OpenAI says Astra delivered 10 AI research results for about $2,000 in compute

OpenAI said the internal reasoning model produced 10 verified results in mathematics and theoretical computer science after earlier disproving the Erdos planar unit distance conjecture.

Summary

OpenAI said its next-generation AI model Astra produced 10 research results across mathematics and theoretical computer science while using about $2,000 of compute at Sol API pricing. The company said the results covered high-dimensional sphere packing, coding theory, non-sofic groups, Connes rigidity, quantum parallel repetition, and post-quantum cryptography. The August 1 announcement followed Astra's May 20 result on the Erdos planar unit distance conjecture, where the model constructed configurations with at least n^(1+delta) unit-distance pairs for infinitely many n and later refined delta to 0.014. OpenAI said external mathematicians verified the new results, and Tim Gowers called the work "a milestone in AI mathematics." Researchers later turned the work into papers with verifiable Lean proof certificates, a formal verification system that can check mathematical proofs.

Terms & Concepts
  • post-quantum cryptography: Encryption methods designed to resist attacks from quantum computers.
  • Lean proof certificates: Machine-checkable formal proofs verified with the Lean theorem prover.
  • quantum parallel repetition: A concept in theoretical computer science that studies how repeating quantum games affects error probabilities.