AI researchers claim technology behind ChatGPT can generate new insights

AI researchers at Google DeepMind have made the world's first scientific discovery using a large language model (LLM), indicating that ChatGPT and similar programs can generate information beyond human knowledge.

The discovery suggests that these large language models can not only repackage training information but also generate new insights.

“When we started the project there was no indication that it would produce something that’s genuinely new,” said Pushmeet Kohli, the head of AI for science at DeepMind.

“As far as we know, this is the first time that a genuine, new scientific discovery has been made by a large language model.”

LLMs are powerful neural networks that learn language patterns from vast amounts of text and data. ChatGPT, introduced last year, has been popular for debugging software and creating various types of content, The Guardian reported.

However, chatbots are not able to generate new knowledge and are prone to confabulation, leading to fluent and plausible answers that are flawed.

DeepMind used an LLM to create "FunSearch", short for “searching in the function space”, by writing computer programs to solve problems. The LLM is paired with an "evaluator" that ranks programs based on performance.

The best programs are combined and fed back to the LLM for improvement, progressively evolving poor programs into powerful ones capable of discovering new knowledge.

The researchers set FunSearch loose on two puzzles.

For the first one, FunSearch developed programs to generate large cap sets beyond the best existing mathematicians' work, addressing the longstanding and arcane challenge of finding the largest set of points in space where no three points form a straight line.

The second one was the bin packing problem — a mathematical problem that aims to efficiently pack items of different sizes into containers — that applies to physical objects like shipping containers and scheduling computing jobs in data centres.

The solution is usually to either pack items into the first bin with space or into the bin with the least available space.

FunSearch found a better approach that avoided leaving small gaps that were unlikely ever to be filled, according to results published in Nature.

“In the last two or three years there have been some exciting examples of human mathematicians collaborating with AI to obtain advances on unsolved problems,” said Sir Tim Gowers, professor of mathematics at Cambridge University, who was not involved in the research.

“This work potentially gives us another very interesting tool for such collaborations, enabling mathematicians to search efficiently for clever and unexpected constructions. Better still, these constructions are humanly interpretable.”