Sep 8, 2026 · 1h 5m · a16z
Inside OpenAI’s Breakthroughs in Mathematical Reasoning
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
OpenAI researchers Mehtaab Sawhney and Mark Selke explore how frontier reasoning models are reshaping pure mathematics by overcoming human cognitive fatigue, automating intricate proof discovery, and resolving longstanding open conjectures in geometry and group theory. Through technical whiteboard breakdowns, they analyze the mechanics of AI deliberation while forecasting a future where human mathematicians shift from mechanical proof verification to conceptual synthesis.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The host as informed peer 5.1 Guest teaching 4.5 Guest disagreement 1.3 The host pushing back 2.4
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 40:04 Sawhney resists mystical definitions of mathematical taste
Sawhney dismisses abstract notions of aesthetic taste, bluntly taking a utilitarian stance that taste simply means solving hard problems faster by making better pruning choices.
Hardest push from the host ▶ 9:51 Li challenges whether models reason or just roll lucky samplesLi interrupts to push back against the assumption that models possess genuine human-like mathematical intuition, asking whether their success is just parallel lucky sampling.
Biggest teaching moment ▶ 23:11 Sawhney explains linear programming bounds and Fourier transformsSawhney educates the host on how Viazovska's Fields Medal work uses Fourier transform constraints on test functions to establish optimal sphere packing bounds.
The host holds their own ▶ 11:45 Li's critique of Rudin and textbook analysis pedagogyLi exhibits deep domain awareness by pointing out that polished textbooks like Baby Rudin strip out the messy heuristics and struggle that actually teach mathematicians how to reason.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The host as informed peer | Guest teaching | Guest disagreement | The host pushing back | Why |
|---|---|---|---|---|---|---|
| Introductions and the Journey from Pure Math to OpenAI | 5 | 3 | 1 | 2 | Li frames the conversation around her own background in mathematics and probes into how the models moved from basic literature search to deep mathematical reasoning. Selke and Sawhney collaboratively explain how GPT models moved from resolving open Erdős problem references to executing detailed technical arguments without getting lost in epsilon-delta subtleties. | |
| How AI Reasons: Tree Search, Backtracking, and Cognitive Bias | 5 | 4 | 2 | 3 | Sawhney explains how mathematical research is often a gamble against problem difficulty and how reasoning models prune search trees without human cognitive biases. Li actively probes whether the model is merely running parallel lucky samples or actively backtracking like a human mathematician, prompting Selke and Sawhney to clarify the nuances of model restarts versus human context pollution. | |
| Training Reasoning Models and Interpreting AI Chains of Thought | 6 | 3 | 1 | 2 | Li offers a sophisticated critique of training on standard mathematical corpora, noting that formal textbooks like Rudin obscure the messy intuition and struggle of discovery. The guests explain that OpenAI focuses on general-purpose reasoning rather than narrow math auto-formalization, producing chains of thought that read like a colleague's candid email notes. | |
| Whiteboard Deep Dive: High-Dimensional Sphere Packing Breakthrough | 4 | 7 | 1 | 2 | Sawhney and Selke conduct an extensive whiteboard lecture covering sphere packing bounds from dimension 2 up to Viazovska's work in dimensions 8 and 24, and Astra's breakthrough asymptotic LP bound. Li asks clarifying technical questions regarding geometry and linear programming while the guests lead the pedagogical exposition. | |
| Whiteboard Deep Dive: Spherical Codes and Information Theory | 5 | 6 | 1 | 2 | Selke illustrates the connection between spherical codes, error-correcting codes, and representation theory on curved surfaces. Li demonstrates quick domain comprehension on Hamming distances and geometry, while Selke reveals the surprising iterative prompt where asking Astra to push the representation theory led to solving the full-space packing bound. | |
| Mathematical Taste, Prompting Harnesses, and Model Collaboration | 6 | 4 | 2 | 4 | Li pushes deeply into the philosophical and engineering implications of mathematical 'taste', debating whether taste is emergent from task execution or requires a separate supervising harness model. Sawhney counters with a utilitarian definition of taste as solving problems faster, and Selke notes that having a supervisor agent check an underling agent mimics human collaborative checks. | |
| Whiteboard Deep Dive: Disproving Gromov's Conjecture on Sofic Groups | 5 | 6 | 1 | 2 | Selke explains group theory fundamentals and how Astra resolved Gromov's question on whether all groups are sofic by constructing a remarkably short counterexample. Li engages with the Cayley graph formulation and asks what combinatorial structure in the literature resisted finite approximation. | |
| The Changing Role of Mathematicians and the Future of the Field | 5 | 3 | 1 | 2 | Li and the guests discuss how AI-generated mathematics will shift the profession from manual proof construction toward synthesizing, communicating, and verifying concepts. Selke and Sawhney reflect optimistically on math becoming more accessible while noting the upper ceiling of problems like P vs NP will keep mathematicians relevant. |
Statements from this episode (19)
Selke: AI models consistently nail detailed mathematical execution where humans get lost
“Another relative strength that's pretty noticeable is just, like, it's very good at executing on some, like, idea once it has it. Like, you know, whenever you have an idea, there's, like, There's usually some amount of, you know, getting everything lined up, l…”
Sawhney: OpenAI models prune search trees instead of brute-forcing proofs
“You can sort of look at it, and it's reasoning like a mathematician, and because it knows a few very correct bits, it makes the right decisions and is eventually able to prune the search tree. It's not really trying everything. It tries a lot of different thin…”
Selke: AI avoids human cognitive bias by easily resetting polluted context
“Like, as a human, if you have some, like, wrong path you go down for a while, it can be hard to, like, rewire your brain to, like, start over and, like, try a different path. Like, you're kind of, the initial idea is kind of linked in your brain with these oth…”
Sawhney: AI models update proof path likelihoods better than humans do
“The model somehow is much better able to, like, it seems, for several of the solutions we've seen, somehow it seems much better able to update the solution, like, how likely the path is to work, like, versus rejecting a path versus a human doing it.”
Sawhney: Backtracking is a general-purpose reasoning tool, not math-specific
“A lot of these behaviors that we're describing mathematically, like backtracking, or kind of starting again, I mean, these are not really specific to mathematics. I mean, we're seeing them specifically in mathematics in these examples, but kind of, they're gen…”
Sawhney: Model reasoning traces resemble a collaborator's rough notes
“It's very much like reading a colleague's, like, notes. I mean, it's a little more disorganized in some way, but kind of, like, especially if you work close enough with a collaborator, sometimes you'll just see them, like, spill out their thoughts in an email …”
Sawhney: Astra proved the asymptotic linear programming bound for sphere packing
“And what the model shows is that Actually the linear programming bound in large dimensions has this extremely nice asymptotic behavior, and the proof kind of explains where this is coming from, and because you understand this LP bound perfectly, this actually …”
Sawhney: Astra's proof of the sphere packing bound is a few pages
“I think also in general, it was one of these solutions which, I knew several people had tried the problem, it's pretty remarkable because, like, the model solution, especially for this being, like, the LP can't do better than this, was, like, quite short. It's…”
Selke: OpenAI models discovered better bounds for spherical and binary codes
“Our models found better bounds for these cases as well.”
Selke: Coding theory was Astra's only proof requiring human interaction
“This was the one case where there was some interactivity involved. So for all of, so except for this pair, it was just, you know, we had some problems, we fed them in, and we you know, the model came back with some solutions.”
Selke: OpenAI's Astra improved bounds further simply when prompted again
“In, in this case, what the model was asked to do originally for codes was to improve the bounds by, like, some exponential factor, so it really, like, shows up in this, like, leading constant up here. And, you know, it improved the bounds, and it didn't try to…”
Sawhney: Solving harder math problems by definition demonstrates better AI taste
“I tend to be pretty utilitarian in my view of taste, and, like, if you're able to solve problems faster by making better judgments, like, I think that's, like, the best, like, general proxy I have for taste, and somehow the fact that solving harder problems me…”
Selke: OpenAI's Astra proved that a non-sofic group exists
“So, so the result that Astra proved is simply that there exists a non-sulfic group.”
Selke: Only humans can generate 200-page mathematical proofs right now
“Like, only humans can generate, like, 200 page proofs right now.”
Sawhney: AI will produce exponentially more math, making it easier to absorb
“Along, I mean, of course models are going to help us produce exponentially more mathematics, but they also make it much easier to absorb it and right now, okay, it's still a bit of a challenge back and forth, but I think it's, for me at least much, much faster…”
Selke: Proving mathematical results is becoming much less of a bottleneck
“Proving the result was, like, so hard that kind of the other stuff was just kind of coming along for the ride, right? You know, like, if you manage to, like, prove this thing yourself, you're automatically gonna understand it quite well. You're kind of respons…”
Sawhney: Fostering human understanding will become a more valuable role in math
“Increasingly it would be a function of, like, you're sort of helping, you're the human who can sort of give this understanding to other people and sort of help them with it. I think that more, sort of, that communal understanding will, I think, become, it was …”
Selke: AI might plausibly never solve problems like P versus NP
“Even if AI get, you know, continues getting, like, exponentially better at math, like, it might, you know, plausible will never solve something like P versus NP.”
Selke: Non-mathematicians will be able to use advanced math without experts
“The ability of someone who's not working on math is, like, their literal job all the time to, like, understand what's going on and, like, you know, learn about some of the mysteries they might have wondered about will go up quite a lot. Also, you know, if you'…”