Notes from Dwarkesh Patel's conversation with Grant Sanderson: Grant Sanderson (3Blue1Brown) – AI and the future of math.
The useful frame is not simply "AI can do math now." It is that math is becoming the cleanest place to watch several different kinds of intelligence separate from each other.
A model can solve a contest problem. A model can connect two existing fields. A model might someday build a new mathematical theory. A model might produce ten thousand correct but unreadable proofs. Those are not the same achievement.
Sanderson's strongest point is that AI progress in math forces a distinction between proof, understanding, and taste.
Patel opens by revisiting an old question: if an AI can get gold at the International Math Olympiad, isn't that basically AGI?
Sanderson's answer from years ago aged well: no. It would be another benchmark. Impressive, but not a magic threshold.
The reason is that math has a spiky internal frontier. Even inside Olympiad math, some subdomains are much more amenable to current methods than others. Sanderson points to geometry as an area where specialized systems can do extremely well, partly because there are brute-force-ish ways to attack it. Combinatorics, by contrast, remains more playful and slippery.
So the lesson is not "AI can do IMO, therefore it can do everything." The lesson is more precise:
> Benchmarks reveal local capability. They do not automatically reveal general intelligence.
This is an important correction because math problems feel like they require creativity. Often they do. But the kind of creativity needed for a contest problem is not identical to the kind needed to run a company, write a great essay, mentor a student, or decide what problem a field should study next.
The conversation gets more interesting when Patel moves the goalpost from IMO to Millennium Prize problems.
If an AI solves the Riemann hypothesis, does that imply it can automate huge parts of the economy?
Sanderson's answer depends on what the solution looks like.
One possibility is that the model connects existing fields. The Riemann hypothesis already has this historical flavor: Hugh Montgomery's work on zeros of the zeta function unexpectedly connected with Freeman Dyson's knowledge of random Hermitian matrices. That kind of lightning bolt between distant domains seems like exactly where LLMs should eventually be strong: they have breadth across many expert literatures and can search for analogies between them.
Another possibility is harder: the solution requires building a new theoretical mountain. Fermat's Last Theorem was not solved by a cute elementary trick. Its known proof depends on deep machinery involving elliptic curves and modular forms. If AI can build a new mountain like that — not just find a clever bridge between existing mountains — that would feel much closer to a world-changing intelligence.
A third possibility is less satisfying: AI produces a huge proof that is technically correct but gives humans little compression. That would settle a theorem but not necessarily create understanding.
The category matters. "AI solved a famous problem" is not one achievement. It could mean cross-domain search, theory invention, brute-force proof production, or something in between.
Once models can solve known problems, the natural next question is whether they can decide which problems are worth solving.
Sanderson cites the old hierarchy:
> Good mathematicians prove theorems. Great mathematicians make conjectures. The greatest mathematicians create definitions.
This is where the benchmark frame starts to break.
A theorem can be verified. A contest answer can be scored. A conjecture or definition is harder. You can say a model proposed a conjecture, but how do you know whether it is a good one today? Some ideas only reveal their value after a field reorganizes around them.
That means the real signal may be social rather than leaderboard-shaped. We may not get a clean headline saying "GPT-6 passed the definition-generation benchmark." Instead, mathematicians may start saying that conversations with models changed what they thought their field should study.
That would be a quieter milestone, but probably a more important one.
The Galois discussion is the best case study in the episode.
The problem begins with formulas for polynomial roots. Quadratics have a familiar formula. Cubics and quartics have formulas too, though they are ugly. The natural question was whether quintics have a general formula.
Lagrange made a crucial move by connecting the problem to symmetry: how expressions change or stay fixed when variables are permuted. Abel proved that a general quintic formula is impossible. Galois went further and exposed a deeper structure behind when equations are solvable.
But the value of Galois's work was not immediately obvious. It was hard to parse, initially rejected, and only later developed into the language that became group theory.
That is the problem for AI evaluation. If the most important mathematical contribution is a new way of seeing, the verification loop may be decades long. A short-term reward signal can check correctness; it cannot easily check whether an idea will become a generative lens for cryptography, physics, and future mathematics.
This is the uncomfortable part: the more profound the contribution, the less it may look like something current RL-style training can reward cleanly.
A recurring distinction in the conversation is proof versus explanation.
A proof can be correct and still fail to make the result feel inevitable. Mathematicians often care about a stronger thing: the compressed structure that reveals why the theorem is true.
That matters for AI because models may soon generate far more correct mathematics than humans can read. If the output is only a flood of formal proofs, the field gets a new bottleneck: attention.
Sanderson and Patel both circle around the same idea: understanding is compression. It is the difference between a thousand-page argument and a conceptual framework that makes the result feel natural.
So the valuable AI mathematician is not merely the one that proves more. It is the one that finds the abstraction that makes many facts smaller.
Lean and formal verification come up as both overhyped and underrated.
Overhyped, because current AI-math progress does not seem to depend entirely on Lean. Models are already doing impressive natural-language mathematical work. Math and code are also useful training domains because they are grindable: you can make many attempts, get feedback, and improve.
Underrated, because once models generate proofs at scale, humans will desperately need trust signals.
If AI systems produce ten mathematical papers a day, even a 1% error rate is exhausting. A mathematician cannot afford to spend weeks finding the subtle flaw in a trash proof. A formal verifier changes that: it gives a green checkmark saying, at minimum, this result is correct.
Lean may matter less as the source of intelligence and more as infrastructure for trust and autonomy. A model could explore a fork of Mathlib, extend formal mathematics, and leave humans with a different question: not "is this true?" but "which true things are interesting?"
The conversation then pivots to writing, and the contrast is useful.
LLMs are already good explainers of known material. Both Patel and Sanderson describe using them as an enhanced search and clarification layer. If the concept is well represented in the training distribution, asking an LLM for an explanation can be better than reading a mediocre human summary.
But good writing is not only clear distillation. It also requires deciding what is worth saying, making the right unpredictable move, and modeling the reader's mind sentence by sentence.
That is much harder to reward than a proof or a passing test. In code, ugly internals can still produce the right behavior. In writing, the output itself is the substance. Every sentence is part of the final artifact.
This helps explain why math and code are racing ahead: they offer better feedback loops. Writing has no equivalent of a compiler, a unit test suite, or a proof checker for insight.
Sanderson's learning advice is refreshingly practical.
Before LLMs, he already believed that *who* teaches often matters more than *what* topic you choose. A great teacher, textbook, or author supplies motivation and ordering. They know which idea should come next and why it matters.
LLMs are useful around that path, not necessarily as the path itself.
The best learning setup is something like:
The model is excellent at pruning branches around a well-designed route. It is weaker at designing the route from scratch, especially when the student is asking the wrong question and needs a reframing rather than an answer.
That is still a real limitation. A great teacher can hear a confused question and say: "Actually, the way you're organizing this topic is off. Try this frame instead." LLMs are often too placating to do that well.
Sanderson is careful not to pretend he can give definitive career advice. But his answer is stronger than the usual "just learn to use AI."
He says students should understand where value comes from.
Many people pursue math because they have been rewarded for being good at academic hoops. That is not the same as understanding the economic and social role of mathematicians. Money and prestige can come from research, grants, university brand value, teaching, public-good basic science, applied work, or explanation.
AI may shift the mix. If theorem proving becomes abundant, other roles become more visible:
Sanderson is especially bullish on teaching as a stable role. Teaching is relational. It is coaching, motivation, taste, trust, and timing — not just explanation. Even if AI becomes an excellent explainer, people may still want human guides.
The museum analogy is apt. If AI creates an overwhelming abundance of mathematical art, humans may still need curators.
The episode avoids an easy answer.
Some areas of math are application-adjacent. Progress in PDEs, simulation, optimization, numerical methods, or materials-adjacent mathematics could plausibly leak into engineering and industry. Sanderson mentions work connected to simulation that saved Boeing huge amounts of money; that kind of math clearly touches the physical world.
Other areas may have long, indirect, or unclear practical paths. A burst of pure math progress might produce enormous intellectual output without immediate economic effect.
That uncertainty itself is revealing. If AI accelerates mathematics by 10x or 100x and nothing else changes, it may force mathematicians to confront which parts of the field are connected to outside value and which parts are mostly internal culture.
But the more likely picture is mixed: some AI-math progress becomes economically useful at the edges, some becomes intellectual infrastructure for later, and some remains beautiful but detached.
The deepest point in the conversation is that AI is not just automating math. It is exposing what math was for.
If the scarce thing was calculation, machines already changed that. If the scarce thing was proof search, AI may change that next. But if the scarce thing is knowing which abstractions matter, which conjectures are fertile, which definitions compress reality, and which true statements deserve human attention, then the future of math is less about replacement and more about a new bottleneck.
The field may soon have more proofs than it can read.
The hard part will be deciding which ones are worth understanding.