OpenAI Researchers on the Future of Mathematical Reasoning

What It Covers

OpenAI mathematicians Mehtab Swani and Mark Selke, interviewed by a16z's Lisha Li, explain how AI system Astra solved 10 open mathematical problems across sphere packing, coding theory, and group theory — producing short, elegant proofs that read like expert human reasoning, and reshaping what mathematical practice looks like.

Key Questions Answered

• •AI mathematical execution advantage: Models eliminate the human cost of executing finicky proof details. Where mathematicians abandon approaches after weeks of failed attempts, AI pursues every path without fatigue or discouragement. The unit distance conjecture proof exemplifies this — the core idea existed previously, but executing the extraordinarily detailed reasoning required was the actual barrier, one AI cleared without the risk-reward calculation humans make.
• •Backtracking without context pollution: A structural advantage AI holds over human mathematicians is the ability to restart problem-solving without carrying failed-path bias. When humans pursue a wrong approach, prior reasoning contaminates subsequent thinking. AI can spawn parallel sessions from a clean state, effectively running multiple independent mathematicians simultaneously — each unaware of the others' dead ends, preserving fresh judgment throughout.
• •Linear programming bound for sphere packing: Astra proved the asymptotic behavior of the LP (linear programming) bound for sphere packing in high dimensions equals d/(2πe)^(1+o(1)) per dimension — matching a numerics-based conjecture by Cohen et al. that had no theoretical explanation. Crucially, the model proved both that this bound is achievable and that no LP-based function can do better, fully resolving the high-dimensional picture.
• •Representation theory in coding problems: For spherical and binary error-correcting codes, Astra improved exponential bounds by leveraging representation theory more aggressively than prior methods. The approach exploited the algebraic symmetry of spheres and hypercubes at a deeper level. Notably, this was the one interactive result — researchers prompted the model to push further after an initial improvement, revealing that task-oriented models benefit from explicit directional nudges.
• •Non-sofic group existence proof: Astra produced a roughly 15-page proof that non-sofic groups exist — disproving the conjecture that every countably infinite group can be approximated by finite groups. This contrasts with the prior disproof of the stronger Aldous-Lyons conjecture, which required 250 pages plus 200 pages of prerequisites and quantum complexity theory. The AI proof stays within group theory using a delicate combinatorial argument.
• •Mathematical bottleneck shift: When proof generation was the primary bottleneck, understanding and communication came automatically — the prover inherently understood the result. As AI removes the proof bottleneck, the scarce resource becomes human comprehension, synthesis, and contextual placement of results. Mathematicians should reorient toward absorbing, explaining, and connecting AI-generated proofs, using models to rapidly parse papers and identify where results fit existing knowledge structures.