Inside OpenAI’s Breakthroughs in Mathematical Reasoning
a16z · 2026-09-08 · 65 min
https://www.youtube.com/watch?v=1JvyLGd2SfsVideo summary
OpenAI mathematicians explain how Astra improved sphere-packing bounds and proved non-sofic groups exist.
a16z Infra Partner Lisha Li speaks with OpenAI mathematicians Mehtaab Sawhney and Mark Sellke about AI’s recent mathematical results and why they reflect more than brute-force search. They describe models pursuing promising ideas through finicky details, backtracking after mistakes, and producing reasoning traces that resemble a mathematician’s notes. In sphere packing, Astra analyzed the linear-programming bound and established its high-dimensional asymptotic behavior, improving known bounds; related work also strengthened bounds for spherical and binary codes using representation theory. Astra also produced a proof that non-sofic groups exist, a result previously approached through a much longer, more technically complex conjecture. The speakers discuss how researchers can build on and absorb AI-generated mathematics, while emphasizing that explaining results and fitting them into human understanding may become more important. They expect AI to accelerate applied mathematics, though problems such as P versus NP may remain beyond reach.
Chapters
- 0:00From Practicing Mathematician to OpenAI: Mark Sellke Joined After the IMO Gold Medal, and Mehtaab Sawhney Tried GPT-5
- 2:43Why GPT-5 Was the Conversion Moment: It Found an Erdős-Problem Reference in Five Minutes
- 4:21Beyond Search & Connections: Unit-Distance Proofs Show Persistence, Mathematical Judgment, and Easier Restarts
- 9:51Reasoning Traces: Models Revisit Mistakes, Reassess Paths, and Can Use Parallel Sessions
- 11:44Why Math Papers Are a Bad Training Set for Real Mathematics: Reasoning Traces Resemble Colleagues’ Notes
- 16:20The Astra 10-Problem Set: Mehtaab’s Sphere-Packing Favorite, from the Hexagonal Lattice to Hales’s Proof
- 21:16The Astra 10-Problem Set: Astra’s Sphere-Packing LP Bound Reaches Roughly 2 to the Minus 0.61d
- 26:18Astra Problem 1: The Model Characterizes the High-Dimensional Linear-Programming Bound
- 31:38Astra Problems 2–3: Better Code Bounds from Representation Theory and Follow-Up Prompts
- 36:17The Harness vs. the Model: Prompts Steered Astra’s Code-Bound Improvements
- 40:01What Is Taste in a Model? Faster Problem-Solving, Judgment, and a Second Perspective
- 43:41What Is Taste in a Model? From Problem Selection to Astra’s Non-Sofic Group Result
- 48:57What Is Taste in a Model? Integers Modulo n, Aldous–Lyons, and a Shorter Group-Theory Proof
- 53:51What Is Taste in a Model? Astra’s Delicate Proof and Follow-Up Mathematics
- 57:32How Should the Math Community Adopt AI? Mehtaab Uses Models to Understand Proofs Faster
- 1:00:01Empirical vs Theoretical Math & the Positive Vision: AI Could Shift Focus to P versus NP and Applied Math
What the full analysis includes
16 chapters · 89 key points · 16 section summaries · mind map
Preview of the first chapter · From Practicing Mathematician to OpenAI: Mark Sellke Joined After the IMO Gold Medal, and Mehtaab Sawhney Tried GPT-5
The remaining 86 key points, section summaries and mind map are in the full analysis. Analyze this video to see everything →
This video is 1.1 hours. The Free plan analyzes videos up to 30 min; plans from Plus cover videos up to 2 hours.
This is a Tier 1 public summary
Whether the chapter key points, section summaries and mind map are public is up to the person who shared it. Want the full analysis?Submit one yourself.
More from this channel
Hip Hop’s Origin Story, Told by the Legends Who Lived Ita16zKurtis Blow recalls building hip-hop from nothing in 1970s New York and mentoring Run before Run-DMC rose to fame.
What iPhone Packaging Reveals About the Robotics Opportunitya16zTutor Intelligence CEO Josh Gruenstein says its 100-robot Boston fleet trains AI while robots tackle factory work for about $16 an hour.
Why AI Could Turn Small Law Firms Into Billion-Dollar Businessesa16zEve CEO Jay Madheswaran says AI helped an LA law firm grow from 100 to 500 clients per attorney by automating plaintiff-law work.
Why Founders Are Betting Big on Factories, Mines, and Energya16za16z investors argue defense demand can rebuild U.S. factories, mines, and energy while creating commercial markets for robotics.
Mark Cuban on Why AI Could Turn Solo Founders Into Trillionairesa16zMark Cuban says AI could expose healthcare contract abuses, while Cost Plus Drugs bets $20 million manufacturing pods can lower drug costs.
Related analyses
OpenAI President on What it Means to Cross Into The Age of AGIa16zOpenAI President Greg Brockman calls Astra AGI after 24-hour tasks and urges urgent AI safety and cyber defense.
Can Open Models Keep Up With OpenAI and Anthropic? | Nathan Lamberta16zNathan Lambert says China matches U.S. AI talent but lacks compute, while open models trail frontier labs on different timelines.
Why OpenAI Is Racing to Make Computer Use Superhumana16zOpenAI’s Ari Weinstein and James Sun say computer-use agents are nearing superhuman speed, automating procurement, finance, and legacy software workflows.
AI Companies Are Setting Money on Fire | OpenAI's Pre-Bailout Bailout PlanGamers NexusOpenAI lost $38.5 billion in 2025 while pursuing a $280 billion revenue target and reportedly offering the U.S. government a 5% stake.
HW News - NVIDIA Employee Detained, CXMT RAM, AMD Has Lost Its Damn MindGamers NexusAMD’s $280 RX 9050 pairs 8GB with a power cable, while CXMT claims 7.67% of global DRAM share.