OpenNashPodcasts

Dwarkesh Podcast

Terence Tao – Kepler, Newton, and the true nature of mathematical discovery

20 Mar 2026 · 1h 23m · Terence Tao, Dwarkesh Patel · 24 quoted facts

▶ Play episodeOriginal source ↗

In brief

Terence Tao argues that AI has made generating scientific ideas almost free, but science depends on checking those ideas, and checking is the bottleneck. Kepler's story shows that verifying a correct idea can take decades, and a correct theory can look worse than the old one for a long time. Judging which ideas matter relies on judgment, persuasion and the test of time, which are hard to turn into a training signal.

The central tension is breadth versus depth. AI tools can try every standard technique on many problems at once and clear the easy ones, but success rates on any single problem are about 1-2%, and they do not build up progress step by step. Tao expects human-AI teams to dominate for a long time, and he wants science redesigned to use AI's breadth.

Key insights

  1. Verification, not idea generation, is now the bottleneck

    Tao says AI has cut the cost of generating theories to almost zero, just as the internet did for communication. Journals are already flooded with AI-generated submissions, and peer review cannot scale to a thousand ideas a day.

  2. Correct theories often start out less accurate

    Copernicus's sun-centered model was simpler but less accurate than Ptolemy's heavily patched Earth-centered one until Kepler improved it. This means a simple test of whether a theory predicts better can reject the right idea.

  3. AI math successes look bigger than they are

    About 50 Erdős problems were solved with AI help, mostly ones with little prior literature. Systematic sweeps show a per-problem success rate of only 1-2%, and the headline wins come from picking the winners.

  4. AI broadens papers but doesn't deepen the core

    Tao says AI helps him add code, plots and deeper literature searches, and would save about five times the effort on those parts. But he still solves the hardest steps with pen and paper, so the papers are richer and broader, not deeper.

  5. Today's AI lacks cumulative, adaptive problem-solving

    Tao describes AI as jumping repeatedly and failing, without grabbing a foothold and building from it. A new session forgets what the last one learned, so there is no accumulating understanding.

Questions

Why does Terence Tao say AI won't automatically speed up science?

Tao says AI makes ideas cheap, but each idea still has to be checked, and checking is slow and hard to automate. Kepler's laws took decades to be explained, and the correct theories sometimes looked worse than the old ones.

How well has AI actually done on the Erdős math problems?

About 50 of roughly 1,100 problems were solved with AI help, mostly ones that had little earlier literature. Tao says that in systematic tests an AI succeeds on a given problem only about 1-2% of the time, and that pure AI solutions have stopped appearing.

Is AI making Terence Tao twice as productive?

He says it is hard to measure that way. Papers he would now write would take about five times longer without AI, but only because of extras like code, plots and literature searches, while the hardest core reasoning is unchanged.

What is Lean and why does it matter for AI in math?

Lean is a formal proof language that checks every step of a proof, so a computer can verify it. Tao says this makes deductive proofs easy to automate and to train AI on, and lets people study each step of a giant AI-made proof on its own.

What career advice does Tao give to people entering math?

He says to expect a lot of change and to stay adaptable, and to keep following curiosity. He thinks traditional education is still worth getting for now, and that AI tools and Lean may let even high schoolers contribute to research.

Facts

Stated by the speakers. Each quote matches the transcript, which shows what was said, not that it is true.

Johannes Kepler proposed that the five Platonic solids could be nested between the spheres of the six known planets, matching their orbit sizes.

“there were six planets, none of the time, five gaps between them, and there were five perfect platonic solids, the cube, the tetrahedron, isochedron, octahedron, and dodecaedron.”

0:01:25 · Terence Tao · Kepler's model

Tycho Brahe persuaded the Danish government to fund an expensive observatory where he recorded decades of planetary observations.

“Tycho Brahe, this Danish astronomer, very wealthy, eccentric astronomer, had managed to convince the Danish government to fund this extremely expensive observatory.”

0:01:53 · Terence Tao · Astronomy data

Tycho Brahe was the last major astronomer to make his observations with the naked eye.

“With the naked eye, actually, he was the last of the naked eye astronomers.”

0:02:22 · Terence Tao · Astronomy data

Tycho Brahe's data did not match Kepler's Platonic-solid theory, which disappointed Kepler.

“he worked out to kind of his disappointment that his beautiful theory didn't quite work like the data was sort of off from his platonic solid theory”

0:02:22 · Terence Tao · Kepler's model

Kepler found that planetary orbits are ellipses rather than circles, which he found shocking.

“he eventually worked out that the ellipses were actually ellipses, not circles, which was shocking for him.”

0:02:58 · Terence Tao · Planetary motion

Kepler gave no explanation for his three laws, and Isaac Newton supplied a theory explaining all three about a century later.

“it took Newton a century later to give a theory that explained all three laws at once.”

0:03:33 · Terence Tao · Planetary motion

Kepler's third law was fitted to only about six data points, one for each known planet's orbital period and distance from the sun.

“Kepler had like six data points um like every planet uh you knew the length of the orbit and the distance of the sun”

0:09:35 · Terence Tao · Kepler's third law

The distance of Uranus, discovered by William Herschel, fit the geometric pattern proposed by Johann Bode.

“when Uranus was discovered by Herschel, the distance Uranus fit exactly this pattern.”

0:10:49 · Terence Tao · Bode's law

Neptune's discovery showed that Bode's distance pattern did not hold, so it was a numerical fluke.

“But then Neptune was discovered and it was completely like way off.”

0:11:19 · Terence Tao · Bode's law

Dwarkesh Patel says Kepler developed his polygon and Platonic-solid theories in 1595 and 1596, and that they were wrong.

“1595 and 96 is where he comes up with first polygons and then platonic objects theory, but they were wrong.”

0:08:48 · Dwarkesh Patel · Kepler's model · 1595

Aristarchus proposed heliocentrism in the third century BC, and critics objected that the stars should show parallax.

“Aristarchus had heliocentrism in the third century BC and then the ancient Athenians were like, this can't be because if the earth is going around the sun, we should see the relative position of the stars change”

0:17:35 · Dwarkesh Patel · Heliocentrism · 3rd century BC

Darwin's On the Origin of Species appeared in 1859, about two centuries after Newton's Principia in 1687.

“The Origin of Species is published in 1859. The Principia Mathematica is published in 1687.”

0:21:35 · Dwarkesh Patel · Scientific history · 1859

Tao says Newton wrote in Latin and invented new mathematics, and that his work spread widely only after other scientists explained it more simply a couple of decades later.

“Newton wrote in Latin. he had invented you know entire new areas of mathematics just to explain what he was doing”

0:23:42 · Terence Tao · Scientific communication

Tao says almost all of the roughly 50 Erdős problems solved by AI had little or no prior literature.

“Erdős posed the problem once or twice. I think maybe some people tried it casually and they couldn't do it, but they never wrote up anything.”

0:43:31 · Terence Tao · Erdős problems

Tao says systematic studies show AI tools succeed on any given hard math problem only about 1-2% of the time.

“any given problem, an AI tool has a success rate of maybe 1 % or 2%.”

0:44:34 · Terence Tao · AI success rate · 2026

Jane Street trained a ResNet, shuffled its 96 layers, and challenged people to restore the order using only the model's outputs and training data.

“Jane Street trained a ResNet and then shuffled all 96 layers and then challenged people to put them back in the right order”

0:27:34 · Dwarkesh Patel · Jane Street puzzle

Tao says the four-color theorem was proved essentially by brute-force computer checking, and no conceptually elegant proof has been found.

“A full-color theorem is a famous example. We have still not found a conceptually elegant proof of this theorem.”

0:53:38 · Terence Tao · Four color theorem

Carl Friedrich Gauss computed the first 100,000 or so primes, creating one of the first mathematical data sets.

“Gauss was interested in the prime numbers, and he computed, he created one of the first mathematical data sets.”

1:03:19 · Terence Tao · Prime numbers

Gauss conjectured the prime number theorem, that the number of primes up to x is roughly x divided by the natural logarithm of x.

“The number of primes up to x is like x divided by the natural log of x.”

1:03:57 · Terence Tao · Prime number theorem

The twin prime conjecture, that infinitely many pairs of primes differ by two, remains unproved, though the random model of primes makes mathematicians confident it is true.

“there's a still open conjecture in number theory called the twin prime conjecture, that there should be infinitely many pairs of primes that are twins”

1:05:28 · Terence Tao · Twin primes

Tao says that if the Riemann hypothesis were false it would seriously undermine the random model of the primes and could threaten cryptography based on them.

“if the real hypothesis failed um we knew it was false it means that it would be a serious blow to this model.”

1:06:37 · Terence Tao · Riemann hypothesis

Tao says the standard axioms of mathematics, known as ZFC, and of first-order logic were formalized only in the early 20th century.

“only in like the early 20th century did we finally list our, here are the axioms of mathematics, well, the standard axioms of what we call ZFC”

0:59:34 · Terence Tao · Mathematical logic · early 20th century

Tao notes that sequencing one organism's genome once took an entire PhD, but now costs about $1,000 and can be sent to a sequencer.

“So in genetics, to sequence the genome of a single organism, that was an entire PhD of a geneticist.”

1:18:52 · Terence Tao · Genetics

Terence Tao spent a year at the Institute for Advanced Study and said he ran out of inspiration after several months there.

“I spent a year once at the Institute for Advanced Study, which is a great place to, you know, there's no distractions.”

1:15:46 · Terence Tao · Tao's career

Chapters

  1. 0:03:33How Kepler found the laws of planetary motion

    Tao retells how Kepler moved from a Platonic-solids theory to three laws using Tycho Brahe's data. It sets up the episode's question about discovery and verification.

    “he had no explanation for them. It was just all driven by experiment”
  2. 0:12:18Is Kepler a model for AI hypothesis generation?

    Dwarkesh argues Kepler was like a high-temperature LLM trying random relationships against data. Tao replies that generation is cheap and verification is the real constraint.

    “AI has basically driven the cost of idea generation down to almost zero.”
  3. 0:18:48Why correct theories can look worse at first

    Copernicus and Newton both had theories that seemed worse or incomplete at first. This shows why judging ideas in the moment is hard to automate.

  4. 0:44:34Erdős problems and the AI success rate

    Tao explains that AI solved about 50 Erdős problems, then stalled, and that success stories hide a 1-2% per-problem hit rate.

    “any given problem, an AI tool has a success rate of maybe 1 % or 2%.”
  5. 0:35:19Breadth versus depth in AI-assisted math

    The pair discuss how AI's breadth and human depth are complementary, and how it could change the way science is organized.

  6. 0:56:25Cleverness versus intelligence, and Lean

    Tao says AI lacks cumulative progress and discusses how formal proofs in Lean could make AI-made proofs understandable.

    “you can take any piece of it and study it atomically.”
  7. 1:03:57Primes, conjectures and formalizing strategy

    Using Gauss and the prime number theorem, Tao explains how mathematicians build confidence in conjectures and why a formal language for strategy is hard.

    “So he conjectured what we now call the prime number theorem.”
  8. 1:19:53Serendipity, the future and career advice

    Tao talks about learning habits and serendipity, predicts human-AI hybrids will dominate math, and advises adaptability.

    “I do believe that hybrid human plus AIs will dominate mathematics for a lot longer.”

Transcript

0:00:00
Dwarkesh Patel

Okay, today I'm chatting with Terence Tao, who needs an introduction. Terence, I want to begin by having you retell the story of how Kepler discovered the laws of planetary motion, because I think this will be a great jumping off point to talk about AI for math.

0:00:15
Terence Tao

Okay, yeah. So I've always had an amateur interest in astronomy, and so I've loved stories of how the early astronomers worked out the nature of the universe. So Kepler was building on the work of Copernicus, who was himself building on the work of Aristarchus. So, Copernicus very famously proposed the heliocentric model, that instead of the planets and the sun going around the Earth, that the sun was at the center of the solar system and the other planets were going around the sun.

0:00:42

And Copernicus proposed that the orbits of the planets were perfect circles. And his theory kind of fit the observations that the Greeks and the Arabs and the Indians had worked out over centuries. theories. I think Kepler got interested, like he learned about these theories in his studies, and he made this observation that the ratios of the size of the orbits that Kuhnick predicted seemed to have some geometric meaning. I think he started proposing that if you take, say, the orbit of, say, the Earth and you enclose it in, I think, maybe a cube, the outer sphere that that encloses the cube, almost matched perfectly the orbit of Mars and so forth.

0:01:25

And there were six planets, none of the time, five gaps between them, and there were five perfect platonic solids, the cube, the tetrahedron, isochedron, octahedron, and dodecaedron. And so he had this theory, which he thought was absolutely beautiful, that he could inscribe these platonic solids between the spheres of the planets, and it seemed to fit. And it seemed to be to him like, you know, God's design of the planets was matching this mathematical perfection of the platonic solids.

0:01:53

So he needed data to confirm this theory. And at the time, there was only one really high-quality data set almost in existence, which was the... So Tycho Brahe, this Danish astronomer, very wealthy, eccentric astronomer, had managed to convince the Danish government to fund this extremely expensive observatory. In fact, an entire island where he had taken decades of observations of all the planets, Mars, Jupiter, every night, at least every night for which the weather was clear.

0:02:22

With the naked eye, actually, he was the last of the naked eye astronomers. And so he had all this data which Kepler could use to confirm his theory. And so Kepler started working with Tycho, but Tycho was very jealous of the data. He only gave little bits of it at a time. And I think Kepler eventually just stole the data. He copied it and had to have a fight with Brahi's descendants. but he did work out he did get the data and then he worked out to kind of his disappointment that his beautiful theory didn't quite work like the data was sort of off from his platonic solid theory by you know about 10 or something and he tried all

0:02:58

kinds of fudges moving the circles around and things it didn't quite work but he worked on this problem for for years and years and eventually he figured out how to use the data to work out the actual orbits of um of the planets and that was an incredibly clever, genius amount of data analysis. And then he eventually worked out that the ellipses were actually ellipses, not circles, which was shocking for him. And then he worked out the two laws of planetary motion, ellipses, also equal areas, sweep out equal times.

0:03:33

And then 10 years later, after collecting a lot of data, the furthest planets like Saturn and Jupiter were the hardest for him to work out. But then he finally worked out this third law also, that the orbits, the time it takes for a planet to complete its orbit was proportional to some power of the distance to the sun. And these are the three famous capital laws of motion, and he had no explanation for them. It was just all driven by experiment, and it took Newton a century later to give a theory that explained all three laws at once.

0:04:10
Dwarkesh Patel

The take I want to try on you is that Kepler was a high-temperature LLM. Where Newton comes up with this explanation of why the three laws of planetary motion must be true. And of course, the way that Kepler discovers the laws of planetary motion or figures out the relative orbits of the different planets is, as you say, a work of genius. But then, you know, through his career, he's just trying random relationships. And in fact, in the book in which he writes down the third law of planetary motion it's sort of an aside on the harmonics of the world which is this book about you know all these different planets have these different harmonies and the reason there's so

0:04:46

much famine and misery on earth is because the earth is me for me that's a note of earth and so all this random astrology but in there is the cube square law which tells you what relationship the uh the period has to a planet's distance from the sun which is as you're detailing uh if you add that to newton's f equals ma and then the equation for centripetal acceleration you get the inverse square law and so newton works that out but the reason i um i think this is an interesting story is i feel like lms can do the kind of thing of like 20 years let's try random relationships some of which make no sense as long as there's

0:05:23

a verifiable data bank like brahi's data set where okay i'm going to try out random things about like musical notes i'm to try out random things about platonic objects. I'm going to, all these different geometries, I have this bias that there's some important thing about the geometry of these orbits. And then one thing works. And as long as you can verify it, it can then drive, these empirical regularities can then drive actual deep scientific progress. Traditionally, when we talk about the

0:05:45
Terence Tao

history of science, idea generation has always been kind of the prestige part of science. So, I mean, a scientific problem comes with, there's many steps, you know, you have to identify a problem and then you have to identify a good problem to work on, a fruitful problem. And then you need to collect data. You need to figure out a strategy to analyze the data, to make a hypothesis. And at this point, you need to propose a good hypothesis.

0:06:07

And then you need to validate. And then you need to write things up and explain. There's a dozen different components. But the ones we celebrate are these sort of eureka genius moments of idea generation. And yeah, so Kepler certainly had to, as you say, cycle through many ideas, and several which didn't work and and i bet many that he didn't even um publish at all um because yeah they just didn't fit and that's an important part of the process um trying all kinds of random things and seeing if they worked um but as you say the um you know the uh it has to be matched by an equal amount of verification otherwise it's it's slop you know i mean

0:06:49

um we celebrate kepler but should also celebrate brahi for for his his assiduous data collection with which was ten times more precise than than any previous observation and it was um that extra decimal point of accuracy was actually essential for for kepler to get uh his um um his results um and you know and he was using you know euclidean geometry and and and like like the most advanced mathematics he could use at the time to to match his his models of the data so like all aspects had to be in play You know, the data and the theory and the hypothesis generation.

0:07:26

I'm not sure nowadays that hypothesis generation is the bottleneck anymore. Science has changed in the centuries since. So classically, sort of the two big paradigms for science were theory and experiment. Then in the 20th century, numerical simulation came along. And so you can also do computer simulations to test theories. But then finally, in the late 20th century, we had big data. We had the era of data analysis. And so a lot of new progress is actually driven now by analyzing massive data sets first, collecting large data sets, and then drawing the patterns from them to deduce laws, which is a little bit different from how science used to work, where you make a few observations or you just have one

0:08:13

out of the blue idea. And then you collect data to test your idea. That's the classic scientific method. Now it's almost reversed. You collect big data first, and then you try to get hypotheses from it. I mean, Kepler was maybe one of the first early data scientists, but even he didn't start with Tycho's data set and analyze it. He had some preconceived theories first. But it seems like this is less and less the way we make progress just because the data is just so much more massive.

0:08:46

is just somewhat more useful.

0:08:48
Dwarkesh Patel

Oh, interesting. I actually feel like the mold of 20th century science that you're describing is actually very well describes what happened in Kepler, where he did have these ideas, 1595 and 96 is where he comes up with first polygons and then platonic objects theory, but they were wrong. And then a few years later, he gets Brahe's data. And it's only after 20 years of just trying random things that he gets this empirical regularity. And so it actually feels closer to Brahe's data is analogous to some massive data bank of simulations.

0:09:22

And then now that you've got the data, you can keep trying random things. But if it wasn't, Kepler would be out there just writing books about harmonics and platonic objects, and there would be nothing to actually verify against.

0:09:35
Terence Tao

Yeah, yeah, yeah. So the data was extremely important. um but the distinction you're trying to make is that sort of traditionally you make a hypothesis and then you test it against data yeah um but um now with um machine learning and data analysis and statistics and something you can you can start with data and um through say statistics work out laws that um were not present before so and so kepler so kepler's third law is a little bit like this except that uh for the third law instead of having the thousand data points that brahi had kepler had like six data points um like every planet uh you knew the length of the orbit and the distance of the sun

0:10:13

and there was like five or six data points and he did uh what we would now call regression you know he could fit a curve to these six data points and he got a square couplau which was amazing uh but actually he was quite lucky i mean that these six data points gave him the right conclusion um you know it's uh that's not enough data to be really reliable um there was a later astronomer, Johannes Borde, who took the same data, actually, the distances to the planets, and inspired by Kepler, I think, he had a prediction that the distances to the planets formed basically a shift to geometric progression, that he also fit a curve.

0:10:49

Except there was one point missing, so there was a big gap between Mars and Jupiter. His law predicted that there was a missing planet, so it was a kind of a crank theory, except when Uranus was discovered by Herschel, the distance Uranus fit exactly this pattern. And then Ceres was discovered, this asteroid between, I think, in the asteroid belt. And it also fit the pattern. So people got really excited that the board had discovered this amazing new law of nature.

0:11:19

But then Neptune was discovered and it was completely like way off. And, you know, and basically it was just a numerical fluke. You know, there were six data points. yeah so maybe one reason why kepler didn't highlight his third law as much as the first two laws is that maybe instinctively even though he didn't have modern statistics he kind of knew that with six data points he had to be somewhat tentative with the conclusions but maybe to

0:11:45
Dwarkesh Patel

ask the question about the analogy more explicitly does this analogy make sense to if we have you know in the future we'll have smarter and smarter ais and we'll have millions of them and then they can go out and hunt for all these empirical regularities. It sounds like you don't think the bottleneck in science is finding more things that are for each given field, they're equivalent of the third law of planetary motion so that then later on somebody can say, oh, we need a way to explain this.

0:12:13

Let's work out the math. Here's the inverse square law of gravity. Right. So I think

0:12:18
Terence Tao

AI has basically 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, which is an amazing thing, but it doesn't create abundance by itself. So now the bottleneck is different. So we're now in a situation where suddenly people can generate thousands of theories for a given scientific problem. And now we have to verify them, evaluate them. And this is something which we have to change our structures of science to actually sort this out.

0:12:51

So in fact, traditionally, we build walls you know so in the past you know before we had uh ai slop you know we had sort of amateur scientists you know create you know have their own theories of the universe uh many of which uh were basically of very little value yeah um and so we brought these like you know peer review publication systems and things to kind of filter out um and try to isolate the high signal um ideas to uh to test um but uh but now that we can generate these these possible explanations at massive scale.

0:13:24

And some of them are good and a lot are terrible. I mean, human reviewers, they're already being overwhelmed, actually. Many, many journals are reporting AI general submissions are just flooding their submissions. So it's great that we can generate all kinds of things now with AI, but it means that the rest of the aspects of science have to catch up. Yeah, so verification, validation, and assessing what ideas actually move the subject forward and which ones are dead ends or red herrings.

0:13:57

And that's not something we know how to do at scale. You know, for each individual paper, we can discuss it with, you know, have a debate among scientists and get to a consensus in a few years. But when we're generating, you know, a thousand of these every day, this doesn't work.

0:14:12
Dwarkesh Patel

Yeah. So I think there is this incredibly interesting question of you have billions of AI scientists. Not only how do you gauge which ones are real progress, but how do you I mean, this is actually a question that human science has had to face and we've solved somehow. And I actually am not sure how we solve this, but in any given field, let's say in the 1940s, and if you're at Bell Labs or if you're just generally trying to, there's these new technologies coming out.

0:14:36

Pulse code modulation, basically how do you transfer signals, how do you digitize signals, how do you transfer them over analog wires. And then, but there's like all these papers about the engineering constraints there and the details. And then there's one which is like, comes up with the idea of the bit, which has implications across many different fields. And you need some system which can then look at that and say, OK, we need to apply this to probability.

0:14:56

We need to apply this to computer science, et cetera. And in the future, the AIs are coming up with the next version of this kind of unifying concept. And how would you identify it among millions of papers which might actually constitute progress, but which have much less general unifying ideas?

0:15:14
Terence Tao

So a lot of it's the test of time. So many great ideas didn't actually get a great reception at the time that they were first proposed. it was only after some other scientists realized that they could take it further and apply them to their own. Deep learning itself was actually a niche area of AI for a long time. The idea of getting answers entirely through training on data and not through first principles, you know, reasoning was very controversial and it took a long time before it actually started bearing fruit.

0:15:43

You know, you mentioned the bit, you know, I mean, there are other proposals for computer architectures than the zero one that is universal today i think there were there were tricks you know zero one three valued logic and you know in an alternate universe maybe a different paradigm would have would have showed up um people have argued that you know the transformer for example is is the foundation of all modern large language models and it was the first um deep learning architecture that really was was sophisticated enough to capture language but it didn't have to be that way there could have been some other architecture that um was the first to do it and once that was adopted it would

0:16:20

become the standard um so i think um one reason why it's hard to assess whether a given idea is going to be fruitful is that it it depends on the future it depends on and it depends on also on the culture and society like like which ones get adopted which ones don't um you know um the base 10 um numeral system in mathematics extremely useful much better than the roman numeral system for instance um but again there's nothing special what 10 it's it's a system that we it's useful for us because everyone else uses it um and we've standardized it and we've brought all our computers and our and our number of representation systems around it um and so we're

0:16:59

stuck with it now actually um you know people are some people occasionally push for other systems than decimal but um it's this this is no this is no uh there's too much inertia um so you can't 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. And so it may never be something that you can just reinforce and learn the same way that you can for much sort of more localized problems.

0:17:31

Yeah. It seems often in the history of science

0:17:35
Dwarkesh Patel

when a new theory comes up that in retrospect we realize is correct, it seems to make implications that just either make no sense because they're wrong and we realize later on why they're wrong or they're correct but seem wildly plausible at the time. So as you've talked about, Aristarchus had heliocentrism in the third century BC and then the ancient Athenians were like, this can't be because if the earth is going around the sun, we should see the relative position of the stars change as we're going around the sun.

0:18:07

And the only way that wouldn't be the case is if they're so far away that you don't notice any parallax, which is actually the correct implication. But there's times when actually the implication isn't correct and we just need to graduate to a better level of understanding. So Leibniz would, you know, chide Newton and disagree with Newton's theory of gravity on the basis that it implied action at a distance. And then there's, we don't know the mechanism.

0:18:30

And Newton himself was sort of stunned that inertial mass and gravitational mass were the same quantity. So all these things which were resolved by Einstein. Yes, yes. But it was still progress. And so the question for a system of peer-reviewed for AI would be, even if you can falsify a theory, how would you notice that it still constitutes progress relative to the thing before?

0:18:48
Terence Tao

Yeah, so often actually the ultimately correct theory initially is worse in many ways. Yeah, so Copernicus's theory of the planets, it was less accurate than Tomli's theory. So geocentrism had been developed for a millennium at that point, and they had made many, many tweaks and increasingly complicated ad hoc fixes to make it more and more accurate. And Copernicus' theory was a lot simpler, but much less accurate. It was only Kepler that made it more accurate than Tomlin's theory.

0:19:20

I mean, science is always a work in progress. So when you only get part of the solution, it looks worse than a theory which is incorrect, but somehow it has been completed to the point where it kind of answers all the questions. As you say, Newton's theory had big mysteries, the equivalence of mass and action at distance, which were only resolved with a very conceptually different approach centuries afterwards. often progress has been made not by adding more theories, but by deleting some assumptions that you have in your mind.

0:20:02

So one reason why geocentrism held on for so long is we had this idea that objects naturally want to stay at rest. This is the Aristotelian notion of physics. And so the idea that the Earth was moving, how come we weren't all falling over? Once you have neutrons in motion, object in motion remains in motion and so forth, then it makes sense. But you had to, so conceptually, it's a very big conceptual leap to realize that the Earth is in motion.

0:20:30

It doesn't feel like it's in motion. And like the biggest advances, you know, Darwin's theory of evolution, you know, is the idea that species are not static. But, you know, it's not obvious because you don't see evolution in your lifetime. Well, now we actually can, but it seems permanent and static. 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 actually seeing that there's very different types of intelligence that are out there with very different strengths and weaknesses and so our assessment of which tasks require intelligence and which ones don't has to be reordered quite a bit.

0:21:21

And so, you know, trying to fit AI into sort of our theories of scientific progress and what is hard and what is easy, we're struggling quite a lot. We have to ask questions that we've never really had to ask before. Or maybe the philosophers had, but now we all have to deal with it.

0:21:35
Dwarkesh Patel

This actually brings up a topic I've been very curious about. So you mentioned Darwin's The Era of Evolution. There's this book, The Clockwork Universe, by Edward Dahlnack, which covers a lot of this era of history we're talking about. And he has this interesting observation in there that The Origin of Species is published in 1859. The Principia Mathematica is published in 1687. So The Origin of Species comes out basically two centuries after the Principia. And conceptually, it seems like Darwin's theory is simpler.

0:22:02

There's a contemporaneous biologist to Darwin who reads The Origin of Species, Thomas Huxley, and he says, how stupid not to have thought of that. And nobody ever says that about Principia. They're chiding themselves for not having beaten Newton to gravity. And so there's a question of, well, why did it take longer? It seems like a big part of the reason is that the evidence for natural selection is cumulative and retrospective, whereas Newton can just like, here's my equations.

0:22:27

Let me see the moon's orbital period and its distance. And if it lines up, then we've made progress. And so Lucretius actually had the idea, this idea that species adapted their environment in the first century BC, but nobody ever really talks about it until Darwin because Lucretius can't run some experiment and people are forced to pay attention. And so 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.

0:23:01
Terence Tao

I think one aspect of science is it's not just creating a new theory and validating it, but communicating it to others. So Darwin was actually an amazing science communicator. He wrote in English, in natural language, I'm speaking like that. In Olin. Okay. Yeah, okay. I have to sort of get out of my technical mindset. He spoke in plain English. You know, didn't use equations. and he synthesized a lot of disparate facts. So, you know, little pieces of evolution had been worked out in the past, but he had this very compelling vision and again, still missing things.

0:23:42

Like he didn't know the mechanism for hereditary. He didn't have DNA. Yeah, but his writing style was persuasive and that helped a lot. Newton wrote in Latin. he had invented you know entire new areas of mathematics just to explain what he was doing he was also from an era which was where scientists were much more secretive and competitive so you know academia is still competitive it was even worse back in Newton's day so he he held back some of his best insights because he didn't want his rivals to get any advantage he was also a somewhat unpleasant person from what I what I what I gather actually so it was actually only a couple decades after Newton where other scientists

0:24:26

explained his work in much simpler terms that they became widespread. So yeah the art of exposition and making a case and creating a narrative is also a very important part of science and if you have the data and it helps but people need to be convinced otherwise they will not push it further or they want to take initial investment to learn your theory and really explore it. And that's another thing which is really hard to reinforce and learn on.

0:25:01

How can you score how persuasive you are? Well, there's the entire marketing departments who are trying to do this. So maybe it's good that AI are not yet optimized to be persuasive. So yeah, there's a social aspect to science. Even though we pride ourselves on having an objective side to it where there's data and there's experiment and validation um we still have to tell stories and convince our fellow scientists um and that's a soft squishy thing like it it's you know it's it's a combination of data and um yeah and uh painting a narrative and it's a narrative of gaps you know i mean as you know so so even darwin as i said there are pieces of his theory

0:25:47

he could not explain. But he could still make a case that in the future, people would find transitional forms, that they would find the mechanism of inheritance. And they did. Yeah. I don't know how you can quantify that in such a precise way that you can start to reinforce some learning. Maybe that will be forever the human side of science.

0:26:11
Dwarkesh Patel

One takeaway I had from reading and watching your stuff on the Cosmic Distance Latter. By the way, I highly, highly, highly recommend people watch your series with through the one round on the cosmic distance ladder. But one takeaway was that the deductive overhang in many fields could be so much bigger than people realize where if you just had the right insight about how to study a problem, you might be surprised at how much more you could learn about the world.

0:26:38

And I wonder if you think that's sort of a product of astronomy at the particular times in history that you're studying, or is this that based on the data that is incident on the Earth right now, we could actually divine a lot more than we happen to know?

0:26:51
Terence Tao

Right. So astronomy was one of the first sciences to really embrace data analysis and squeezing every last possible drop of information out of the information they had because data was the bottleneck. I mean, it still is a bottom. I mean, it's really hard to collect astronomical data. So astronomers are the best, almost, world-class in extracting, almost like Sherlock, extracting all kinds of conclusions from little traces of data. I hear that a lot of quant hedge funds, they're preferred hires in astronomy PhD.

0:27:28

They also are very interested for other reasons in extracting signals from various randomness of data.

0:27:34
Dwarkesh Patel

Okay, speaking of clever ideas, one of my listeners, Sean, solved the puzzle that Jane Street made for my audience and posted a great walkthrough on X. For context, Jane Street trained a ResNet and then shuffled all 96 layers and then challenged people to put them back in the right order using only the model's outputs and training data. You can't brute force this. There's more possible orderings than atoms in the universe. So Sean broke the problem into two different parts.

0:28:00

First, pair the layers into 48 different blocks, and second, put those blocks in the right order. For pairing, Sean realized that in a well-trained resonant, the product of two weight matrices in a residual block should have a distinctive negative diagonal pattern. And this arises as a way for the model to keep the residual stream from growing out of control. From this insight, he was able to recover the right pairings. For ordering, Sean noticed that the model seemed to improve if he sorted the blocks by the size of their residual contributions.

0:28:29

Starting with that rough approximation, he combined a clever ranking heuristic with local swaps to recover the exact right order. His full walkthrough is linked in the description. Don't worry if you didn't get to this puzzle in time, though. There's still one up about backdoored LLMs that even Jane Street doesn't know how to solve. You can find it at janestreet.com/dwarkesh. All right, back to Terrence.

0:28:50
Terence Tao

We do underexplore sort of how to extract extra information from various signals. like just to pick one random study I remember reading once that people had discovered were trying to measure how often scientists actually read these citations the papers that they cite so how do you measure this you could try to survey different scientists but they had some clever tricks so many citations have little typos like a number is wrong or punctuation symbol is wrong And they measured how often a type of word got copied from one reference to the next.

0:29:32

And they could infer whether an author was actually just copying it, cutting and pasting a reference without actually checking it. And so from that, they were able to infer some measure of sort of how much attention people were paying. So there are also clever tricks to extract. you know so these questions you posed earlier of you know how can we assess whether um a scientific development is fruitful or uh or interesting or represents real progress you know maybe there are um really useful metrics and or um footprints of this of this of this um of this phenomenon in in a data data so we can we can examine citations and um and like how often something is mentioned in a

0:30:16

conference or something and maybe that there's there's a lot of uh social sociology of science research to be to be done and and that could actually um detect these things um yeah maybe we usually get some astronomers on the case actually um okay so i think this brings

0:30:32
Dwarkesh Patel

us uh nicely to the progress that from the outside it seems like ai for math is making and i think you had a post recently where you pointed out that over the last few months AI programs have solved 50 out of the 1 ,100-odd ERDOs problems. But then I think, I don't know if it's still correct, but as of a month ago, you said that there had been a pause because the low-hanging fruit had been picked.

0:30:55

First of all, I'm curious if actually that is still the case, that we have picked the low-hanging fruit and now we're at this plateau currently.

0:31:02
Terence Tao

It does seem so. I mean, there's still activity at the ERDOs. Yeah, so 50-odd problems have been solved with AI assistance, which is great, but there's like 600 to go. and people are still chipping away at one or two of these right now. We're seeing a lot fewer sort of pure AI solutions now where the AI just one-shots the problem. So there was a month where that happened and that has stopped. Not for lack of trying.

0:31:28

I know three separate attempts to get frontier model AIs to just attack every single one of the problems simultaneously. And they picked up some minor observations or maybe they found that some problems are already solved in the literature, but there hasn't been any further AI purely powered solution yet. People are using AI a lot currently. So someone might use AI to generate a possible proof strategy and then another person will use a separate AI tool to critique it or rewrite it or generate some numerical data for it or do a literature survey.

0:32:04

And some problems have been solved by an ongoing conversation between lots of humans and lots of ai tools um but uh it it does seem like it it was this this one-off thing um so maybe one analogy to for these problems is like um imagine like um there's this there's all these that you're in some sort of mountain range with all kinds of of cliffs and walls and and uh maybe there's a there's a there's a little um wall which is maybe like three feet high and one there's six feet high and then there's 15 feet high and then there's there's there's some mile high cliffs um and you're trying to climb as many of these cliffs as possible

0:32:40

but it's in the dark uh we don't know which ones are tall which ones are short and um so you know we try to light some candles and make some maps and and slowly we kind of figure out uh some of them are climbable some of them we can identify some some partial um track in the wall that you can reach first um and then these these ai tools they're kind of like these jumping machines that can kind of jump you know two meters in the air you know higher than any human and sometimes they jump in the wrong direction and sometimes they crash but sometimes they they can reach um um the tops of of the lowest um you

0:33:15

know um walls that we couldn't reach before and so we basically set them loose in this mountain range hopping around and you know and then there's this exciting period where they could actually find all the um all the low ones um and they could reach them um But then there's been no, I mean, maybe if the next time there's a big advance in the models, then they will try it again and maybe a few more will be breached.

0:33:42

But it's a different style of doing mathematics than sort of the, you know, so normally we would hill climb and, you know, we would make little markers and try to identify partial things. And, you know, these tools, they either succeed or they fail. And they've been really bad at creating sort of partial progress or identifying intermediate stages that you should focus on first. Again, going back to this previous discussion, you know, we don't have a way of evaluating partial progress.

0:34:14

The same way you can evaluate a one-shot success or failure of solving a problem.

0:34:19
Dwarkesh Patel

So there's two different ways to think through what you've just said. And one of them is more bearish on AI progress and one of them is more bullish and bearish on being, oh, they're only getting to a certain height of wall, which is not as high as humans are reaching. And the second is that, well, they have this powerful property that once they achieve a certain waterline, they can fill every single problem that is available at that waterline, which we simply can't do with humans, where we can't make a million copies of you and give each of them $1 million of inference compute.

0:34:50

and have you do 100 years of subjective time research on 100 different problems at the same time or a million different problems at the same time. But once AIs reach Terence Tower level, they could do that. And once they reach intermediate levels, they could do the intermediate version of that. So the same reason that we should be bearish now is the reason we should be especially bullish, not even when they achieve superhuman intelligence, but just when they achieve human level intelligence because their human level intelligence is qualitatively wider and more powerful

0:35:19
Terence Tao

than our human level intelligence i i agree yeah so they excel at breadth uh and humans excel at depth um like human experts at least yeah so um i think they're very complementary um but our current uh way of doing math and science is focused on depth because that's where the human expertise is because humans can't do breath um but uh yeah so we have to redesign uh the way we do science to take full advantage of of this breadth capability that we now have um so as i said we do uh we should have a lot more effort in creating very broad classes of problems to work on rather than one or two um really um deep important problems

0:35:58

i mean we should still have the deep important problems um and humans should still be working on them um but but now now we have this other way of of of doing um of doing science you know i mean uh we can explore entire new fields of science by first getting these broad, moderately competent AIs to sort of map it out and clear out all the easy, make all the easy observations, and then identify certain islands of difficulty, which, you know, then human experts can come and work on.

0:36:28

So I see very much a future of very complementary science. Eventually, you would hope to get both breadth and depth, you know, and somehow get the best of both worlds. But I think we need practice with the breadth side. It's too new. We don't even have the paradigms really to make full advantage of it. But we will. And then science will be unrecognizable after that.

0:36:55
Dwarkesh Patel

To this point about complementarity, programmers have noticed that they're way more productive as a result of these AI tools. And I don't know if you as a mathematician feel the same way, but it does seem like one big difference between vibe coding and vibe researching is that with software, the whole point of the thing is to have some effect on the world through your work. And if it leads to you better understanding a problem or you coming up with some clean abstraction to embody in your code, that is instrumental to the end goal.

0:37:29

Whereas maybe with research, the reason we care about solving the Millennium Pipers problems is presumably that in the process of solving them, we discover new mathematical objects or new techniques and those who understand our civilization's understanding of mathematics. And so the proof is sort of instrumental to the intermediate work. I don't know if you agree with that dichotomy or if that in any way will explain the relative uplift we'll see in software versus

0:37:58
Terence Tao

research. Right. Yeah. So certainly in math, the process is often more important than the problem itself um the problem is kind of a proxy for for measuring the progress and i think even in software there's there's different types of software tasks i mean that you know like if you just kind of create a web page that does the same thing that a thousand other web pages do um there's there's sort of no skill to be learned well um there's there's still some skill maybe that the individual programmer could pick up um but you know for for kind of a boilerplate type code definitely um um you know, it's something that you should definitely offload to AI.

0:38:34

But, you know, sometimes once you make the code, you know, you still have to maintain it and there's issues with upgrading it and making it compatible with other things. And that, I think, I've heard that programmers are reporting, you know, that even if an AI can create the first prototype of a tool, making it mesh with everything else and making it interact with the real world in the way they want, I mean, that's an ongoing process.

0:38:57

And if you didn't have the skills that you pick up from writing the code, that may impact your ability to maintain it down the road. So certainly mathematicians, we've used problems to build intuition and to train people to have a good idea as what's true, what to expect, what is provable, what is difficult. and so just getting the answers right away may actually inhibit that process. I mean, so I made a distinction between theory and experiment before.

0:39:37

So in most sciences, there's an equal division between there's a theoretical side and experimental side. But math has been almost unique. It's almost entirely theoretical. We pay a premium on sort of trying to have coherent, clean theories of why things are true and false. And we haven't done much experiments as to, like, you know, maybe we have two different ways to solve a problem. Which one is more effective? We have some intuition, but we haven't done large-scale studies where we take a thousand problems and we just test them.

0:40:10

But we can do that now. So I think AI-type tools, we really will actually revolutionize the experimental side of math, where you don't care so much about individual problems and the process of solving them, but you want to gather just large-scale data about what things work, what things don't. Same way that if you're a software company and you want to roll out a thousand pieces of software, you don't really want to handcraft each one and learn lessons from each.

0:40:39

You just want to find what are the workflows that let you scale. So we don't yet... The idea of doing mathematics at scale is at its infancy, but that's where AI is really going to revolutionize the subject.

0:40:52
Dwarkesh Patel

Interesting. I feel like a big crux in these conversations about how good AI will be for science is, I think you said this, it's like, oh, they're using existing techniques and modifying them. And it would be interesting to understand how much progress one can make simply from using existing techniques. Like how much of, if I looked at the top math journals, how many of them are, how many of the papers are coming up with, whatever coming up with a new technique means, doing that versus using existing techniques in new problems.

0:41:23

And what the overhang is, where if you just applied every known technique to every open problem, would that just constitute a humongous uplift in our civilization's knowledge, or would that not be that impressive and useful?

0:41:34
Terence Tao

This is a great question. We don't have the data to fully answer it yet. Certainly, a lot of work that human mathematicians do, you know when you when you take a new problem one of the first things we do is we just find we we look at all the standard things that have worked on similar problems in the past and we try them one by one um and sometimes that works um and that's still worth publishing sometimes because the question was important um sometimes they almost work and you have to add one more wrinkle um to it and that's also um interesting um but then you know the papers that go into the top journals are usually ones where you

0:42:10

you know the existing methods can kind of solve you know 80 % of the problem, but then this is 20 % which is resistant, and a new technique has to be invented to fill in the gaps. It's very, very rare now that a problem gets solved with sort of no reliance on past literature, where all the ideas come out of nowhere. That was more common in the past, but math is so mature now that it's just so much of a handicap to not use literature first um so yeah ai tools are really good at getting really good at the first part of that just trying all the standard techniques on a problem um often now actually making fewer mistakes in

0:42:53

implementing them than than humans they still make mistakes but but um um i've i've tested these tools you know on on on um like little tasks that i can do and and sometimes they pick up errors that i make sometimes i pick up errors that they make it's it's about a tie right out uh for um but um yeah i i haven't yet seen them take the next step you know so so when there are holes in in in the argument where none of the things are working to to how then what do you do um and then they can kind of suggest random things and it but it it um often i find that trying to chase them down

0:43:31

and make them work and finding they don't work it wastes more time than it saves So I think some fraction of problems that we currently think are hard will fall from this method. I mean, especially the ones that haven't received enough attention. So with the Erdős problems, almost all of the 50 problems that were solved by AIs were ones for which basically there was no literature. I mean, Erdős posed the problem once or twice. I think maybe some people tried it casually and they couldn't do it, but they never wrote up anything.

0:44:03

um but it turned out that there was a solution and it was just you know maybe combining this one obscure technique that that not many people know about with some other result in the literature and that's the kind of love the median level of what ai can accomplish and that's really great it clears out 50 of these problems um so i think you'll see some isolated successes um but the six but what we found so people have to have done large scale sweeps of these early problems and If you only focus on the success stories, the ones that get broadcast on social media, it looks amazing.

0:44:34

All these problems that haven't been solved before for decades, now they're falling. But whenever we do a systematic study, any given problem, an AI tool has a success rate of maybe 1 % or 2%. It's just that they can buy a scale, and if you just pick the winners, it looks great. So I think there'll be a similar thing happening with, there are hundreds of really prestigious, difficult math problems out there a couple may make you know some AMA get lucky and I keep solve them and there was there was some some backdoor to solve the problem that everyone else missed and that will get a lot of publicity but then people will try these fancy tools on their own

0:45:13

favorite problem and they will again experience the one to 2% success rate right so there'll be a lot of noise and amongst the signal of sort of when they're working when they're not. We have to do, it's increasingly important to collect these really standardized data sets. You know, there are efforts now to create a standard set of challenge problems for AI to solve and not just rely on the AI companies to only publish their wins and not disclose their negative results.

0:45:43

So that will maybe give more clarity as to where we're actually at.

0:45:48
Dwarkesh Patel

Well, I think it's worth emphasizing how much progress in AI constitutes already to have models that are capable of applying some technique that nobody had written down as applicable to this particular problem.

0:45:59
Terence Tao

The progress is simultaneously amazing and disappointing. It is a very strange feeling to see these tools in action. But also people acclimatize really quickly. I remember when Google's web search came out 20 years ago and it just blew all the searches out of the water. You're just getting relevant hits on the front page perfectly, almost exactly what you wanted. And it was amazing. And then after a few years, you just took for granted that you could just Google anything.

0:46:29

And yeah, so a lot of, yeah, I mean, 2026 level AI would be stunning in 2021. And a lot of it, you know, face recognition, natural speech, doing, you know, college level math problems, we just take for granted now. Yeah.

0:46:43
Dwarkesh Patel

Okay. So speaking of 2026 AI, you made a prediction in 2023 that I think by 2026, what was it? It would be like a colleague in mathematics or?

0:46:52
Terence Tao

Yeah, a trustworthy co-author, if used correctly.

0:46:55
Dwarkesh Patel

Which is looking pretty good in retrospect. Yeah, I'm pretty pleased. So, you know, let's see if we can continue the streak. You personally are 2x more productive as a result of AI. What year would you say that?

0:47:10
Terence Tao

Yeah, so productivity, I think, is not quite a one-dimensional quantity. I'm definitely noticing that the style in which I do mathematics is changing quite a bit. and the type of things I do. So, for example, my papers now have a lot more code, a lot more pictures, because it's so easy to generate these things now. So some plot, which have taken me hours to do now, I can do in minutes. But in the past, I just wouldn't have put the plot in my paper in the first place.

0:47:38

I would just talk about it in words. So it's hard to measure what 2x means. So, yeah, on the one hand, I think the type of papers that I would write today, if I had to do them without AI assistance, they would definitely take five times longer. Interesting. But I would not write my papers that way. 5X? Yeah, but it's because these are sort of auxiliary types. I mean, so things like doing a much deeper literature search, supplying a lot more numerics, I mean, they enrich the paper.

0:48:17

so yeah the core of what I do like actually solving the most difficult part of a math problem that hasn't changed too much I still use pen and paper for that but you know there's lots of silly things I use an AI agent now to reformat like sometimes all my parentheses are not quite the right size you know I usually manually change in my hand and I can get an AI agent to sort of do all that quite nicely now in the background um so yeah they've really sped up lots of secondary tasks uh they haven't yet sort of uh um sped up the core thing that i do but it's allowed me to sort of add more things to

0:48:56

to to my papers um yeah but um by the same token like if i were to write a paper i wrote in 2020 again and not add all these extra features but just have something of the same sort of level of functionality, then that actually doesn't save that much, to be honest. Yeah, so it's made the papers sort of richer and broader, but not necessarily deeper.

0:49:19
Dwarkesh Patel

You made this distinction between artificial cleverness and artificial intelligence, and I would like to better understand those concepts. What is an example of intelligence that is not just cleverness?

0:49:35
Terence Tao

Yeah, so intelligence is famously hard to define. It's one of these things that you kind of know it when you see it. But when I talk to someone and we're trying to collaboratively solve a math problem together, there's this conversation where, you know, neither of us knows how to solve the problem initially, but one of us has some idea and it looks promising. And so then we have some sort of prototype strategy and then we test it and then it doesn't work, but then we modify it.

0:50:10

And there's some adaptivity and continual improvement of the idea over time. And eventually, you know, we've systematically mapped out what doesn't work, what does work, and we can kind of see a path forward, but it's evolving with our discussion. And this is not quite what the AI is. The ads can kind of mimic this a little bit. So to go back to this analogy of these jumping robots, so they can jump and fail and jump and fail and jump and fail, but what they can't do is they kind of jump a little bit and they reach some handhold, but then they sort of stay there and then they pull other people up and then they try to jump from there.

0:50:55

There isn't this cumulative process which is sort of built up interactively. it seems to be a lot more trial and error and just repetition brute force, which it scales and it can work amazingly well in certain contexts. But yeah, this idea is sort of building up cumulatively from partial progress is kind of what's still not quite there yet.

0:51:23
Dwarkesh Patel

Interesting. You were saying if Gemini 3 or Cloud 4.5, whatever, solves a problem, it is not the case that its own understanding of math has progressed. Or even if it works on a problem without solving it, it's not that its own understanding of math has progressed.

0:51:36
Terence Tao

Yeah, you run a new session and it's forgotten what it just did. It has no new skills to attach to, to build on related problems. Maybe what you just did is part of 0.01 % of the training data for the next generation. So maybe eventually some of it gets absorbed.

0:51:54
Dwarkesh Patel

So Terence talks about the importance of decomposing, particularly gnarly problems into a series of easier chunks. Even if this doesn't result in the full solution, approaching problems in this way helps you build up the intuitions and practice the techniques that you'll need to keep making progress. But models today tend to struggle with these kinds of problem-solving techniques. That's where LabelBox comes in. LabelBox helps you train models not just to get the right answer, but to think the right way.

0:52:19

They've operationalized these reasoning behaviors into rubrics, giving you the ability to evaluate every important dimension of a model's output. These rubrics go beyond simple correctness. Did the model reach for the right tools? Did it check its own work and explore alternative paths? How clear was its response? These skills are useful across domains. Math, physics, finance, psychology, and more. And they're becoming increasingly important as models take on harder, open-ended problems, some of which have multiple solutions and some of which we don't even know the solutions to.

0:52:48

LabelBox can get your rubrics tailored to your domain, helping you systematically measure and shape how your models think. Learn more at labelbox.com/dwarkesh. like totally doing it in Lean, that the constructions which are made, the definitions which are created, even in the Lean program, have to advance our understanding of mathematics? Or do you think it could just be assembly code googly gook?

0:53:38
Terence Tao

Oh, yeah, we don't know. I mean, some problems have been basically solved by pure brute force. A full-color theorem is a famous example. We have still not found a conceptually elegant proof of this theorem. It basically, and maybe we never will. I mean, some problems may only be solvable by just splitting into some enormous number of cases and doing a brute force, uninsightable computer analysis on each case. I mean, part of the reason we prize problems like the real hypothesis is that we're pretty sure that something amazing has to, a new type of mathematics has to be created, or a new connection between two previously unconnected areas of mathematics has to be discovered to make this work.

0:54:21

Like we don't even know what the shape of the solution is, but it doesn't feel like a problem that will be solved just by exhaustively checking cases or something. I mean, it could be false, actually. We could actually, okay, there is an unlikely scenario that the hypothesis is false and there's just this, you can just compute, oh, here's a zero off the line and a massive computer calculation verifies it. That would be very disappointing. I don't know.

0:54:49

I do feel that fully autonomous one-shot approaches are not the right approach for these problems. I mean, I think you'll get a lot more mileage and the interplay between humans collaborating with these tools. And I can see one of these problems being solved by some smart humans assisted by some extremely powerful AI tools. but the exact dynamic may be very different from what we envision right now. I mean, it could be a collaboration of a type that just doesn't exist yet.

0:55:29

Yeah, I mean, there may be a way to generate a million variants of the human's data function and do some data analysis, AI-assisted data analysis, and we discover some pattern between connecting them, which we didn't know about before, and this lets you transform the problem into a different area of mathematics. I mean, there could be all kinds of scenarios. So suppose the AI figures it out

0:55:53
Dwarkesh Patel

and latent in the lean is some brand new construction, which, you know, if you realize the significance, we would be able to apply it in all these different situations. How do you even recognize it, right? Like if you just, again, a very naive question, but if you come up with the equivalent of like Descartes comes with this idea, oh, you can have this coordinate system where you can unify algebra and geometry. But in Lean code, it would just look like R to R and it wouldn't look that significant or something.

0:56:21

Or similarly, I'm sure there's other constructions which have this kind of property.

0:56:25
Terence Tao

Well, the beauty of formalizing a proof in something like Lean is that you can take any piece of it and study it atomically. So when I read a paper with my humans, which shows some difficult problem, you know there's often some big sequence of lemmas and theorems and things and so ideally the author will talk it talk their way through you know what's important what's not but sometimes they don't reveal what what what steps were the important ones and which ones are just kind of boilerplate standard steps but you can study each lemma in isolation and some of them I can say oh this looks fairly standard this this this resembles something I'm familiar with I'm pretty sure there's nothing

0:57:05

interesting going on here but this lemma oh that's that's something I haven't seen before and i could see why if you could if you had this result that would really help prove the main result like you could you know you can assess whether some things are are really sort of key to your um uh to your argument or not and lean really facilitates that you know you can you can you can you know the individual steps are really precisely um i think in the future there'll be um you know there'll be entire professions of of mathematicians who I take a giant lean generated proof and maybe do some ablation on it or something.

0:57:39

I try to remove parts of it and try to find more elegant ways, maybe some other AIs to sort of do some reinforcement learning. How can you make the proof more elegant and maybe other AIs will grade whether this proof looks better or not. One thing that will change quite a bit in the near future is that until recently, writing papers was the most time-consuming and expensive part of the job. And so you did it very rarely.

0:58:07

You only wrote up your results once everything was, all the other parts of your argument were checked out and things, because just rewriting it again, refactoring was a total pain. But that's one thing that's become a lot easier now with modern AI tools. So you don't have to have just one version of your paper. Once you have one, people can generate hundreds more. So yeah, one giant messy lean proof may not be very meaningful or understand it on its own, but other people can refactor it and do all kinds of things with it.

0:58:39

We have seen with the Erdős problem website, an AI will generate a proof and then here's 3 ,000 lines of code that verify the proof, but then people got other AIs to summarize the proof and people write their own proofs. There's actually post-processing, once you actually have one proof, we actually have a lot of tools now to deconstruct it and interpret it. It's a very nascent area of science or mathematics, but I'm not as worried about, so some people are concerned, what if the real hypothesis is proven with a completely incomprehensible proof?

0:59:14

I think once you have the artifact of a proof, we can do a lot of analysis on it.

0:59:21
Dwarkesh Patel

You posted recently that it would be helpful to have a formal or semi-formal language for mathematical strategies as opposed to just mathematical proofs, which is what Lean specializes in. I would love to learn more about what that would involve or look like.

0:59:34
Terence Tao

We don't really know. I mean, we've been very lucky in mathematics that we have worked out the laws of logic and mathematics. But this is actually a fairly recent accomplishment. I mean, it was started by Euclid, you know, millennia ago. But only in like the early 20th century did we finally list our, here are the axioms of mathematics, well, the standard axioms of what we call ZFC and the axioms of first order logic. And this is what a proof is.

0:59:59

and this we've managed to automate and have formal language for. But there could be some way to assess plausibility of certain, you know, so you have a conjecture that something is true, you test a few examples and it works out, like how does this increase your confidence that the conjecture is true? We have a few sort of mathematical ways to model this, Bayesian probability, for example, but they're not but you often have to set certain base assumptions and there's a lot of subjectivity still in these tasks so it is it's not clear I mean this is more of a wish than a plan to develop these languages but just seeing how successful having a formal framework in place like lean

1:00:54

has made deductive proofs so much easier to automate and train AI on. If there was some similar framework, so the bottleneck for using AI to create strategies and make conjectures is we have to rely on human experts and the test of time to validate whether something's plausible or not. If there was some semi-formal framework where this could be done semi-automatically in a way that that uh um isn't sort of easily hackable um to you know so of course yeah it's really important with these formal proof assistants that that that there are just no um uh there's no back doors or exploits that um that you can do to somehow get your your certified proof without actually proving it

1:01:42

because reinforcement learning is just so so good at finding these these back doors um but um yeah if it's not very good sort of mimics how scientists talk to each other in a semi-formal way you know using data and an argument but but also um you know constructing narratives and and there's some such there's some subjective aspect of science that we don't know how to capture in a way that that that uh we can insert ai into them in a useful way.

1:02:16

So yeah, this is a future problem. There are research efforts to try to create automated conjectures and maybe there are ways to benchmark these and get some way to simulate this. But it's all very, very new science. Can you help me get

1:02:36
Dwarkesh Patel

some intuition for... I have two step questions. One, it would be very helpful to have a tangible sense of it'd be helpful to have a specific example of what something like this would look like that the way scientists communicate that we can't formalize yet and two it seems almost definitionally paradoxical to say building up some narrative or building up some natural language explanation and then also having something which you could have formalized And I'm sure there's some intuition behind where that overlap is, and I'd love to understand that better.

1:03:19
Terence Tao

So an example of conjecture. So Gauss was interested in the prime numbers, and he computed, he created one of the first mathematical data sets. He just computed the first 100 ,000 prime numbers or so, hoping to find patterns. And he did find a pattern, but maybe not the pattern he was expecting. he found a statistical pattern in the primes that if you count how many primes there are up to 100, 1 ,000, 1 million and so forth, they get sparser and sparser but the drop-off in the density was inversely proportional to the natural logarithm of the range of numbers.

1:03:57

So he conjectured what we now call the prime number theorem. The number of primes up to x is like x divided by the natural log of x. And he had no way to prove this. It was data-driven. So this was a conjecture. It was revolutionary for its time because it was maybe the first really important conjecture of math that was statistical in nature. So normally you talk about patterns like maybe the spacing between the primes has a certain regularity or something.

1:04:26

But this was really something which it didn't tell you exactly how many primes there were in any given range. it just gave you an approximation that got better and better as you went further and further out. But it helped. So it started the field of what we call analytic number theory. But it was the first in many conjectures like this, many of which got proved, which sort of started consolidating the idea that the prime numbers actually didn't really have a pattern, that they behaved like random sets of numbers with a certain density.

1:05:02

I mean, they had some patterns, like they're almost all odd. Okay, so there was something, and they're not actually random. They're what's called pseudo-random. I mean, there's no random number generation involved in creating the prime numbers. But 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 just creating this random set. And this allowed us to make all these other predictions.

1:05:28

So there's a still open conjecture in number theory called the twin prime conjecture, that there should be infinitely many pairs of primes that are twins, distance to a path, like 11 and 13. We can't prove that, and there's actually good reasons why we can't prove it. But because of this statistical random model of the primes, we are absolutely convinced it's true. We know that if the primes were sort of generated by flipping coins or something, that we would, just by random charge, just like infinite monkeys at a typewriter, we would see twin primes appear over and over again.

1:05:56

and we have over time developed this very accurate conceptual model of what the primes should behave like based on statistics and probability but it's all mostly heuristic and non-rigorous but extremely accurate so the few times when we actually can prove things about the primes it has matched up with the predictions of this what we call the random model of the primes so we we have this conjectural concept framework for understanding the primes that we everyone believes in and you know it's the same reason why we believe the real hypothesis is true why we believe that cryptography based on the primes is basically um is mathematically secure things like that it's it's all part of this this belief

1:06:37

um in fact one reason why we care about the reuben hypothesis is that if the real hypothesis failed um we knew it was false it means that it would be a serious blow to this model. That it would mean there's a secret patent to the primes that we were not aware of. And I think we would very rapidly abandon any cryptography based on the primes, because if there was one patent that we didn't know about, there's probably more.

1:07:00

And these patents can lead to exploits in crypto. And yeah, it's going to be a big, big shock. So we really want to make sure that doesn't happen. so yeah it's it's um so we've been convinced of of things like agreement hypothesis and things over time but uh some of it is experimental evidence some is uh the few times we've been able to make theoretical results they've always aligned um you know it is possible that the consensus is wrong and and we've all just missed something very basic um you know there have been paradigm shifts in the past in scientific history um yeah but we don't really have a way of measuring this um i think partly because we don't

1:07:43

have enough data on on on how math or science develops we have one timeline of history and you know we have like you know a hundred stories of turning points in history if if we had access to a million alien civilizations and each of the different development of history and of science in different orders then maybe we'd actually have a have a have a decent shot at an understanding of how do we measure what is progress and what is a good strategy and we could maybe start formalizing it and actually having a framework.

1:08:17

Maybe if what we need to do is actually start creating lots of mini universes or simulations of AI solving very basic problems you know in arithmetic or whatever but coming over their own strategies for doing these things and having these little laboratories to test. I mean, there are people who investigate, like what's the smallest neural network that can do 10-digit multiplication and things like that. I think we could actually learn a lot just from evolving small AIs on simple problems.

1:08:47

We could learn a lot.

1:08:48
Dwarkesh Patel

I was super excited when Mercury reached out about sponsoring the podcast because I've been banking with them for years. I think I opened my first account with them in 2023. Something I've come to appreciate over the last few years is that Mercury is constantly updating things and adding new features. Take their newest feature, Insights. Insights summarizes your money in and out, showing you your biggest transactions and calling out anything that deserves extra attention. Like maybe your revenue from a particular partner has gone down, or you've got a big uncategorized purchase that needs to be investigated.

1:09:15

It's a super low friction way for me to keep tabs on my business and make quick decisions. For example, I tried to invest any cash that I don't need on hand to keep running the business. With Insights, with just a couple of clicks, I was able to see exactly how much money I spent in each month of 2025. And that lets me know exactly how much cash I'll need for the next year or so of operations.

1:09:33

And then I can go invest the rest. Mercury just keeps adding new features like this. Go to mercury.com to check it out. Mercury is a fintech company, not an FDIC insured bank. Banking services provided through Choice Financial Group and Column N.A., members FDIC. You have to learn about new fields, not only very rapidly, but deeply enough to contribute to the frontier. So in some sense, you're also one of the world's greatest autodidacts. What is your process of learning about a new subfield in math?

1:10:05

What does that look like?

1:10:06
Terence Tao

Yeah, so I certainly identify with kind of the, yeah, we talked about depth and breadth before. And it's not purely human AI distinction. I mean, humans also split. So I think it was Irving who split them into hedgehogs and foxes. and a hedgehog knows one thing very very well and a fox knows a little bit about everything so i definitely i didn't you know i i think of myself as a fox um you know i i work with hedgehogs a lot and sometimes i can be a hedgehog if need be but um yeah so um i've always had a little bit of an uh obsessive streak if there's something which i read about which i feel like i should understand

1:10:46

i i have the capability to understand this but i don't understand why it works there's some magic in it that um you know so someone was able to use a type of mathematics i'm not familiar with and get that which i would like to prove and i can't do it by myself but they could do it by their method then i wanted to find out what was their trick um it bugs me that they someone else can can do something which i think i can do but but i can't um so i've always had that kind of obsessive completionist type type streak um i've had to wean myself off computer games because I start a game, I want to play

1:11:18

it to completion, so both the levels. So that's one way in which I learn new fields. I collaborate with a lot of people who have taught me other types of mathematics. I just make friends with another mathematician who's working on another area of mathematics and I find their problems interesting, but they have to teach me some of the basic tricks and what's known, what's not known. and I learned a lot from that. I found that writing about what I've learned, I have a blog where I sometimes record things that I've learned because in the past, when I was younger, I would learn something and do this cold trick and say, okay, I'm going to remember this.

1:12:02

And then six months later, I've forgotten. I remember remembering it, but I can't reconstruct my arguments. And the first few times it was so frustrating to have understood something and then lost it. I sort of resolved I should always write down anything cool that I've learned. And this is part of how this blog came about. How long does it take you to write a blog post? It's something I often do when I don't want to do other work.

1:12:28

You know, like there's some referee report or something. There's something that feels slightly unpleasant for me to do at the time. And so writing a blog, it feels creative and fun. Like it's something that I do for myself. So maybe depending on the topic, it could be a quick, you know, half an hour or several hours. But it doesn't, because it's something that I do sort of voluntarily, it doesn't feel like it, it doesn't feel, time flies when I write these things.

1:12:57

As opposed to sort of doing something which I have to do for administrative reasons, but it's just that it's drudgery. Those are tasks that AI is really helping with nowadays, actually.

1:13:06
Dwarkesh Patel

Is it, if like civilization could from first principles decide how to use Terry Tau's time? You know, it's like a limited resource. What is the biggest diff between if the veil of ignorance got to decide how to use Terry Tau's time versus what it does now? This podcast wouldn't be happening.

1:13:28
Terence Tao

Yeah, as much as I complain about certain tasks that I don't want to do, but I have to do. So as you get more senior in academia, you get more and more responsibilities, and I get some more committees and whatever. But I have also found that a lot of events that I kind of reluctantly went to because I was obliged to for one reason or another, because it's outside my comfort zone, I often find interactions with people who I wouldn't normally talk to, like you, for instance.

1:13:56

And I would learn interesting things and have interesting experiences, and I would have opportunities to then network with other people that I would never have done before. So I do believe a lot in semantivity. I mean, I do optimize my time when I... So there's some portions of a day where I do schedule very carefully. But I have been willing to sort of leave some portions just, okay, I'm going to do something which is not my usual thing and maybe it'll be a waste of my time, but maybe I will learn something.

1:14:32

And more often than not, I feel like I've gotten a positive experience, which is not something I would have planned for. And yeah, so I believe a lot in serendipity. And maybe there's a danger actually that, you know, in modern societies, it's not just AI, but we've become really good at optimizing everything. And maybe we are optimizing, we're not optimizing a lot of optimization. that you know with with with covert for example um we we switched like we switched a lot to remote meetings um and so everything was scheduled now and so uh we kept busy at least in academia you know we met almost the same number of people that we met in person but everything had to be

1:15:15

planned um you had to schedule things in advance and what we lost out on was sort of the the casual like knocking on the hallway, just meeting someone while getting a coffee. And there's serendipitous interactions that you may think are not optimal, but actually are really important. When I was a grad student, I would go down to the library to look for a journal article. I had to physically go down to the library, check out the journal, and read your article.

1:15:46

And sometimes the next article, you can just browse through, and the next article is also interesting. sometimes it wasn't but but you could accidentally find interesting things um which is something which has basically been lost now because you can just type in you know if you if you want to access an article now you just type it into to a search engine or even an ai and you can get instantly what you want but you don't get sort of the accidental things that you you might have have gotten if you'd done it more inefficiently um so um yeah there have been times when i'm um I spent a year once at the Institute for Advanced Study, which is a great

1:16:24

place to, you know, there's no distractions. You're there to just do research. And like the first few weeks you're there, like it's great. You're getting all these papers written up that you've been wanting to do for a long time. You've been thinking about problems for blocks of hours of a time. But I find out if I'd stay there for more than several months, like I run out of inspiration somehow. Like I get bored. I actually, you know, surf the internet a lot more.

1:16:46

You actually do need a certain level of distraction in your life, it somehow adds enough randomness and temperature, high temperature if you need. So yeah, I don't know the optimal way to schedule my life. It just seems to work.

1:17:04
Dwarkesh Patel

I'm very curious when you expect AIs that can actually do frontier math at least as well as the best human mathematicians.

1:17:15
Terence Tao

I mean, in some ways, they're already doing frontier math that is super intelligent that humans can't do, but it's a different frontier from what we're used to. I mean, you could argue that calculus were doing frontier math that humans could not accomplish, but it wasn't, you know, number crunching. But replacing Terry Tao completely. I mean, what do you want me for?

1:17:42
Dwarkesh Patel

You'll just go on all the podcasts after.

1:17:47
Terence Tao

I'm not sure we've it might not be the right question to ask I think with within a decade a lot of things that mathematicians currently do where we spend a lot of the bulk of our time doing it and a lot of stuff we put in our papers today can be done by AI but we will find that that actually wasn't the most important part of what we do. You know, 100 years ago, a lot of mathematicians were just solving differential equations.

1:18:22

People needed, physicists needed some exact solution to some system, and they were just, they hired a mathematician to go through the calculus and work out the solution to this fluid equation or whatever. A lot of what a 19th century mathematician would do, you could make a call to Mathematica or Wolfram Alpha or a computer algebra package, or now more recently an AI, and it would just solve the problem in a few minutes. But we moved on.

1:18:52

We worked on different types of problems after that. Once computers came along, computers used to be human. People used to laboriously create log tables and work out primes, as Gauss did, and that has all been our source to computers. But we moved on. So in genetics, to sequence the genome of a single organism, that was an entire PhD of a geneticist. So carefully separating all the chromosomes and whatever. And now you can just spend $1 ,000 and send it to a sequencer and get it done.

1:19:26

But genetics is not dead as a subject. You move to a different scale. Maybe you study whole ecosystems rather than individuals.

1:19:33
Dwarkesh Patel

I take your point. But on the question of, well, when is most mathematical progress, almost all mathematical progress happening by AI? So if you find out, oh, this year, a millennium price problem has been solved, you would put, you know, a 95 % odds that an AI did it autonomously. Surely there will be such a year.

1:19:53
Terence Tao

I guess. I mean, I do believe that hybrid human plus AIs will dominate mathematics for a lot longer. Yeah, it will depend, it will require some additional breakthroughs beyond what we already have. So it's going to be stochastic. You know, I think, you know, AI is currently very good at certain things, but really terrible at others. And while you can sort of add more and more frameworks on top to kind of reduce the error rates and make them work with each other a bit more and so forth, it feels like we don't have all the ingredients to really have a truly satisfactory replacement for all intellectual tasks.

1:20:37

It is complementary currently. It is not a replacement. But maybe, I mean, because current-level AIs will accelerate science in so many ways, hopefully new discoveries, new breakthroughs will happen more quickly. I mean, it's possible that also by somehow destroying serendipity, we actually inhibit certain types of progress. Anything is possible really at this point. I think the world is very, very unpredictable

1:21:11
Dwarkesh Patel

at this point in time. What is your advice to somebody who would consider a career in math or is early in a career in math, especially in light of AI progress? How should they be thinking about the career differently, if at all, as a result of AI progress? Yeah, so we live in a time of change.

1:21:31
Terence Tao

It is, as I said, we live in a particularly unpredictable era. And I think things that we've taken for granted for centuries may not hold anymore um so um yeah the way we uh do everything and not just mathematics um will change and um you know so i i think which is you know i mean in many ways i would prefer the much more boring quiet era where things are much the same as they were 10 years ago 20 years ago but um so i think one just has to embrace this that that there's going to be a lot of change and that the things that you study, some of them may become obsolete or revolutionized, but some things will

1:22:22

be retained. And so you somehow always have to keep an eye on... There'll be a lot of opportunities for things that you wouldn't be able to do before. So in math, you previously had to basically go through years and years of education, 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 and lean and everything else.

1:22:56

So there'll be a lot of non-traditional opportunities to learn. So you need a very adaptable mindset. There'll be one for pursuing things just for curiosity, for playing around. And I mean, you still need to get your credentials. I mean, I think for a while it would still be important to sort of still go through traditional education and learn math and science and so forth the old-fashioned way for a while. But you should also be open to very, very different ways of doing science, some of which don't exist yet.

1:23:36

Yeah, so it's a scary time, but also very exciting.

1:23:40
Dwarkesh Patel

Awesome. That's a great note to close on. Thanks so much. Yeah, pleasure.