OpenAI Researchers on the Future of Mathematical Reasoning

Key Themes

The Bottleneck in Mathematics Has Shifted from Proving to Understanding

For decades, the scarce resource in mathematics was the ability to prove a result — and understanding/communicating that result came "along for the ride" because the same person who struggled to prove it necessarily understood it deeply. AI breaks this coupling. Mark Sellke explains: "a couple of years ago, like, proving the result was, like, so hard that kind of the other stuff was just kind of coming along for the ride... if you manage to, like, prove this thing yourself, you're automatically going to understand it quite well... And, yeah, now this kind of, what was the main bottleneck before is kind of much less of a bottleneck." [00:59:46] This implies the valuable human skill going forward is synthesis and pedagogy — explaining and integrating new results into humanity's knowledge base — not just generating proofs.

Risk/Reward Asymmetry Explains Why AI Is Unlocking "Reachable" Results

Many of the problems AI is solving were not fundamentally beyond human ability — they were beyond what a human would rationally invest time in given uncertain payoff. Leisha Li frames it: "for a human, you're like, well, I only have a limited amount of time. And if after so many steps, it is still not clear... it's not clear that the risk reward is not good enough. Whereas for GPT, okay, a human told me to do this. Let's just do this. And so that's why we're sort of in this renaissance of like reachable results." [00:05:55] Mehtaab Sawhney adds that mathematicians often abandon promising ideas only to learn years later someone else made them work: "somehow getting an idea to work can even be a large portion of the battle." [00:07:19]

AI Doesn't Get "Polluted" by Failed Approaches the Way Humans Do

A distinctly human cognitive limitation — that a failed line of reasoning contaminates your mental context and makes it hard to restart cleanly — doesn't bind AI in the same way. Mark Sellke: "your context window is like 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... But it's very easy to do this with an AI." [00:08:49] This is presented as a structural, not just capability-based, advantage.

AI Proofs Are Reasoning Like Mathematicians, Not Brute-Forcing

Both researchers stress the reasoning traces resemble expert human thought rather than exhaustive search. Mehtaab Sawhney: "It's very much like reading a colleague's like notes... they kind of like, especially if you were close enough with the collaborator, sometimes you'll just see them like spill out their thoughts in an email to you. And it kind of feels like reading a lot of those chained together." [00:15:11] OpenAI released summarized chains of thought specifically to counter the assumption that models are "guessing in some insane way": Mark Sellke, "is the model just guessing in some insane way?... But actually it's reasoning kind of shockingly like an expert human would." [00:00:17]

AI-Generated Proofs Are Surprisingly Short and Elegant, Not Brute-Force Sprawl

Contrary to fears that AI would produce unreadable, massive proofs, the opposite has occurred. Mark Sellke: "I think it's, like, a year ago, I would have been very surprised to learn that, like, all of these AI proofs are, like, very short and elegant... you're kind of afraid that they're going to generate all these thousand-page things... But it's been kind of the opposite... only humans can generate, like, 200-page proofs right now." [00:55:28]

Emergent Mathematical Reasoning Is Not a Product of Training on Math-Specific Corpora

The strong mathematical judgment displayed by the models is not because they were narrowly trained to auto-formalize proofs in systems like Lean — it's a byproduct of general-purpose reasoning training. Mehtaab Sawhney: "we're training general purpose reasoning models and... a lot of these behaviors that we're describing mathematically like backtracking or kind of starting again... are not really specific to mathematics... they're general purpose tools for reasoning." [00:12:33] Leisha Li notes this is philosophically strange given how poor math papers/textbooks are as training data for "how to think," since they present cleaned-up final results rather than the messy motivating struggle: "papers also... most people don't write papers with the context of I need to educate somebody to be a mathematician." [00:11:24]

Mathematics Will Bifurcate into "Big Mysteries" vs. Routine Problems

Mark Sellke predicts the field's center of gravity shifts upward in difficulty as AI clears out the solvable-but-labor-intensive layer: "a nice thing about math is that the ceiling for difficulty of a math problem is pretty high. So even if... AI continues getting, like, exponentially better at math, like, it might... plausibly never solve something like P versus NP... it could be that the field kind of becomes more attached to these big mysteries and less to these smaller mysteries that are more routine now." [01:01:54]

Democratization of Mathematical Understanding and Applied Math Acceleration

Both researchers see a future where non-specialists can access deep mathematical machinery without needing a world expert. Mark Sellke: "if you're working on something that requires some math, you know, suddenly you don't need to... find a world expert on this topic to be able to... use it in your own work." [01:03:13] Mehtaab Sawhney explicitly ties this to real-world economic impact: "It'd be nice for the world if applied mathematics went a lot faster." [00:00:27]

Taste Is Best Understood Utilitarianly — As Whatever Produces Faster Problem-Solving

Rather than treating mathematical "taste" as an ineffable aesthetic quality, Mehtaab Sawhney defines it operationally: "if you're able to solve problems faster by making better judgments... that's the best general proxy I have for taste... the fact that solving harder problems means it has — kind of by definition means it has better taste." [00:39:15]

Contrarian Perspectives

The Hardest Part of Math Isn't Having Ideas — It's Having the Courage/Judgment to Execute Them

Most people assume the bottleneck in unsolved math problems is a missing insight. Mehtaab Sawhney argues instead that many "unsolved" problems have already had their correct approach identified by humans — the barrier was executing under uncertainty: "somehow getting an idea to work can even be a large portion of the battle... very often when you're doing mathematics, you're kind of gambling against the problem. You're like, maybe I should try this approach, but it seems really unlikely and just not worth my time." [00:07:19]/[00:07:49]

A Single "Taste Model" Separate from a "Grinding Model" May Be the Right Architecture

Rather than one model needing to embody both long-horizon execution and big-picture judgment simultaneously, Mark Sellke suggests these should possibly be separated to avoid cognitive interference: "if you have, like, one model that's responsible for taste and one model that's responsible for going out and... working for a long time at solving a hard problem kind of as the... underling of the supervising AI, I feel like that's kind of going to be fine currently." [00:40:10] This runs counter to the more common industry framing of a single model scaling toward general research intelligence.

The Legendary "Ugly" Human Proofs (Hales' Sphere Packing) May Reflect a Failure Mode, Not True Difficulty

Mehtaab Sawhney notes the human proof of the D=3 Kepler conjecture is "famously ugly," running hundreds of pages using messy linear programming/geometry [00:17:59], while the AI-derived asymptotic sphere-packing bound was described as "a few pages... but it's kind of exactly the right approach... it's unbelievable. Like, why hasn't somebody done this before?" [00:27:45] The implication, understated by the speakers, is that decades of human mathematical effort in this area may have been organizationally or cognitively locked into suboptimal frameworks rather than facing genuine intractability.

Auto-Formalization / Lean-Based Training Is Not the Reason Models Are Good at Math

Given how much of the applied-AI-for-math world (Lean, formal verification communities) assumes formal proof corpora are the path forward, it's notable that OpenAI's mathematicians attribute the model's mathematical judgment to general reasoning training rather than formal-math-specific data: "it doesn't rely on... doing auto formalization in order to... guide the reasoning." [00:12:58] This is a quiet but significant divergence from a popular industry thesis.

Mathematicians' Discomfort with AI Co-Authorship Is a Sideshow, Not the Real Story

While Leisha Li raises the sociological friction ("people... would find it hard to either take AI as a co-author or... how do you even do kind of attribution" [00:57:06]), the researchers largely wave this off in favor of a much more radical claim: that the entire value chain of mathematical labor — from proving to understanding to disseminating — is being restructured, making credit-attribution debates a relatively minor concern next to the bigger shift.

Companies Identified

OpenAI — AI research lab; developer of GPT-5 and the "Astra" mathematical reasoning system discussed throughout. Mentioned as having produced ten new mathematical results including advances in sphere packing, coding theory, and group theory that resisted human mathematicians for decades. Mark Sellke: "we're training general purpose reasoning models and... if you work hard at reasoning, you should see these patterns eventually." [00:12:33] Mehtaab Sawhney on Astra's sphere packing result: "the model shows that... the linear programming bound in large dimensions has this extremely nice asymptotic behavior. And the proof kind of explains where this is come from." [00:26:19]

People Identified

Mark Sellke — Mathematician, now at OpenAI, working on Astra and mathematical reasoning models. Joined OpenAI after being inspired by the IMO gold medal result: "I saw the IMO gold medal last summer, basically. I thought, this is amazing. I want to see what the heck they did." [00:02:41] Contributed key framing on sofic groups, sphere packing, and codes results.

Mehtaab Sawhney — Mathematician (combinatorics), now at OpenAI, converted to AI-for-math after Mark Sellke gave him GPT-5 access: "very quickly became convinced that, yeah, it was extremely exciting to play with them." [00:02:58] Described using GPT-5 to instantly resolve open-problem status on an Erdős problem site that had stumped him and colleagues for hours: "I just plugged it into GPT-5 and five minutes later it found a reference." [00:03:45]

Leisha Li — a16z infra partner and former practicing mathematician, hosts the conversation, connects the dots between OpenAI's results and broader implications for the practice of mathematics, and personally frames several of the conceptual questions (taste, harness vs. model, democratization of math).

Yufei (Zhao) — Referenced as the shared academic advisor connecting Leisha Li, Mark Sellke, and Mehtaab Sawhney: "We're connected actually because Yufei was actually your advisor." [00:01:47]

Paul Erdős — Famous mathematician whose collected open problems (on a dedicated site) are a key proving ground; Mehtaab Sawhney describes using GPT-5 to check the status of these problems. [00:03:23]

Andrew Wiles — Referenced as the archetype of a mathematician willing to spend a decade alone on a single hard problem, contrasted with the AI's "just try it" approach to reachable results. Leisha Li: "maybe you're like Andrew Wiles and you actually spent 10 years alone and do something." [00:06:25]

Maryna Viazovska — (referred to as "Villazosca" in transcript) Fields Medalist (2022) credited with constructing the linear programming functions that gave optimal sphere-packing bounds in 8 and 24 dimensions. Mehtaab Sawhney: "a large part of the reason she won a Fields Medal in 2020... or in 2022 was that she constructed a function in 8 and 24 dimensions such that this upper bound matches exactly these two very special lattices." [01:25:26 / 00:25:26]

Henry Cohn — Mathematician referenced as co-author of the earlier numerics paper that conjectured (without proof) the asymptotic sphere-packing bound Astra later proved: "There was a numerics paper by Cohn and several others which conjectured that, just based on doing numerics, that this was the answer. But they had no idea why this would be the answer." [00:25:56]

Kabatjanskii and Levenshtein ("Kaptyonski" in transcript) — Two Russian mathematicians who in the 1970s established the previous best bound on sphere packing density using an ugly optimization argument. Mark Sellke: "It's the answer to some extremely ugly optimization problem." [00:20:35]

Henry Cohn / Noga Kun and Tom ("Kun and Tom") — Mathematicians whose earlier work the AI's non-sofic group counterexample built upon, and who have since produced human follow-up work extending the AI's result. Mark Sellke: "there has been some follow-up on this already, actually, by Kun and Tom, who this was always built on." [00:56:17]

David Aldous and Russell Lyons ("Aldous and Lyons") — Mathematicians who popularized the stronger Aldous-Lyons conjecture (about unimodular random graphs) related to the sofic groups result; this conjecture was disproved two years prior via a 250-page proof using quantum complexity theory, described as "a really tour de force... work." [00:51:06]

W.H. Gottschalk — Mathematician credited with originating the 1970s question of whether every group is surjunctive, which motivated the sofic groups conjecture. [00:47:57]

Operating Insights

Ask the Model to "Push Further" — Interactivity Can Unlock Compounding Gains

When OpenAI researchers interactively re-prompted Astra to extend an initial partial improvement on coding-theory bounds, the model discovered progressively more sophisticated representation-theory techniques that ultimately reproduced the conjectured optimal sphere-packing constant — revealing a previously unnoticed connection between two problem domains. Mark Sellke: "we first asked it to improve the bounds for the codes... And then we kind of asked it, hey, can you push this further?... then it came back with some much more sophisticated representation theory. And it turned out that you got this conjectured value for full space sphere packing." [00:34:28] The lesson for operators: don't treat a model's first output as its ceiling — the marginal cost of asking "go further" is low and the marginal yield can be a qualitative jump.

Release Reasoning Traces, Not Just Results, to Build Trust and Enable Verification

OpenAI deliberately published summarized chains of thought alongside the proofs specifically because bare results invite suspicion of "black box" guessing. Mark Sellke: "I think this kind of thing is one reason we decided it was important to release like these summarized chains of thought for these kinds of results. Because if you've never seen these and you just see all these proofs coming out, you're... not sure what it means." [00:10:41] This is a broader operating principle for any AI product making novel claims: transparency into process, not just output, is what converts skeptics.

Use Models to Compress the "Absorption" Time for New Knowledge, Not Just Generation Time

Mehtaab Sawhney describes a personal workflow shift: instead of reading an arXiv paper's introduction line by line, he feeds the PDF into a model to extract the proof strategy. "It's much, much faster to understand a piece of mathematics with a model than without it. So it's helping solve the problem it creates anyways." [00:58:03] This reframes AI's value proposition beyond "doing the work for you" to "shortening the loop of understanding work done by others" — directly applicable to any domain drowning in a rising volume of outputs (research papers, codebases, legal filings).

Task-Specificity of Prompts Can Mask or Reveal Capability — Design Prompts to Separate the Two

The researchers found that when a model didn't extend a result further, it wasn't a capability ceiling — it was literal task compliance. Mehtaab Sawhney: "models tend to be pretty task-oriented. But if you tell it to do a task, it accomplishes the task." [00:36:41] Operators evaluating model capability should be wary of concluding "the model can't do X" when the real issue may be "we didn't ask for X."

Overlooked Insights

The Real Constraint Was Never Intelligence — It Was Risk-Adjusted Time Allocation

Buried in the discussion of "reachable results" is a much bigger claim about the economics of human intellectual labor: talented mathematicians have been leaving correct ideas on the table for decades not because they lacked the ability to develop them, but because the expected value calculation ("try for a few hours, a few weeks... at some point you give up" [00:00:00]) didn't justify the investment given uncertain payoff and career/reputation risk. This reframes AI's contribution to any knowledge-work field: the unlock isn't necessarily raw capability beyond the best humans, but the removal of opportunity-cost constraints on trying ideas that were plausible but not "worth it" for a human to gamble career time on. This has direct implications for R&D-heavy industries (drug discovery, materials science, theoretical physics) where similarly "abandoned but plausible" ideas likely litter the literature.

Proprietary Harnesses May Be a Deliberate Moat, Not Just a Technical Necessity

In a single aside, Leisha Li speculates that the need to train models jointly with their harness ("training regimes where we require the harness to be... trained with... maybe part of this is to keep things more proprietary and harder for other people to use it" [00:42:00]) hints at an underappreciated competitive dynamic: as harnesses become co-trained with models rather than being thin, swappable wrappers, the defensibility of frontier AI labs shifts from "who has the best base model" to "who has the best model+harness bundle that cannot be easily replicated by prompting a competitor's raw model." This is a subtle but important moat thesis for anyone evaluating AI infrastructure investments — it suggests the "wrapper companies are doomed" narrative may invert at the frontier, where harness-model co-design becomes the durable asset.