OpenAI teases Astra, its next major AI model, after it solves 10 long-standing math problems
Key Points:
- OpenAI unveiled Astra, an unreleased AI model designed to solve complex, long-running mathematical and theoretical computer science problems, achieving ten significant advances in areas untouched for over a decade.
- Astra's internal research covered diverse fields such as high-dimensional geometry, coding theory, quantum complexity, and lattice cryptography, producing breakthroughs like disproving Connes’s rigidity conjecture and resolving problems posed by Paul Erdős.
- The model formalized its mathematical proofs using the Lean verification system, allowing for rigorous checking of arguments prepared by human researchers, with computational costs estimated at around $2,000 in tokens.
- Astra is expected to enable AI agents to collaborate on different parts of large problems and may launch as GPT-5.7, GPT-6, or under a different name, with potential restrictions similar to Anthropic's tiered release policies.
- Independent sources confirm Astra's development, highlighting it as a major AI breakthrough aimed at advancing scientific research through long-duration computational workloads.