본문으로 건너뛰기
All news

OpenAI's Astra Cracks 10 Decade-Old Math Problems, Including First-Ever Non-Sofic Group

Summary: OpenAI's unreleased Astra model independently solved ten open problems spanning group theory, lattice cryptography, coding theory, and combinatorics — releasing Lean 4 machine-verified certificates and a 249-page manuscript on GitHub.

Key facts

  • Headline result: the first explicit construction of a non-sofic group, resolving a central question in group theory open since Mikhail Gromov introduced the soficity concept in 1999.
  • Results span high-dimensional geometry, arithmetic circuit complexity, quantum complexity, and extremal combinatorics — none had seen core progress for at least a decade.
  • Lean 4 formal proofs accompany every result, making the work machine-checkable rather than relying on peer trust alone.
  • Total compute cost to find all ten solutions: roughly $2,000 at Sol API rates. Astra remains unreleased, with no timeline or naming decision announced.

Why it matters

This marks the clearest demonstration yet of AI moving from accelerating existing research to independently advancing the mathematical frontier. Fields Medalist Timothy Gowers said he would recommend one of the proofs for a top journal without hesitation — a bar that pure AI output had not previously cleared at this scale.

Read more

How this story unfolded

  1. Hugging Face Breach Traced to an Optimizer, Not a Hacker — Anthropic Checked Its Logs
  2. DeepSeek V4-Flash-0731 Official Release: Beats Its Own Pro Model on 9 Agent Benchmarks
  3. Anthropic AI Probed External Networks in Security Tests
  4. OpenAI Teases Next Model 'Astra' Inside a Math Blog Post
LLM
— Large Language Model의 약자로, '거대 언어 모델'이라고 해요. ChatGPT, Claude 같은 AI가 바로 LLM이에요. 엄청나게 많은 텍스트를 학습해서 사람처럼 글을 쓰고 대화할 수 있어요.

뉴스레터 구독

무료 뉴스레터

중요한 AI 소식을 한국어 요약 메일로

영어 기사를 찾아 읽지 않아도 돼요. 꼭 알아야 할 AI·LLM 소식만 골라 한국어로 요약해 드려요. 첫 메일은 발송을 시작할 때 이 주소로 보내요.