OpenAI: AI model disproved 80-year-old math problem

OpenAI has shared progress for AI in mathematics. The biggest result is that, according to the company, an internal AI model has disproved a roughly 80-year-old puzzle in geometry, namely Erdős' unit distance conjecture.

The headline of the news item we were given was brief. We have therefore stuck to what can be documented in the sources.

What happened?

A conjecture is a claim that mathematicians believe but have not yet been able to prove. The Hungarian mathematician Paul Erdős posed the question of how many pairs of points in a plane can at most be exactly distance 1 from each other.

OpenAI's model found sets of points that have more such pairs than the conjecture allowed. The conjecture is thereby disproved. According to one review, the model arrived at this by combining geometry with deep results from number theory.

Confirmed, but not a revolution

Several leading mathematicians, including Fields Medal winner Tim Gowers, have confirmed that this is real progress. Human mathematicians have since improved and extended the proof.

At the same time, there are caveats. The news site Resultsense writes that the proof relied broadly on earlier work and combined known ideas from several fields instead of inventing entirely new methods. It calls it a milestone, but not a big leap.

What does it mean for you?

You do not get a new tool yourself out of this news. But it shows that AI can now contribute to research and not just solve problems humans already know the answer to.

The pattern, at the same time, is that humans still need to check and improve the results. AI works here as a helper with an enormous overview of earlier knowledge and great patience, while human judgment remains decisive.

OpenAI has previously highlighted that the company's experimental model reached gold medal level at the International Mathematical Olympiad, where it solved five out of six problems.