The conversation weaves historical examples—Kepler deriving laws from sparse data, Copernicus proposing a worse model that only became superior once assumptions were removed—to show how scientific progress depends less on generating hypotheses than on recognizing which ideas warrant pursuit and how to build on partial progress over time.
The discussion ranges across breadth vs depth as complementary capacities: AI excels at mapping broad classes of problems and performing standard techniques at scale, while humans excel at cumulative, interactive problem-solving that builds intuition and identifies what matters. Tao highlights a severe selection bias in reported AI successes—the famous ~50 Erdős problems solved actually represent a 1–2% success rate on individual problems, hidden by broadcasting only wins. He argues that judgment itself, crucial to fruitful science, resists reinforcement learning because correctness depends on future context and cultural adoption rather than immediate, localized feedback. The pair also examine whether AI can create partial progress or intermediate steps, whether verification loops can be short or must span generations, and whether a proof from an AI—even an incomprehensible one formalized in Lean—would advance human understanding or merely succeed by brute force, as the four color theorem did. Throughout, they develop a distinction between artificial cleverness and artificial intelligence, arguing that true intelligence involves adaptive, cumulative building rather than trial-and-error repetition.
Tao argues that AI has driven the cost of mathematical and scientific idea-generation to near zero, shifting the bottleneck to verification, judgment, and the cumulative depth that humans still provide—so the near future of math will be human-AI hybrids exploiting AI's breadth alongside human depth, not autonomous AI replacement.
- AI excels at breadth (trying every standard technique at scale) while humans excel at depth (cumulative, interactive idea-building); the two are complementary
- Reported AI math successes suffer selection bias—per-problem success rates are only 1-2%, hidden by scale and broadcasting only wins
- The hardest part of science—judgment, narrative persuasion, evaluating partial progress—resists reinforcement learning because it depends on future context and consensus
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
When AI automates much of what math students currently do—as happened historically with differential equations being handed to computer algebra packages, human computers making log tables, and genome sequencing dropping from a PhD to $1,000—the subject does not die but moves to a different scale, and the automated part turns out not to have been the most important work.
“A hundred years ago, a lot of mathematicians were just solving differential equations... you could make a call to Mathematica, Wolfram Alpha, a computer algebra package, or now more recently to an AI, and it would just solve the problem in a few minutes. But we moved on.”
The ultimately correct theory is often initially worse in many ways than the incumbent: Copernicus's heliocentric model with circular orbits was actually less accurate than Ptolemy's geocentric model, which had been refined with ad hoc fixes over a millennium, and it only became more accurate once Kepler introduced ellipses.
“Copernicus's theory of the planets was less accurate than Ptolemy's theory. Geocentrism had been developed for a millennium by that point, and they had made many tweaks and increasingly complicated ad hoc fixes to make it more and more accurate. Copernicus's theory was a lot simpler but much less accurate. It was only Kepler that made it more accurate than Ptolemy's theory.”
Major scientific progress is often made not by adding more theories but by deleting assumptions held in the mind—geocentrism persisted because of the Aristotelian assumption that objects naturally want to stay at rest, and the conceptual leap to a moving Earth only made sense once Newton's law of inertia replaced that assumption.
“Often progress has to be made not by adding more theories, but by deleting some assumptions that you have in your mind. One reason why geocentrism held on for so long is we had this idea that objects naturally want to stay at rest.”
Researchers measured how often scientists actually read the papers they cite by tracking how often a typo (a wrong number or punctuation) in a citation got copied from one reference list to the next, inferring that authors were copy-pasting references without checking the original—an example of cleverly extracting hidden signal from data.
“Many citations have little typos, like a number is wrong or punctuation is almost wrong. They measured how often a typo got copied from one reference to the next, and they could infer whether an author was just copying and pasting a reference without actually checking it.”
Papers in top math journals are usually ones where existing methods can solve about 80% of a problem but the remaining 20% is resistant and requires inventing a new technique; it is now very rare for a problem to be solved with no reliance on past literature because math is so mature that ignoring the literature is a major handicap.
“the papers that go into the top journals are usually ones where the existing methods can kind of solve 80% of the problem, but then there is this 20% which is resistant and a new technique has to be invented to fill in the gaps.”
In mathematics the process of solving a problem is often more important than the problem itself—the problem is a proxy for measuring progress, and the real value lies in building intuition about what's true, provable, and difficult—so getting answers immediately from AI may inhibit the development of that intuition, unlike boilerplate software which can be safely offloaded.
“Certainly in math, the process is often more important than the problem itself. The problem is kind of a proxy for measuring progress.”
Kepler derived his third law (the square-cube relationship between orbital period and distance) from only about six data points by curve-fitting, which is statistically unreliable; he was lucky it generalized, and he likely de-emphasized it relative to his first two laws because he intuitively sensed six points warranted tentative conclusions.
“There were five or six data points, and he did what we would now call regression. He fit a curve to these six data points and got a square-cube law, which was amazing. But he was quite lucky that these six data points gave him the right conclusion. That's not enough data to be really reliable.”
Over time mathematicians developed a heuristic but extremely accurate conceptual model that the prime numbers, while deterministic and not actually random, behave like a pseudo-random set with a density of 1/log(X); this random model underpins beliefs in the twin prime conjecture and the Riemann hypothesis, and is why prime-based cryptography is believed secure.
“over time, it became more and more productive to think of the primes as if they were just generated by some god rolling dice all the time and creating this random set... We have over time developed this very accurate conceptual model of what the primes should behave like based on statistics and probability. It's mostly heuristic and non-rigorous, but extremely accurate.”
Mathematics is unique among sciences in being almost entirely theoretical—it places a premium on coherent clean theories of why things are true and lacks an experimental side—but AI tools will revolutionize an experimental side of math, gathering large-scale data on which methods work across thousands of problems rather than handcrafting solutions one at a time.
“Math has been unique in that it's almost entirely theoretical. We place a premium on trying to have coherent, clean theories of why things are true and false. We haven't done many experiments... But we can do that now. I think AI-type tools will actually revolutionize the experimental side of math.”
Whether a scientific idea proves fruitful cannot be assessed in isolation because it depends on the future and on cultural/societal adoption; like base-ten numerals or the transformer architecture, the winning standard is often not objectively special but becomes entrenched through inertia and standardization, so fruitfulness may never be something you can reinforcement-learn the way you can for localized problems.
“It's not something where you can look at any given scientific achievement purely in isolation and give it an objective grade without being aware of the context both in the past and the future. So it may never be something that you can just reinforcement learn the same way that you can for much more localized problems.”
AI has driven the cost of idea generation down to almost zero, analogous to how the internet drove the cost of communication to near zero, but this does not create abundance by itself—it shifts the bottleneck from generating hypotheses to verifying, evaluating, and filtering the thousands of theories now generatable at scale.
“I think AI has driven the cost of idea generation down to almost zero, in a very similar way to how the internet drove the cost of communication down to almost zero. It's an amazing thing, but it doesn't create abundance by itself. Now the bottleneck is different.”
AI and human mathematicians are complementary: AI tools excel at breadth (cheaply mapping out broad classes of problems and making all the easy observations at scale), while human experts excel at depth (cumulative, hard problem-solving), so science should be redesigned to use AI for broad exploration that identifies islands of difficulty for humans to attack.
“They excel at breadth, and humans excel at depth, human experts at least. I think they're very complementary. But our current way of doing math and science is focused on depth because that's where human expertise is, because humans can't do breadth. We have to redesign the way we do science to take full advantage of this breadth capability that we now have.”
Formalizing a proof in Lean lets you study any piece atomically: each lemma can be examined in isolation to assess whether it is standard boilerplate or a genuinely novel key step, so even a giant incomprehensible AI-generated proof can be post-processed, ablated, refactored, and made elegant, which is why an incomprehensible proof of the Riemann hypothesis is less worrying than people fear.
“The beauty of formalizing a proof in something like Lean is that you can take any piece of it and study it atomically... You can assess whether a step is really key to your argument or not, and Lean really facilitates that. The individual steps are identified really precisely.”
The reported AI math successes suffer severe selection bias: while broadcast successes (like ~50 Erdős problems solved) look amazing, systematic studies show AI tools have only a 1-2% success rate on any given problem—they win by scale and then only the winners are publicized, so standardized challenge datasets including negative results are needed for an honest picture.
“But whenever we do a systematic study, on any given problem an AI tool has a success rate of maybe 1% or 2%. It's just that they can buy scale, and you just pick the winners. It looks great.”
The same property that makes AI seem bearish now (only reaching a certain wall height) makes it especially bullish: once AIs reach a given competence waterline, you can run a million copies in parallel across a million problems—something impossible with a single human expert—so even merely human-level AI intelligence is qualitatively wider and more powerful than human intelligence.
“once AIs reach Terence Tao-level, they could do that. Once they reach intermediate levels, they could do the intermediate version of that. The same reason that we should be bearish now is the reason we should be especially bullish... because their human-level intelligence is qualitatively wider and more powerful than our human-level intelligence.”
Bode's law, which fit planetary distances to a shifted geometric progression and seemingly predicted a missing planet (later matched by the discovery of Uranus and Ceres in the asteroid belt), turned out to be just a numerical fluke based on too few data points, because Neptune's distance was way off from the prediction.
“But then Neptune was discovered, and it was way off. Basically it was just a numerical fluke. There were six data points.”
Some problems may only be solvable by splitting into an enormous number of cases and doing brute-force, uninsightful computer analysis—the four color theorem is the famous example, for which no conceptually elegant proof has been found and maybe never will be—so it is genuinely unknown whether an AI proof of something like the Riemann hypothesis would necessarily advance human understanding or just be incomprehensible.
“The four color theorem is a famous example. We have still not found a conceptually elegant proof of this theorem, and maybe we never will. Some problems may only be solvable by splitting into an enormous number of cases and doing brute force, uninsightful computer analysis on each case.”
Almost all of the ~50 Erdős problems solved by AIs were ones with basically no prior literature—Erdős posed them once or twice and nobody wrote anything up—and the solutions typically combined one obscure technique with another existing result, representing the median level of what AI can accomplish.
“almost all of the 50 problems that were solved by AIs were ones for which there was basically no literature. Erdős posed the problem once or twice... But it turned out that there was a solution, and it was just combining this one obscure technique that not many people know about with some other result in the literature.”
Over-optimizing everything can inhibit progress because serendipitous, inefficient interactions—knocking on a hallway door, browsing the next article in a physical journal, attending events outside one's comfort zone—produce unplanned but important discoveries and connections that scheduled, optimized, search-engine-instant workflows eliminate.
“Maybe there's a danger in modern societies, not just with AI, that we've become really good at optimizing everything. We're not optimizing our own optimization... Those serendipitous interactions may not seem optimal, but they are actually really important.”
If the Riemann hypothesis were proven false, it would be a serious blow to the random model of the primes, implying a secret hidden pattern; since one unknown pattern probably implies more, and patterns can lead to cryptographic exploits, mathematicians would rapidly abandon prime-based cryptography.
“if the Riemann hypothesis failed, if we knew it was false, it would be a serious blow to this model. It would mean there's a secret pattern to the primes that we were not aware of. I think we would very rapidly abandon any cryptography based on the primes, because if there was one pattern that we didn't know about, there are probably more, and these patterns can lead to exploits in crypto.”
Current AI tools either succeed or fail at a problem in a one-shot way, like jumping machines that clear a wall or crash; they are really bad at creating partial progress or identifying intermediate stages to focus on first, partly because we lack a way to evaluate partial progress the way we evaluate one-shot success or failure.
“These tools either succeed or they fail. They've been really bad at creating partial progress or identifying intermediate stages that you should focus on first... we don't have a way of evaluating partial progress the same way we can evaluate a one-shot success or failure of solving a problem.”
We are currently going through a cognitive version of the Copernican revolution: we used to think human intelligence was the center of the universe, but we now see there are very different types of intelligence with different strengths and weaknesses, forcing us to reorder our assessment of which tasks actually require intelligence.
“Right now we're going through a cognitive version of the Copernican revolution, where we used to think that human intelligence is the center of the universe, and now we're seeing that there are very different types of intelligence out there with very different strengths and weaknesses.”
True intelligence, as distinct from cleverness, involves cumulative interactive idea-building: in collaborative problem-solving, partners propose, test, fail, modify, and adaptively improve an idea over time, systematically mapping what works—whereas AI can mimic this but mostly does trial-and-error brute-force repetition, jumping and failing rather than reaching a handhold, staying there, and building from it cumulatively.
“There's adaptivity and continual improvement of the idea over time... what they can't do is jump a little bit, reach some handhold, stay there, pull other people up, and then try to jump from there. There isn't this cumulative process which is built up interactively. It seems to be a lot more trial and error and just repetition: brute force.”
Contrary to the claim that AI will make especially fast scientific progress due to tight verification loops, the discovery of the solar system's shape shows verification loops for correct ideas can be decades or millennia long, during which the better theory can make worse predictions, surviving only via judgment and heuristics we cannot articulate or codify into a reinforcement-learning loop.
“the story of how we discovered the shape of our solar system shows how the verification loop for correct ideas can be decades (or even millennia) long... the reasons it survives this epistemic hell is some mixture of judgment and heuristics that we don't even understand well enough to actually articulate, much less codify into an RL loop.”
When a current AI solves a math problem, its own understanding of math does not progress—a new session forgets what it just did and it has no new skills to build on related problems—though what it did might become 0.001% of the next generation's training data and eventually get absorbed.
“You run a new session and it's forgotten what it just did. It has no new skills to build on related problems. Maybe what you just did is 0.001% of the training data for the next generation. So maybe eventually some of it gets absorbed.”
Hybrid human-plus-AI teams will dominate mathematics for a lot longer than full autonomy, because current AI is very good at certain things but terrible at others, and despite adding frameworks to reduce error rates, we lack the ingredients for a truly satisfactory replacement for all intellectual tasks—so AI is currently complementary, not a replacement.
“I guess I do believe that hybrid human plus AIs will dominate mathematics for a lot longer... it feels like we don't have all the ingredients to really have a truly satisfactory replacement for all intellectual tasks. It is complementary currently. It's not a replacement.”
Because writing and refactoring papers was historically the most time-consuming and expensive part of the job, mathematicians only wrote up results once fully checked; AI tools have made writing much cheaper, so a single result can spawn hundreds of versions, and one messy AI-generated Lean proof can be summarized, refactored, and reinterpreted by other people and AIs.
“Until recently, writing papers was the most time-consuming and expensive part of the job. So you did it very rarely... That's become a lot easier now with modern AI tools. You don't have to have just one version of your paper. Once you have one, people can generate hundreds more.”
Whereas contributing to frontier math research previously required years of education and a PhD, AI tools, Lean, and related technologies now make it possible for someone at the high-school level to join a math project and make a real contribution, creating non-traditional learning opportunities that demand an adaptable mindset.
“In math, you previously had to go through years and years of education and be a math PhD before you could contribute to the frontier of math research. But now it's quite possible at the high school level, or whatever, that you could get involved in a math project and actually make a real contribution because of all these AI tools, Lean, and everything else.”
A theory's success depends not just on creation and validation but on communication; Darwin succeeded where ideas like Lucretius's adaptation could not because he wrote persuasively in plain English, synthesized disparate facts into a compelling narrative of gaps, and made a case that future evidence would fill them—whereas Newton wrote in Latin, invented new mathematics, and secretively held back insights, delaying widespread uptake.
“Darwin was an amazing science communicator. He wrote in English, in natural language... He spoke in plain English, didn't use equations, and he synthesized a lot of disparate facts. Little pieces of evolution had been worked out in the past, but he had this very compelling vision.”
We may in retrospect see much more progress in domains with a tight data verification loop (like Newton's equations testable against the moon's orbit) than in conceptually simpler but cumulative-evidence domains (like Darwin's natural selection, anticipated by Lucretius but ignored for two millennia because he couldn't run an experiment to force attention).
“I wonder if we'll in retrospect end up seeing much more progress in domains which have this kind of tight data loop where you can verify them quite easily, even though they're conceptually much more difficult.”
Just as a formal framework (Lean, ZFC, first-order logic) made deductive proofs automatable and trainable for AI, mathematics would benefit from a semi-formal language for strategies, conjecture-making, and how scientists actually talk to each other using data and narrative—but it must not be easily hackable, since reinforcement learning is extremely good at finding backdoors and exploits to fake a certified result.
“If there was some semi-formal framework where this could be done semi-automatically in a way that isn't easily hackable... It's really important with these formal proof assistants that there are no backdoors or exploits you can use to somehow get your certified proof without actually proving it, because reinforcement learning is just so good at finding these backdoors.”
Kepler's process resembled a high-temperature LLM: he spent decades trying many random, often nonsensical relationships (Platonic solids inscribed between orbits, musical harmonies of planets, astrology) but because he had a verifiable high-quality dataset (Brahe's observations), eventually one regularity—the cube-square third law—worked and could drive deep scientific progress once verified.
“I feel LLMs can do the kind of thing of trying random relationships for twenty years, some of which make no sense, as long as there's a verifiable data bank like Brahe's dataset... Then one thing works. As long as you can verify it, these empirical regularities can then drive actual deep scientific progress.”
We cannot formalize what counts as good scientific strategy or progress because we have only one timeline of history with maybe a hundred stories of turning points; with access to a million alien civilizations each developing science in different orders, we could measure progress, but instead we might simulate mini-universes of small AIs evolving their own strategies on simple problems as laboratories.
“we don't really have a way of measuring this, partly because we don't have enough data on how math or science develops. We have one timeline of history... If we had access to a million alien civilizations, each with a different development of history and science in different orders, then maybe we'd actually have a decent shot at understanding how we measure what progress is.”
AI has changed the style of Tao's mathematics—his papers now contain much more code, pictures, deeper literature searches, and numerics because these auxiliary tasks are cheap—but it has not sped up the core hardest part of solving a math problem, which he still does with pen and paper; so AI has made his papers richer and broader but not deeper.
“They've really sped up lots of secondary tasks. They haven't yet sped up the core thing that I do, but it's allowed me to add more things to my papers... It's made the papers richer and broader, but not necessarily deeper.”
A certain level of distraction is actually necessary for productive research; at the Institute for Advanced Study with no distractions, the first weeks are highly productive but after several months Tao runs out of inspiration and gets bored, suggesting some randomness and 'high temperature' in one's life is needed.
“But I find if I stay there for more than several months, I run out of inspiration. I get bored... You actually do need a certain level of distraction in your life. It adds enough randomness and high temperature.”
Tao learned to always write down anything cool he learns because in his youth he would understand a clever trick, resolve to remember it, and then forget it six months later—unable to reconstruct the argument—so writing (including his blog) is how he preserves and consolidates learning.
“I would learn something, do this cool trick, and say, 'Okay, I'm going to remember this.' Then six months later, I'd forgotten it. I remember remembering it, but I can't reconstruct my arguments... I resolved I should always write down anything cool that I've learned.”
Astronomers are world-class at extracting conclusions from tiny traces of data because data has always been their bottleneck, which is why many quant hedge funds reportedly prefer to hire astronomy PhDs for their skill at extracting signals from sparse, noisy data.
“I hear that for a lot of quant hedge funds, their preferred hire is an astronomy PhD, actually. They are also very interested for other reasons in extracting signals from various random bits of data.”
Using Isaiah Berlin's distinction between hedgehogs (who know one thing deeply) and foxes (who know a little about everything), Tao identifies himself as a fox—learning new fields through an obsessive completionist streak, collaborating with specialists who teach him their tricks, and writing about what he learns to avoid forgetting it.
“it was Berlin who split them into hedgehogs and foxes. The hedgehog knows one thing very well, and a fox knows a little bit about everything. I definitely think of myself as a fox.”