The company indicated that Astra produced the mathematical proofs, which were subsequently formed into research papers by humans with aid from the model and formalized using Lean, a software system designed to verify mathematical proofs. OpenAI also provided the model’s reasoning narratives along with each proof.
This announcement comes after OpenAI’s previous revelation in May, where another AI-produced result reportedly disproved the Erdős unit-distance conjecture during the evaluation of an unreleased model.
Findings encompass cryptography, quantum computing, and group theory
As stated by OpenAI, Astra yielded new insights on high-dimensional sphere packing, binary and spherical codes, arithmetic circuit complexity, the closest vector problem, quantum parallel repetition, and Ehrhart’s volume conjecture.
The model also demonstrated the existence of non-sofic groups, provided a disproof of Connes’s rigidity conjecture, and resolved three open problems proposed by mathematician Paul Erdős in extremal graph theory and Ramsey theory.
Also read: Why AI firms are acquiring old books by the million
OpenAI remarked that generating these solutions would have cost around $2,000 based on its Sol application programming interface (API) pricing. Additionally, they noted that the accompanying manuscripts were prepared by researchers utilizing the same model before the proofs were formalized in Lean.
Along with this announcement, the company introduced ChatGPT for Academic Researchers, an initiative aimed at providing 100,000 scientists and mathematicians with complimentary access to its latest ChatGPT models.
Researchers advise caution regarding broader AI capabilities
While this announcement underscores advancements in the application of AI in mathematical research, some researchers caution that it should not be seen as proof that general reliability issues with generative AI have been resolved.
“Astra is (apparently—we still haven’t seen the methodology) quite proficient in math, or at least certain areas of math, but that does not imply it will be immune to hallucinations or overcome the reliability challenges faced by other generative AI systems. It doesn’t mean it will correctly interpret a PDF either. It also doesn’t guarantee that it will adhere to strict rules,” cognitive scientist and AI researcher Gary Marcus stated in a post on X.
OpenAI has yet to release Astra to the public, and the results are anticipated to be evaluated by the broader mathematics community through peer review and independent verification.
Also Read: Alphabet scales back AI image generation in Google Earth due to policy violations
