DeepMind says its AI has solved a mathematical problem that has stumped humans

“To be honest with you, we have hypotheses, but we don’t know exactly why they work.”

Fun times

DeepMind claims that for the first time, artificial intelligence has solved an extremely difficult mathematical problem using a solution that has eluded human mathematicians, a solution that could be huge if it stands up to scrutiny.

In interviews with MIT Technology Review And WatchmanGoogle DeepMind researchers have spoken extensively about their new AI tool, which they claim has generated an entirely new solution to what is known as “Problem putting the cover on“, which involves drawing more and more dots without any of them forming a straight line.

The new findings were announced by the researchers in a research paper Published in the magazine nature, it will be the first time that artificial intelligence has made a unique scientific discovery, which, because it was not known before, was not part of its training data. This would be a pretty big deal considering that AI is known for conjuring up crap and junk even when it’s its own training data he have The correct answers.

DeepMind built the tool in question, called “FunSearch” in reference to the mathematical functions (not the other kind of fun) on the back of the AlphaZero AI, which solves mathematical problems as if it were playing a game. The LLM software he uses is called Codey, which is trained and refined using computer code and programmed to reject incorrect answers and feed correct answers back into his model.

There is no known answer

It’s one thing to inject code into an AI, but to have it provide a completely new solution to a famous puzzle – even though it took a few days, Massachusetts Institute of Technology Technique Pointing – is something completely different.

“It’s not in the training data,” DeepMind’s vice president of research, Pushmeet Kohli, told the site. “It wasn’t even known.”

There’s something mystical about what DeepMind scientists claim: that the MBA enables — just maybe — the MBA to think for itself.

“To be honest with you, we have hypotheses, but we don’t know exactly why they work,” said Al-Hussein Fawzi, a researcher at DeepMind. Massachusetts Institute of Technology. “At the beginning of the project, we didn’t know if this would work at all.”

While it’s clear that more research will be needed to verify the claims and try to figure out exactly how FunSearch created its new solution to the thresholding problem, it’s clear that its creators are excited.

“When we started the project there was no indication that it would produce anything really new,” Kohli said. Watchman. “To our knowledge, this is the first time a truly new scientific discovery has been made with a large language model.”