OpenAI has introduced the Astra model, which has solved 10 open problems in mathematics and computer science.

OpenAI has officially confirmed the existence of a new model, Astra. According to the report, an internal version of this model helped solve 10 complex mathematical problems that scientists had been unable to solve for decades.

OpenAI describes Astra as its next major family of models. The problems solved by the model span various branches of mathematics and computer science, ranging from geometry and cryptography to group theory and quantum computing. Progress had stalled on each of these problems for at least 10 years, and in some cases, much longer.

One of the key results is that Astra constructed a mathematical object whose existence had been discussed by scientists for many years but could not be proven. This resolves a long-standing question in group theory. The model also disproved a well-known hypothesis about mathematical groups and improved several important estimates in coding theory and computational complexity.

According to OpenAI's calculations, solving all 10 problems would have cost approximately $2,000 if using the Sol model API at standard rates. Following this, researchers collaborated with Astra to draft scientific papers based on the model's reasoning. The model itself was asked to verify each proof using a specialized program called Lean, which ensures the correctness of proofs without manual human verification. OpenAI also provided step-by-step records of how the model reasoned through each problem.

The company emphasized that while humans are responsible for the final text of the publications, the mathematical ideas and the course of the proofs were generated by the model. Earlier, in May, OpenAI had already demonstrated a similar case, when an unannounced model disproved a known mathematical hypothesis, and this work has already inspired scientists to write new papers.

Thomas Bloom, a mathematician at the University of Manchester, called the new results a significant event for the scientific community. He does not believe that artificial intelligence will soon replace mathematicians, as such models are trained on works and scientific texts created by humans.

Previously, OpenAI had stated that it was working on a new family of models for complex and lengthy tasks. This week, OpenAI CEO Sam Altman presented Astra to U.S. government officials and regulators. He explained that the model can coordinate the work of multiple AI agents when solving scientific and engineering problems. According to available information, Astra will be one of the models alongside Sol, Terra, and Luna, but it is not yet known under what name it will be released to the market.

Source: OpenAI