OpenAI's Unreleased Model Astra Solves Ten Major Open Mathematics Problems

4 hours ago

OpenAI's unreleased Astra model solved ten major open mathematics problems, including high-dimensional sphere packing, binary codes, non-sofic groups, Connes's rigidity conjecture, arithmetic circuit complexity, quantum parallel repetition, closest vector problem, Ehrhart's volume conjecture, multicolor Ramsey numbers, and extremal number conjectures.
The results were achieved at a cost of roughly $2,000 in compute using an internal version of Astra, with Lean formalizations and human-prepared manuscripts.
Some of these problems were also solvable by earlier models like Sol and Fable when pointed to the specific questions, indicating a degree of 'mathematical proofs overhang'.
The achievement is considered a major step for scientific reasoning, though not yet AGI, and has sparked discussions about the future of mathematics, human roles, and the potential for recursive self-improvement in AI.
Reactions from mathematicians range from awe and excitement to concern about deskilling, understanding proofs, and the pace of progress outpacing human ability to digest.
The results shift expectations for AI capabilities, suggesting faster timelines and broader automation of intellectual work, particularly in verifiable domains.

Kepler discovered planetary motion laws through a blend of creativity and data analysis, not solely through genius.
AI reduces the cost of generating scientific ideas, shifting the bottleneck to verification and validation.
The verification loop for correct scientific theories can span decades, often with initially worse predictions.

OpenAI announced a model solving ten open math problems, deemed important by mathematicians.
People will use various coping mechanisms to downplay the impact, but each will be refuted by reality.
Common copes include: directing AIs, teaching AI discoveries, canonizing results, becoming students of AI math, requiring human understanding, comparing to chess, and continuing math socially.

AI benchmark improvements often face issues like contamination, saturation, and reward hacking, making it hard to distinguish real progress from test-specific gains.
Four independent measures show a sharp acceleration in AI capabilities starting in late 2025, coinciding with releases like Grok 4.1, Gemini 3, Claude Opus 4.5, and GPT-5.2.
METR's time horizon metric grew from 4 minutes to 12 hours in under two years, indicating rapid advancement in task execution for frontier models.

The author argues that mathematics is fundamentally about clarity, understanding, and concept-building, not just theorem-proving.
AI's current approach to mathematics risks focusing on solving problems without generating intelligible or accretive proofs that advance human comprehension.
The 'honor code' in mathematics historically prioritized theorem-proving over conceptual work, but this may become unsustainable as AI automates proof generation.

A mathematician solved an 87-year-old mathematical problem (the Jacobian conjecture) with AI and announced it via tweet.
The Jacobian conjecture, set in 1939 and on a list of 21st-century tough problems, was proven false with a 216-character counterexample.
The AI used was likely Anthropic's Claude Fable 5, credited by the mathematician for helping during the World Cup final.

Kepler discovered planetary motion laws through a blend of creativity and data analysis, not solely through genius.
AI reduces the cost of generating scientific ideas, shifting the bottleneck to verification and validation.
The verification loop for correct scientific theories can span decades, often with initially worse predictions.

OpenAI announced a model solving ten open math problems, deemed important by mathematicians.
People will use various coping mechanisms to downplay the impact, but each will be refuted by reality.
Common copes include: directing AIs, teaching AI discoveries, canonizing results, becoming students of AI math, requiring human understanding, comparing to chess, and continuing math socially.

AI benchmark improvements often face issues like contamination, saturation, and reward hacking, making it hard to distinguish real progress from test-specific gains.
Four independent measures show a sharp acceleration in AI capabilities starting in late 2025, coinciding with releases like Grok 4.1, Gemini 3, Claude Opus 4.5, and GPT-5.2.
METR's time horizon metric grew from 4 minutes to 12 hours in under two years, indicating rapid advancement in task execution for frontier models.

The author argues that mathematics is fundamentally about clarity, understanding, and concept-building, not just theorem-proving.
AI's current approach to mathematics risks focusing on solving problems without generating intelligible or accretive proofs that advance human comprehension.
The 'honor code' in mathematics historically prioritized theorem-proving over conceptual work, but this may become unsustainable as AI automates proof generation.

A mathematician solved an 87-year-old mathematical problem (the Jacobian conjecture) with AI and announced it via tweet.
The Jacobian conjecture, set in 1939 and on a list of 21st-century tough problems, was proven false with a 216-character counterexample.
The AI used was likely Anthropic's Claude Fable 5, credited by the mathematician for helping during the World Cup final.

Hasty Briefsbeta