Grant Sanderson – AI and the future of math

1. Key Themes

AI Progress in Math Is Fractal, Not Uniform

AI's advancement in mathematics isn't monolithic — it has a spiky, fractal nature where some subfields are essentially solved while others remain intractable. Geometry problems at the IMO were effectively brute-forced, while combinatorics remained elusive.

"They're very good. They just like cold solved geometry, basically... Combinatorics is the one that's the wild card of much more like playful, puzzly seeming problems." — Grant Sanderson 00:02:05

The Next Frontier: Conjecture Generation, Not Theorem Proving

The field's real unsolved challenge for AI is moving from proving known problems to generating interesting new conjectures and definitions. This is inherently non-benchmarkable.

"How good mathematicians prove theorems, great mathematicians come up with conjectures, and the greatest mathematicians come up with definitions." — Grant Sanderson 00:09:28

"The way that you'd measure conjecture generating ability is going to be more subjective on like that tone shift where it'll be mathematicians saying they're not just using it to like solve their problems. But as they step back and decide what their research field should even be, that a conversation with such and such model like was genuinely helpful for that." — Grant Sanderson 00:11:16

Grindability, Not Just Verifiability, Drives AI Math Progress

The common explanation for AI's fast math progress — verifiability — is incomplete. The deeper driver is "grindability": the ability to run massive parallel rollouts in a containerized, deterministic environment.

"Math, of course, is the exception. And I feel like this is actually an important driver of progress in this domain and also in coding. It's not just verifiability. It has to be grindable." — Dwarkesh Patel 00:56:01

"With code also, you can containerize a given level of progress in a repository and then just spin out thousands of parallel containers or hundreds of parallel containers and say like, try to implement this feature. And it's totally deterministic. And because it's deterministic, you can solve the credit assignment problem." — Dwarkesh Patel 00:55:39

Parallelization as AI's Most Underrated Advantage

The ability to run billions of simultaneous AI instances — each with the same knowledge base — is a qualitatively different kind of intelligence multiplier than raw model capability.

"It's just like one idiosyncratic genius in the history of mathematics who makes a few connections and then dies in a duel. Right. It's just universally applying the waterline across all problems that are accessible at the level of capability. I feel like this is among the many advantages that digital minds inherently have that we don't think enough about." — Dwarkesh Patel 00:46:13

AI as a "Lightning Bolt" Connector Between Fields

The most near-term productive use of AI in mathematics will be drawing connections between fields that humans miss due to siloed expertise — analogous to the famous Hugh Montgomery / Freeman Dyson lunch conversation connecting number theory and random matrix theory.

"They're an expert at the quantum physics, they're an expert at the analytic number theory. They should be able to see that similarity in a way that doesn't require like Montgomery and Dyson to be having lunch and like happening to talk about that." — Grant Sanderson 00:05:16

"My guess on what most of the useful progress from these models will look like. Like in the next five years is just really filling in that landscape of like connections that you can draw if you're an expert in multiple fields." — Grant Sanderson 00:41:18

The 100-Year Verification Loop Problem

Some of the most valuable mathematical insights in history — like Galois's group theory — weren't recognized as valuable for a century. This long verification loop is fundamentally incompatible with current RLVR training paradigms, and poses a deep challenge for training AI to generate truly novel mathematical structures.

"Between the time of Lagrange, like having this inkling of maybe symmetries of roots is the right way to go to where it all looks like modern group theory. Like you've got this long span. A lot of the time, it's like not even passing the like verified reward of human reviewers." — Grant Sanderson 00:21:48

"He anticipates quarks based on a purely group theoretic question. And like that's one of the more interesting applications of group theory is that like to even predict the existence of quarks is a group theoretic like question. And that's so long after Lagrange before you have anything like that." — Grant Sanderson 00:22:07

Systematically Increasing Entropy May Be AI's Key Research Advantage

Rather than the entropy-collapse concern (all AIs thinking the same), deliberate prompt-level entropy injection — giving different agents different biases, having one try to prove and one to disprove — could be AI's unique structural advantage over human research institutions.

"Systematically increasing entropy at the prompt level, even though you have this like inevitable collapse at the like auto regression level." — Grant Sanderson 00:51:36

"It sounded like one of the reasons the unit distance problem conjecture took so long to be disproven, which because people assumed the conjecture was actually true. So mostly they were trying to figure out ways in which to prove it." — Dwarkesh Patel 00:50:54

Lean's Real Value Is "Press Go and Walk Away" Math Exploration

Lean's importance is not primarily as a verifier for current progress, but as an infrastructure that allows completely automated, indefinitely running mathematical exploration — analogous to AlphaZero playing Go without human supervision.

"You could imagine having a basically endlessly running program that's constantly trying to extend Mathlib... It might come up with its own conjectures. It might come up with its own theories and like different definitions. Maybe many of them are useless, but it just has this infinite tree that it can like grow out. That's a very unique thing that math has that nothing else has where you could press go and then just like poor compute at it and like look away for 10 years." — Grant Sanderson 00:58:43

Human Mathematicians Will Become Curators, Not Solvers

The lasting role of human mathematicians — and educators like Grant — will be analogous to art museum curators: navigating an AI-generated near-infinite landscape of ideas to surface what is worth engaging with.

"One interesting take that I've heard about like what mathematicians will end up being is actually more analogous to art museum curators than anything else. Where the AI is all the things, so the art exists, right? They even know how to explain it really well, you know, all there. But like you still, you still want someone to help you navigate in this like nearly infinite space of like what ideas are worth engaging with." — Grant Sanderson 00:36:46

---

2. Contrarian Perspectives

Solving a Millennium Prize Problem Would NOT Signal General AGI Capability — Unless It Requires "Mountain Building"

Most assume that solving the Riemann hypothesis would imply near-general AI. Grant argues this depends entirely on the character of the solution. A cross-field lightning bolt is different from building a wholly new mathematical framework, and only the latter would imply the kind of intelligence that would obviously generalize.

"If the nature of the solution to the Riemann hypothesis was something like that, that feels pretty distinct to me than what's necessary to get good at white collar work... But that like if it's capable of building mountains that are, you know, the correct new theory that like crystallizes how we should be thinking about a subject. That's just such a level of intelligence that then it starts to feel like it would be surprising if that didn't permeate into other aspects of the economy." — Grant Sanderson 00:05:16 / 00:07:03

Lean Is Overrated as a Training Signal for Current AI Math Progress

The consensus view in AI math circles is that formal verification via Lean is central to the progress story. Grant and Dwarkesh argue it is largely not the reason for recent breakthroughs.

"The initial attempts... it's like initially DeepMind basically does that. It's like everything in Lean. And then the next year it's all in natural language. So to your point, not needed." — Grant Sanderson 00:57:29

"Lean just doesn't matter that much for like the current level of progress in AI... It just seems like less relevant than just having this grindable outcome that is verifiable." — Dwarkesh Patel 00:56:30

AI Will Also Be Better Than Humans at Explanation — Not Just Proof

The comforting narrative for mathematicians is that even if AI proves things, humans will remain valuable as explainers and educators. Grant has changed his mind on this.

"I kind of used to think that AIs will become these automated theorem provers, but like the role of the mathematicians is going to shift towards like my job, like explain these things. I kind of suspect that actually they'll also be like quite good at doing that and probably just like better than most humans are at like doing the explanation half and distilling half." — Grant Sanderson 00:35:05

AI's Weak Theory of Mind Is Structurally Inevitable, Not a Bug to Be Fixed

People assume AI's poor mentalizing is a training or scale issue. Grant argues it is structurally expected: understanding others' emotions requires embodied mimicry (face muscle feedback), and AI has no such mechanism.

"Part of understanding like this emotion that you're looking at is doing it yourself... you move your face muscles. And it's like you see that, you mimic that, and you're like, oh, yeah, that's anxiety, right?... how could it have theory of mind? It would be like this very emergent thing to have theory of mind. Right. Whereas we can just like plug it into our own minds." — Grant Sanderson 00:14:19 / 00:15:03

Who You Learn From Matters More Than What You Study

Counterintuitively, Grant argues that the identity of the teacher or author matters more than the subject matter when choosing what to learn — and this principle should govern how you use LLMs.

"Who matters more than what. So, like, advice to any college student when they're choosing what courses to take, care a little bit less about your pre-existing interests, because they're kind of arbitrary right now. And care a little bit more about whether, like, the person teaching it is a good educator and someone you resonate with." — Grant Sanderson 00:17:10

---

3. Companies Identified

Jane Street Quantitative trading and research firm. Mentioned for extraordinary employee retention, unusual cross-functional culture where roles blur, and general excellence as an employer of researchers, engineers, and traders.

"In the industry, they're known as having like a pretty wild retention rate... everyone's doing a little bit of everything else. Like even if you're officially a trader, you're doing a lot of research. Even if you're officially a researcher, you're doing a lot of coding. And I suspect maybe that's part of like why they have the insane retention that they do." — Grant Sanderson 00:15:10

OpenAI AI research company. Mentioned for publishing the result disproving the unit distance conjecture and for its IMO performance, framing how AI companies use math benchmarks as PR headlines.

"OpenAI can have their headline on disproving the unit distance conjecture because it's a clear distinct. It's like it did it, right?" — Grant Sanderson 00:09:57

DeepMind AI research lab. Mentioned for their initial approach to IMO using formal Lean proofs, then pivoting to natural language — a data point used to argue Lean is not essential.

"Initially DeepMind basically does that. It's like everything in Lean. And then the next year it's all in natural language." — Grant Sanderson 00:57:29

DeepSeek Chinese AI lab. Mentioned for their DeepSeek math model and their innovative training approach using a natural-language verifier trained by a meta-verifier, showing process-based supervision can work without Lean.

"DeepSeek had their DeepSeek math model and they released a paper on how they trained it... they have a verifier and then the verifier is trained by a meta-verifier that makes sure that all the problems that they're training this model to solve... the verifier is giving good feedback on that." — Dwarkesh Patel 01:03:08

Cursor AI-powered coding environment. Mentioned as a highly effective harness for real-world tasks including audio editing automation and research, illustrating the gap between raw model capability and practical productivity with good tooling.

"I don't think people appreciate the kinds of things that these models can just go handle for you when you equip them with a good harness like Cursor." — Dwarkesh Patel 00:52:49

Google (Gemini) AI division. Mentioned for releasing Gemini 3.5 Live Translate, which supports real-time translation across 70+ languages — used as a mid-roll sponsor illustration.

"Google just released Gemini 3.5 Live Translate... automatically detects more than 70 different languages and translates them in almost real time." — Dwarkesh Patel 00:25:16

Polylog (YouTube channel) Mathematics education YouTube channel. Mentioned favorably for producing a high-quality video explaining the unit distance conjecture breakthrough.

"For anyone curious about the unit distance conjecture, there's this really nice video by a math channel called Polylog where they talk about it." — Grant Sanderson 00:09:28

---

4. People Identified

Grant Sanderson Creator of 3Blue1Brown, mathematics educator and now documentary filmmaker covering AI progress in mathematics. Praised throughout for his ability to distill complex mathematical concepts and his nuanced understanding of the AI/math frontier.

"I will probably be doing something like what I am until I die." — Grant Sanderson 00:35:44

Terence Tao Fields Medal-winning mathematician. Mentioned as someone who failed an IMO problem that AI also struggled on, and separately for proposing an exhaustive search of possible algebraic axiom systems as a research project.

"Terry Tao also failed on it. And the nature of it is basically that people were very mad at the problem because they called it a troll problem." — Grant Sanderson 00:48:47 "Terry Tao was talking about one like research project that's basically try to exhaustively search the space of possible like algebras." — Grant Sanderson 01:01:25

Alexander Grothendieck (implied via Galois/mountain-building discussion) Implicitly referenced via the "mountain building" archetype — the kind of mathematician who creates new theoretical frameworks rather than solves existing problems.

Évariste Galois 19th-century French mathematician who invented group theory as a teenager while in prison, died in a duel at 20. Used as the central historical example of an idea with a 100-year verification loop.

"He's like this teenager. He's in prison. He had tried to submit his math papers and they had been rejected... He has some instinct that there's something there." — Grant Sanderson 00:19:12

Joseph-Louis Lagrange 18th-century mathematician who first recognized symmetry as the right lens for studying polynomials — planting the seed that took 50+ years to flower into group theory.

"Lagrange found the right kind of question to ask about this... the first time in history that people had the instinct that some kind of question about symmetry was the right way to be studying these polynomials." — Grant Sanderson 00:17:46

Niels Henrik Abel Norwegian mathematician who proved the quintic has no general radical solution, died at 26 from tuberculosis. Mentioned as the tragic parallel to Galois.

"Abel definitely read Lagrange and was influenced by it... he died young. He died at 26 from tuberculosis." — Grant Sanderson 00:18:12

Hugh Montgomery Number theorist who discovered the statistical correlation of Riemann zeta zeros — whose chance conversation with Freeman Dyson revealed a link to random matrix theory. Used as the canonical example of serendipitous cross-field connection.

"You have this number theorist who is pointing out, just trying to understand the statistical correlation between pairs of zeros of the Riemann's eta function... He writes down a formula that looks like one over sine squared or something like that." — Grant Sanderson 00:04:19

Freeman Dyson Physicist who recognized Montgomery's number theory formula as identical to one from random matrix theory / nuclear energy level statistics. Emblematic of the cross-field connection AIs should be able to make systematically.

"Freeman Dyson, a physicist, is like, I know that expression. That expression comes up in studying the eigenvalues for random Hermitian matrices, which was something that comes up in studying the energy levels of like a nucleus." — Grant Sanderson 00:04:47

Timothy Chow Mathematician who proposed the concept of "unsolved expository problems" — where a result is proven but not truly understood. Cited by Grant as a favorite intellectual framing.

"I want to propose the idea of an unsolved expository problem. We're like, sure, we've proven it, but we don't really know why it's true." — Grant Sanderson 00:32:45

Alex Konturovich Mathematician. Mentioned for articulating why a high error rate from AI-generated papers would be devastating to working mathematicians even if 99% correct.

"Alex Konturovich has talked about this it becomes insufferable like as a mathematician because you would basically be like every single time I see one of these I kind of don't know if it's worth my time." — Grant Sanderson 01:05:36

David Bessis Mathematician and author. Cited for his blog post "The Fall of the Theorem Economy," which argues theorem proving is parasitic on the more valuable work of coming up with definitions and insights.

"David Bessis had a really great blog post called the Fall of the Theorem Economy, where he's talking about this... really the theorem proving stuff is what gets all the credit, but it's like really a parasite on the coming up with the definition and stuff." — Dwarkesh Patel 00:29:53

Andrey Kolmogorov Mathematician. Mentioned in the context of Kolmogorov complexity as a potential formal tool for measuring elegance and compression in mathematical proofs.

"I don't know, Kolmogorov complexity. Like maybe you throw that into your like your attempt to quantify what you mean by elegance." — Grant Sanderson 00:23:56

Robert Langlands Mathematician. Mentioned as the originator of the Langlands program — a vast research ethos built on finding connections between mathematical structures, highly relevant as an AI target.

"Langlands was a mathematician. He has this like famous letter now essentially spelling out how it seems likely that there's a lot more connections like that." — Grant Sanderson 00:39:30

Andrey Wiles (implied) Implicitly referenced as the solver of Fermat's Last Theorem via elliptic curves and modular forms, used as the canonical example of "mountain building" mathematics.

Claude Shannon Founder of information theory. Cited as an example of a revolutionary thinker who was also an unusually clear writer — evidence that deep originality and clarity of exposition are correlated.

"People who are really coming up with something quite novel, so you've got like Einstein or like Claude Shannon or something there, you read their papers, they're really lucid papers." — Grant Sanderson 00:34:20

Richard Feynman Physicist. Cited alongside Shannon and Einstein as an example of someone whose depth of understanding correlated with exceptional expository ability.

"Feynman has this characteristic too, like very good expositor." — Grant Sanderson 00:34:42

Andrei Karpathy AI researcher. Mentioned for his "auto research" concept — a single Python file for LLM training where LLM agents try to make modifications, keeping those that improve performance.

"Karpathy's auto research idea. He wrote this basically one Python file that does basic LLM training and then just had a repo where LLM agents would like try to make modifications to the file." — Dwarkesh Patel 00:59:57

Eric Jang AI/robotics researcher. Mentioned for implementing a similar self-improving Go bot and observing that AI agents are good at going deep but bad at stopping dead ends and doing truly parallel exploration.

"Eric Jang, who came on to explain how AlphaGo works, did a similar thing when he was trying to build in a very strong Go bot. And he had interesting observations about the kinds of like it's really good at just running an experiment and going down that path. But it's bad at stopping at dead ends and just doing extremely parallel things." — Dwarkesh Patel 01:00:26

Annie Matushak Learning researcher. Mentioned for work attempting to train LLMs to write good spaced repetition flashcards — and concluding this requires sophisticated mental modeling of the learner's future state.

"Annie Matushak and another collaborator whose name I'm forgetting right now did an interesting report where they tried to teach LLMs to write good space repetition prompts." — Dwarkesh Patel 01:11:44

Stephen Strogatz Mathematician and author. Mentioned favorably for his textbook on nonlinear dynamics and chaos, which Dwarkesh was using as a learning resource.

"I was going through, I think you might have recommended Stephen Strogatz's textbook on... The Chaos one? Yeah. Chaos and Nonlinear Dynamics? I love that book." — Dwarkesh Patel 01:19:51

---

5. Operating Insights

Use LLMs as a Souped-Up Oracle for Finding the Right Human Resource, Not as the Primary Teacher

Grant's framework for using AI in learning: rather than trying to extract explanations from an LLM, use it to identify the best human-written book, video, or article, then use the LLM only to prune around branches that specific resource has identified.

"Basically using it like a very souped-up version of Google on, like, zero in on the right human-written resource... So often I like to just ask an LLM, like, who should I read?" — Grant Sanderson 01:18:37

"The most productive learning sessions I've had is when there's some artifact that a human has produced, whether it's an article, a book, a video, that organizes the relevant concepts in the correct way and builds up the motivation... And then using the LLMs to just do a little bit pruning around this branch that the book has identified." — Dwarkesh Patel 01:19:30

Jane Street's Cross-Functional Role Blurring as a Retention and Talent Flywheel

Jane Street's documented practice of having employees work across functions regardless of official title creates a continuous growth environment. This is a structural design choice, not just a culture statement, and correlates with exceptional retention.

"Even though the people have role titles, like, you know, researcher or trader or engineer, they often don't know what their colleague's actual role is because everyone's doing a little bit of everything else... I suspect maybe that's part of like why they have the insane retention that they do. Because anyone who wants to be growing, they just have the chance to do a lot of different kinds of things." — Grant Sanderson 01:15:40

Credit Assignment in Complex Systems: Containerize and Diff

The reason AI is making fast progress in coding and math but not in real-world tasks is a precise operating insight: if you can containerize a starting state and run deterministic parallel rollouts, you can solve credit assignment cleanly. Apply this principle when designing any AI-assisted workflow or research process.

"You can containerize a given level of progress in a repository and then just spin out thousands of parallel containers... And because it's deterministic, you can solve the credit assignment problem because you know that whatever caused this rollout to succeed and this one to fail, the diff is the thing that like worked." — Dwarkesh Patel 00:55:39

---

6. Overlooked Insights

Autoregression Is a Structurally Weird Way to Generate Insight — and This Explains Many AI Limitations at Once

Grant makes a brief but profound analogy: imagine you're locked in a box, receive a slip of paper, predict the next token, your memory is wiped, and this repeats. The resulting "essay" would be nothing like what you'd compose with normal cognition. This isn't a casual observation — it's a unified explanation for why AI struggles at writing, at making unlikely cross-field connections, and at theory building. It reframes many separate AI limitations as one structural problem, and implies that architectural innovation (beyond just scale or data) may be necessary for certain kinds of insight.

"Imagine I've locked you in a box, right? And then the only way that you have of interacting with the world is that you receive a slip of paper. And then someone says, can you like predict what will come next, right?... Imagine that was done a whole bunch. And then what comes out on the other end, they're like, look at this essay that you wrote. You might look at that and be like, this is awful. That's not the essay that I would have written, right?... the connection that actually is where all the substance is going to come from is like by its nature a very like unlikely one." — Grant Sanderson 00:42:09

The Fully Automated Mathlib Extension: An Entirely Novel Research Paradigm Nobody Is Fully Pursuing

Grant briefly floats the idea of a continuously running AI that simply tries to extend Mathlib — with no human checking in, no outcome-based supervision, just process correctness via Lean — and then you look at it in 10 years. This is not just a theoretical curiosity. It is a literally implementable research program today that would be structurally analogous to AlphaZero's self-play — but for all of mathematics. The potential to discover entirely new algebras, generate conjectures, and build new mountains of theory through pure compute investment, with zero human bottleneck, was mentioned only in passing but is arguably one of the most actionable and high-upside ideas in the conversation.

"You could imagine having a basically endlessly running program that's constantly trying to extend Mathlib... It might come up with its own conjectures. It might come up with its own theories... That's a very unique thing that math has that nothing else has where you could press go and then just like poor compute at it and like look away for 10 years and then come back and say like, what do you have?" — Grant Sanderson 00:58:43