Summary

OpenAI mathematicians Mark Sellke and Mehtaab Sawhney described how Astra was applied to a ten-problem set, headlined by determining the asymptotic Cohn–Elkies linear-programming bound for sphere packing in high dimensions. Astra constructed a function attaining the bound and proved that no function in that framework does better—an equality previously conjectured from numerics by Cohn and collaborators, who “had no idea why this would be the answer”—improving a long-standing density upper bound. The prompt was essentially “analyze this linear program in high dimensions… go have fun.”
The model’s edge is execution, correct bets, and the ability to restart without a polluted context. Sellke says models are unusually good at getting technical details right once they have an idea. In the unit-distance case, the model made the correct bet on an approach mathematicians had tried and pruned the search tree rather than brute-forcing; it still makes mistakes and goes back to reconsider them. “It’s not really trying everything.”
The guests linked mathematician-like reasoning to general-purpose reasoning work, rather than establishing a math-specific recipe. Sellke says behaviors such as backtracking and starting again “aren’t really specific to mathematics.” OpenAI released summarized reasoning traces partly to show that the model was not simply guessing; Sawhney compared them to reading a colleague’s notes.
Astra produced a roughly 15-page proof that a non-sofic group exists, while the related disproof of the Aldous–Lyons conjecture about two years earlier took ~250 pages atop another 200 and used quantum complexity theory. One guest’s surprise inverts the feared scenario: “a year ago I would have been very surprised to learn that all of these AI proofs are very short and elegant.” Kun and Thom have already followed up with further work.
The main visible limitation is task-orientedness, not capability—the model stops when the assigned task is done. In the one interactive result, involving spherical and binary codes, asking “can you push this further?” elicited more sophisticated representation theory and connected to the sphere-packing value. One guest takes a utilitarian view of taste: solving harder problems through better judgments is evidence of better taste; another suggests separating taste from long-horizon execution, with a supervising model directing the underlying system.
The economics of mathematics may invert: proving was the bottleneck that made understanding come “along for the ride,” and now understanding, absorbing, and communicating results may become explicitly valued scarce work. Models can generate exponentially more mathematics but also help digest it—“it’s helping solve the problem it creates”—and applied work may no longer require finding a world expert. The ceiling remains high: “it’s plausible we’ll never solve something like P versus NP.”

Deep dive

1. Two working mathematicians converted by an IMO gold and a five-minute literature search

Sellke joined after seeing the IMO gold medal result last summer: “this is amazing… I want to see what the heck they did.” In the fall, he gave his prior collaborator Sawhney a GPT-5 account, and Sawhney “very quickly became convinced” that the models were extremely exciting to play with.
Sawhney’s conversion story: browsing the site collecting Erdős’s problems in combinatorics, his chronic frustration was seeing a problem marked open without knowing if it was actually still unsolved—“the literature is often quite hard to search.” He plugged one into GPT-5 and five minutes later it found a reference, for a problem friends had already spent hours on without knowing whether it was in reach. “We found like 10 more cases sort of like this.”

2. Why the model wins: execution, correct bets, and a clean context window

Sellke on a relative strength: humans get lost in the “is epsilon smaller than delta” bookkeeping, but “the AIs just kind of always nail these kinds of arguments” once an idea is in hand.
On the unit-distance result, Sellke says a mathematician may try an idea for hours, days, or weeks, conclude it does not work, and later discover that someone else made it work. “You’re kind of gambling against the problem,” while the model, “extremely dogged” but unable to try everything, “made the correct bet” and pruned the search tree.
Sawhney adds that the idea was not foreign, suggesting serious mathematicians had tried it. He also notes that, when humans go down a wrong path, the original plan can become linked to everything that failed; with an AI, it is easy to start another session in a different direction.
Sellke says the model does make mistakes and then thinks about them. Humans tend to downgrade an approach after a failure and keep doing so over several failures; the model’s ability to reconsider and restart is part of why execution becomes less of a barrier.

3. Good mathematical reasoning is linked to general-purpose reasoning

Host Lisha Li’s puzzle is that papers and textbooks are poor training sets for how to think: “don’t learn real analysis from Rudin,” because polished artifacts hide “the struggle that made us formulate definitions in a certain way.” So where does mathematician-like thought come from?
Sellke’s answer stays general: “we’re training general-purpose reasoning models,” and behaviors like backtracking and starting again “aren’t really specific to mathematics… if you work hard at reasoning you should see these patterns eventually.”
One reason OpenAI released summarized reasoning traces for these results was to answer whether the model was simply guessing or thinking in a totally foreign way. Sawhney said the traces were “very much like reading a colleague’s notes”—collaborators spilling thoughts into an email, chained together. “It’s quite surprising the first few times.”

4. Sphere packing: Astra resolves the LP bound with an exact asymptotic

The sphere-packing discussion illustrates how little is known. The transcript identifies five dimensions with known optimal answers: d=1 trivially, d=2 with the hexagonal lattice, d=3 with Hales’s grocery-store-orange result from the 2000s, and d=8 and 24 through the E8 and Leech lattices. The d=2 proof was only obtained in the 1960s; the d=3 proof is a few hundred pages and “famously ugly.” Viazovska’s related work in dimensions 8 and 24 was a large part of the reason she won a Fields Medal in 2022.
The best intuition offered for the planar answer was simply that bees use a honeycomb: “if there was a more efficient way then probably bees would pack honeycomb some other way.” In high dimensions, an easy lower bound is 2^-d: a saturated packing, with every sphere blown up by a factor of two, must cover all space. The discussion emphasizes that even placing spheres at cube vertices leaves vanishingly little covered volume and can leave room for another sphere.
For a long time, the best upper bound was roughly 2^(-0.599d), from Kabatiansky and Levenshtein in the 1970s. Their paper was only one or two pages and notably terse.
The Cohn–Elkies linear-programming method optimizes over functions satisfying two conditions: f(x) ≤ 0 when ||x|| > 1, and a nonnegative Fourier transform. Astra showed that the best possible bound within this framework has an asymptotic form essentially like (e^(-c+o(1)))^d, roughly 2^(-0.6d), and proved that no admissible function does better. This is an exact result about the LP bound, not a claim that the actual sphere-packing density itself has been determined in all high dimensions. It explains a value Cohn and collaborators had previously conjectured from numerical work.
One guest had thought about the problem for about six months as a graduate student and made “absolutely zero progress.” The part showing that the LP bound cannot be improved is only a few pages of complex analysis: “Once you see it… it’s unbelievable—why hadn’t somebody done this before?”

5. Codes: the one interactive result, and what “push this further” exposed

Spherical codes are sphere packings on the surface of a sphere; binary codes are the analogue on a hypercube. They formalize error correction by requiring large Hamming distance between allowable code words. For binary codes, this is a practical information-theoretic question: with a 1% error rate, how much longer must an n-bit message be to tolerate the errors?
The models found better bounds for these cases through representation theory rather than the complex analysis used in the full-space LP analysis, leaning heavily into the symmetry of the sphere and cube.
This was the only problem pair with interactivity. Asked to improve the code bounds by an exponential factor, the model used some representation theory. Asked, “Hey, can you push this further?”, it returned much more sophisticated representation theory; in the small-sphere limit, that recovered part of the full-space sphere-packing result and its value. The researchers then asked it to analyze the full-space case directly and complete the picture.
The diagnosis matters for extrapolation: “it wasn’t a capabilities issue.” The model had been asked to improve the bounds, did that task, and stopped. “Models tend to be pretty task oriented. If you tell it to do a task and it accomplishes the task, it’s pretty happy.”

6. The taste debate: utilitarian metrics and a supervisor/underlying-system split

One guest takes a utilitarian view: taste is “if you’re able to solve problems faster by making better judgments,” so solving harder problems “kind of by definition means it has better taste.” The task-oriented symptom—making a breakthrough, recognizing it, and not pushing to the limit because that was not requested—“seems rather minor compared to the state of progress.”
Another guest suggests that a concrete long-horizon task and taste may be better separated: one model could be responsible for taste while another works for a long time on the problem, as the “underlying system of the supervising AI.”
The discussion’s human analogy is that a friend looking over your shoulder and asking, “What are you doing?” can provide a useful ten-second reset. The guests saw no reason models should be fundamentally different from humans here; several agents or collaborators may be more powerful than one.

7. A non-sofic group exists: roughly 15 pages where humans needed 450 and quantum complexity

Sofic groups are, roughly, groups that can be approximated by finite groups. The integers and the integers mod n provide the canonical picture: locally, a large finite cycle looks like the infinite line. Soficity implies surjunctivity, and the hope that every group was sofic was motivated by the way finite-group proofs can sometimes be transferred through such approximations, tracing back to Gottschalk’s question from the 1970s.
The stronger Aldous–Lyons conjecture was disproved about two years ago in a tour de force—“like 250 pages building on another 200 pages”—using quantum complexity theory and a complicated bridge that not many people could understand. Astra’s direct proof that a non-sofic group exists is perhaps 15 pages, stays in group theory, and builds on results by Kun and Thom. It is “a very, very delicate combinatorics argument” that adds one algebraic fact to rule out the “weird conspiracy” prior work could not exclude.
One guest’s genuine update was that, a year earlier, they would have expected AI proofs to become thousand-page objects that nobody could understand. Instead, “all of these AI proofs are very short and elegant.” “Only humans can generate 200-page proofs right now.”
Kun and Thom have already followed up with further examples of non-sofic groups. “We don’t want to be writing lots of follow-up papers ourselves.”

8. Math’s new economics: understanding becomes the explicitly valued scarce work

Sawhney on the absorption side: he now drops arXiv PDFs into his favorite model for a rough proof strategy rather than reading only the introduction. Models “are going to help us produce exponentially more mathematics, but they also make it much easier to absorb it… it’s helping solve the problem it creates.”
Sawhney’s structural point is that proving a result used to be so hard that understanding, maintaining, and explaining it came along for the ride: if you proved it yourself, you would automatically understand it quite well. With that bottleneck relaxed, “the optimal structuring for organizing knowledge could look rather different.”
Sawhney predicts that communal understanding—helping other humans put mathematics into a framework and appreciate it—will become “an increasingly more explicit and valuable part of the subject.”
Sellke notes that the ceiling for mathematical difficulty is high, so “it’s plausible we’ll never solve something like P versus NP.” The field may become more attached to large mysteries than to smaller routine ones.
The guests also expect broader access: someone doing work that requires mathematics may no longer need to find a world expert on a topic. One guest’s closer was, “it’d be nice for the world if applied mathematics went a lot faster.”

AI:AM Highlights: Welcome to the AGI Era