AI solved a math problem that humans could not: it has probably started to "think for itself"

For the first time in history, artificial intelligence has solved a very complex mathematical problem, obtaining a solution that human mathematicians could not think of. If a thorough check reveals that AI has found the right answer, this could be a new step towards a revolution.

The announcement of the AI breakthrough was made by Google DeepMind researchers. The results were published in the journal Nature.

The researchers said that it is about solving the so-called "cap set problem". The task is to draw the maximum possible number of points in such a way that they do not create a straight line in any case.

A tool called FunSearch (we are talking about mathematical functions, not entertainment) managed to cope with the task. This tool was built on the basis of AlphaZero AI, which solves math problems as if it were playing a game. This AI uses a large language model (LLM) called Codey. This LLM is trained on computer code and is programmed to discard incorrect answers and insert the correct ones back into its model.

But when an AI copes with a code, it is no longer a sensation, but the ability to get a completely new solution to a known puzzle - even if it took several days to find this solution - which is something else.

As explained by Pushmeet Kohli, Vice President of Research at DeepMind, the sensation is that AI managed to find a solution that was not only "not in the training data" but also "not even known to exist."

Scientists suggest that AI has probably reached a new level and managed to "think for itself."

"To be honest, we have hypotheses, but we don't know exactly why it works," said DeepMind research scientist Alhussein Fawzi.

Fawzi and Kohli admitted that when the project was launched, they weren't even sure "if it would work at all."

"As far as we know, this is the first time that a real, new scientific discovery has been made using a large language model," Kohli said.

Obviously, the resulting solution to the "cap set problem" will require a large number of tests to prove the statement correct, but the fact that AI has created something it didn't know before is already promising.

Futurism notes that the confirmation of AI data will be quite a big event as previously, LLMs often made up nonsense, even if the training data contained correct answers.

Earlier, OBOZ.UA reported that AI created a robot that walks on air in 26 seconds.

Subscribe to OBOZ.UA on Telegram and Viber to keep up with the latest events.