OpenAI Astra: Ten Advances in Mathematics and Theoretical Computer Science

OpenAI Astra Solves Ten Long-Standing Mathematical Problems

OpenAI has announced that an internal version of its next major model, Astra, has successfully solved ten open problems in mathematics and theoretical computer science. These problems had seen no progress on their main results for at least a decade, and in many cases, significantly longer. To ensure correctness, OpenAI has released Lean certificates for each proof via a public GitHub repository.

Summary of Mathematical Breakthroughs

The results span several complex domains, including high-dimensional geometry, group theory, and quantum complexity. The ten solved problems are:

  • High-dimensional sphere packing: Established new upper bounds on sphere-packing density down to the Cohn–Elkies threshold.
  • Binary and spherical codes: Produced exponentially improved bounds on the maximum size of binary codes at any prescribed minimum distance, with analogous results for high-dimensional spherical codes.
  • Non-sofic groups: Provided a construction establishing the existence of non-sofic groups, resolving a central open question in group theory.
  • Connes’s rigidity conjecture: Disproved the conjecture that certain groups are uniquely determined by their von Neumann algebras.
  • Arithmetic circuit complexity: Established new lower bounds for computing the permanent using arithmetic circuits and formulas, including an arithmetic-formula lower bound of order $n^{4/\log n}$.
  • Quantum parallel repetition: Developed an exponential parallel repetition theorem for general two-player quantum games, extending classical complexity theory principles.
  • Closest vector problem: Found polynomial-factor hardness of approximation for the closest vector problem, a key lattice question for post-quantum cryptography.
  • Ehrhart’s volume conjecture: Determined the maximum possible volume of a convex body whose centroid is its only interior lattice point across every dimension.
  • Multicolor Ramsey numbers: Resolved Erdős problem 183 by providing a superexponential lower bound for multicolor triangle Ramsey numbers.
  • Extremal number conjectures: Resolved Erdős problems 146 and 180 regarding compactness and degeneracy conjectures in extremal graph theory.

Methodology and Resource Consumption

The solutions were generated by an internal version of the Astra model. OpenAI reports that the total token cost to find these solutions would be approximately $2,000 at current Sol API rates.

Following the AI's discovery of the mathematical arguments, human researchers worked with the model to prepare the results into manuscripts. Finally, the model formalized each argument into a Lean certificate to provide a machine-verifiable proof of correctness. OpenAI has also released narrations of the model's thinking process for each solution.

Responsibility and Attribution in AI Mathematics

OpenAI asserts that attribution should honestly reflect the origin of a result, stating that claiming human authorship for AI-generated proofs would misrepresent the nature of intellectual work. While humans assisted in manuscript preparation and Lean formalization, OpenAI attributes the core mathematical arguments to the Astra system.

Community Analysis and Counterpoints

The announcement has sparked significant debate within the technical community regarding the nature of AI-driven discovery and the transparency of the process.

Transparency and Experimental Rigor

Some critics argue that the reported cost of $2,000 is potentially misleading without full transparency of the experimental setup.

"I want to know: 1. How many total problems were given to the model, and what percent were left unsolved at what cost before giving up? 2. How many attempts did you give the model at solving these problems?"

Heuristic Search vs. Novel Theory

There is ongoing discussion about whether the model is generating genuinely novel mathematical theories or simply performing an exhaustive search of existing tools to find the right combination for a known problem.

"Do these proofs contribute new ideas to the mathematical corpus, or are they simply an effective method to exhaustively search the literature for the right combination of existing tools to apply to the problem?"

Impact on the Mathematical Profession

While some view these results as a threat to the profession, others suggest AI will act as a catalyst, opening new avenues for human researchers.

"For every conjecture defeated some seven or eight new ideas open up. Our path through that combination will be set by creative and curious human mathematicians."

Technical Observations

Observers noted that some of the resulting Lean proofs are exceptionally long—reaching up to 50,000 lines—highlighting the difference between AI-generated formalization and concise human-written proofs.

Sources