Google DeepMind claims a groundbreaking achievement, asserting that its new AI tool, named “FunSearch,” has successfully solved a notoriously complex mathematical problem that had previously stumped human mathematicians. This problem, known as the “cap set problem,” involves the challenge of plotting an increasing number of dots without any of them forming a straight line.
According to DeepMind researchers, this represents a milestone as it marks the first instance where an AI has made a unique scientific discovery not present in its training data. The findings, detailed in a paper published in the journal Nature, demonstrate the AI’s ability to generate a completely novel solution to a problem that was previously unknown.
FunSearch, the AI tool in question, was built upon the foundation of DeepMind’s AlphaZero AI, which approaches mathematical problem-solving akin to playing a game. The underlying language model, named Codey, is trained on computer code and programmed to reject incorrect answers while incorporating correct ones into its model.
The key distinction in this achievement lies in the fact that the AI produced a solution that was not part of its training data, challenging the common occurrence of AI generating nonsensical or fabricated outputs even when trained on accurate information.
“It’s not in the training data,” DeepMind research VP Pushmeet Kohli told the website. “It wasn’t even known.”
The researchers expressed a sense of mystery regarding the AI’s capabilities, suggesting that the AI might have, to some extent, engaged in independent thinking. Despite having hypotheses, the team acknowledges uncertainty about the exact mechanisms behind the AI’s success, emphasizing that the outcome exceeded their initial expectations.
“To be very honest with you, we have hypotheses, but we don’t know exactly why this works,” DeepMind researcher scientist Alhussein Fawzi told MIT Tech. “In the beginning of the project, we didn’t know whether this would work at all.”
The novelty of the AI’s solution to the cap set problem introduces a mystical element, with the researchers acknowledging their lack of complete understanding of why the AI produced such results. While further research is necessary to validate these claims and comprehend the underlying processes of FunSearch, the creators are enthusiastic about what appears to be the first instance of a large language model making a genuine, new scientific discovery.
“When we started the project there was no indication that it would produce something that’s genuinely new,” Kohli told The Guardian. “As far as we know, this is the first time that a genuine, new scientific discovery has been made by a large language model.”