OpenAI mathematical reasoning breakthrough is ten hard problems solved by Astra, walked through on a16z by mathematicians Mark Selki and Matab Swani after they joined OpenAI following GPT-5's IMO gold.
Nothing here is an exploit. No system is affected, no remote or local attacker is required, and no patch exists. The demo is pure mathematics.
How the attack actually works
Think of Astra as a dogged colleague who never tires of pruning a search tree. It tries an idea, backtracks without polluted context, and nails epsilon delta details where humans get lost.
For sphere packing it constructed the optimal Cohn Elkies linear programming function and proved the bound tight at e to the 2 pi plus 0.01 to the d, about 2 to the minus 0.61 d, beating the 1970s Kabatiansky Levenshtein 2 to the minus 0.599 d constant, then reused representation theory for spherical and binary codes.
Why this matters and what is not fixed
OpenAI trains general purpose reasoning, not Lean autoformalization, so the same trace that solves packing can also explain a PDF you upload. That helps Astra adoption but leaves judgement as the gap: Astra improved codes then stopped until prompted to push further.
P versus NP and other deep mysteries still look out of reach, so verification and taste become the new bottleneck, not derivation.
Frequently Asked Questions
What is the sphere packing problem Astra solved?
It asks how densely unit spheres can fill d-dimensional space. Exact answers are known only in dimensions 1, 2, 3, 8 and 24, so Astra's result settles the asymptotic value of the Cohn Elkies linear programming bound for large d.
What is a sofic group?
A countable group is sofic if it can be approximated by finite groups. Astra gave a short group theory proof that a non-sofic group exists, a stronger and simpler result than the recent 250 page disproof of the Aldous Lyons conjecture.
Did Astra just search or guess?
No. The released summarized chains of thought read like a human collaborator's notes, making choices, discarding bad paths and updating correctly. The strength is persistence plus good pruning, not exhaustive brute force.
