Mathematicians Feared AI Would Write a Thousand-Page Proof. It Wrote 15 Pages.
Humans disproved a conjecture about random graphs in 250 pages, built on another 200; inside group theory, AI proved the weaker result in a dozen-odd pages. AI's math proofs are generally short and elegant — long proofs, it turns out, only humans write.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The first thing models replace is literature search
Erdős's problems were collected onto a website, and a problem marked open isn't necessarily unsolved — the mathematical literature itself is hard to search. One guest was unsure about one of them, threw it into GPT-5, and five minutes later the model produced a reference, which amounted to telling him the problem was within reach and this was how to do it. He and a few friends had already spent hours on it without being sure they could solve it. They then found ten-odd more cases of the same kind. The model's first beachhead isn't proving, but the retrieval and detail execution that eats the most time for mathematicians.
The model isn't smarter, it's more willing to bet
A human mathematician's day is "gambling with problems": you have an idea, you think it might work, you try it for hours, days, weeks, then give up; a year or two later you learn someone else got the approach you'd thought of to work. The model's difference isn't inspiration — it's that it doesn't compute risk and reward. If a human tells it to do it, it does it to the end. And it isn't brute-forcing: it tries many different paths, very persistently, but it can only try finitely many ideas, pruning the search tree with existing knowledge plus mathematical judgment. On the unit distance problem, the idea may have been thought of long ago; what's genuinely hard is aligning all the extremely tedious execution details, and that's the part humans get lost in and the model gets almost never wrong.
Humans get polluted by their own thinking; models don't
Once a human has gone down a wrong path for a while, the earliest idea gets bound up with everything that failed afterward — in the guest's words, the context window is polluted, and you can't clone a clean session off last week's self and say, don't do this, rebuild your intuition in another direction. A model can, and worst case it just starts a new game. Its way of updating "how likely is this path to work" is also more computational: a human automatically downgrades confidence in a method after the first failure, while the model seems better at repeatedly adjusting probabilities rather than killing a path outright. So it does backtrack and make mistakes — it just doesn't get dragged down by them.
The model turned a numerical guess into an equals sign
The sphere packing problem has exact answers in only five dimensions: 1, 2, 3, 8, 24; the 3-dimensional proof is at minimum several hundred pages and famously ugly. The long-standing best upper bound came from a two-page 1970s paper by two Russian mathematicians, with a constant of 2 to the -0.599d — the solution to some extremely ugly optimization problem. Viazovska's linear programming framework hits the optimum exactly only in dimensions 8 and 24, which is a major reason she won the Fields Medal. What the model did was not shrink the constant a little, but prove that the framework's optimal value in large dimensions is itself that beautiful asymptotic expression: an equals sign, not an inequality. One guest spent about six months on this problem during his PhD with zero progress, and finally watched the model finish it off with a few pages of complex analysis.
One "push it further" buys a conjectured value
Of the ten problems, only the pair of spherical codes and binary codes interacted: they first had the model improve the code bound using representation theory, and the model stopped once it was done; so they asked whether it could push further, and the model produced a far more complicated representation-theoretic version — and taking that path to its limit actually recovered the conjectured value for sphere packing in large dimensions. The two pieces of work weren't a coincidence but two ends of the same method. That also prompted the host's question: is this the model lacking judgment, or a harness problem? The guest's answer is task orientation: it finishes what you assigned and clocks out, not that it's incapable — ask again and it goes further.
It's not that it can't push, it's that it wasn't asked to
The more honest summary is: this isn't a capability problem, it just didn't want to at the time. Task orientation is a double-edged knife — the model does what you assigned, then stops; sometimes it clearly realizes it has made a breakthrough and still won't push to the limit on its own, because that isn't what you asked. From there comes the discussion of taste: one utilitarian definition holds that taste is being able to solve problems faster through better judgment; another holds that taste is knowing which problems and which methods are worth investing in. One compromise is to split it: let one model own taste and another act as its subordinate doing the long-horizon grunt work, the two not needing to share context and therefore not polluting each other.
AI does in a dozen-odd pages what takes humans hundreds
A sofic group, roughly, is a group that can be approximated by finite groups; until now no one knew whether non-sofic groups exist. A stronger conjecture (Aldous–Lyons) was just disproved two years ago: that work was called a tour de force, 250 pages, built on another 200, drawing on quantum complexity theory, and few people can read it. Proving the weaker conclusion that non-sofic groups exist took the model only about 15 pages, staying inside group theory the whole way and borrowing only some existing group-theoretic results. The guest said that a year ago he would have assumed an AI proof would be a thousand pages of incomprehensible stuff; the opposite turned out to be true — now only humans write two-hundred-page proofs.
Proof is no longer the bottleneck; understanding is
The old logic was: proving a result is so hard that the understanding, absorption and explanation that follow come along for free — what you proved yourself, you naturally understand deeply and naturally take responsibility for explaining to others. Now the proof bottleneck has been largely shaved away, and understanding, putting a result into the right framework, and explaining it to others have instead become explicit contributions that get rewarded. But the ceiling remains: the upper limit on the difficulty of mathematical problems is very high, and even if models keep getting exponentially stronger, problems like P versus NP may never be solved. So mathematics is more likely to organize itself around these grand puzzles, with the small puzzles handed to machines — and by the way, models both produce exponentially more mathematics and make absorption faster; they are solving the problems they themselves create.
In their own words · checked verbatim
It's not really trying everything. It tries a lot of different things. It's extremely dogged.
It's sort of like your context window is a little polluted, and you can't just make another clone of yourself from last week and say, don't do this, try something else, build your intuition in another direction.
I had actually thought about this problem for about six months at some point when I was a graduate student. And, yeah, I just, I remember making, like, absolutely zero progress on it.
So it wasn't like a capabilities issue. It just didn't feel like it at the time.
It was like 250 pages building on another 200 pages. It uses like quantum complexity theory.
It's like 15 pages maybe. And it doesn't have any of this like very complicated connection with quantum complexity.
Only humans can generate 200-page proofs right now.
I mean, a nice thing about math is that the ceiling for difficulty of a math problem is pretty high.
Figures
| Time for GPT-5 to find a reference | 5 minutes | 4:08 |
| Similar successful literature-search cases | 10 more found | 4:08 |
| Number of open problems in Astra | 10 | 15:18 |
| Length of the shortest proof for 3-dimensional sphere packing | several hundred pages | 17:22 |
| Number of dimensions with exact answers for sphere packing | 5 (dimensions 1, 2, 3, 8, 24) | 18:23 |
| Constant of the long-standing best upper bound for sphere packing | 2 to the -0.599d | 20:27 |
| Asymptotic constant of the linear programming bound given by the model | about 2 to the -0.601…d | 22:27 |
| Length of the human disproof of the Aldous–Lyons conjecture | 250 pages, built on another 200 | 50:57 |
| Length of the AI proof that non-sofic groups exist | about 15 pages | 51:57 |
Glossary
- sphere packing
- Finding the densest arrangement of spheres in d-dimensional space; a classic geometry problem.
- linear programming bound
- Constructing a function satisfying two properties to turn sphere packing density into an optimizable upper bound.
- spherical code
- Moving sphere packing onto the sphere, equivalent to studying the rate limit of error-correcting codes.
- sofic group
- A group that can in some sense be approximated by finite groups; until now it was unknown whether non-sofic groups exist.
- Aldous–Lyons conjecture
- Any infinite random graph satisfying unimodularity can be approximated by large finite graphs.
- harness
- The prompts, tools and training pipeline designed by humans outside the model, which together determine final performance.
How to listen
Readers building AI applications or investing, especially those trying to judge whether models can only brute-force search; and researchers wondering how the scientific workflow gets rewritten.
From [43:45], the introductory groundwork on "what is a group" can be fast-forwarded.