AI experts assert that ChatGPT's technology can produce fresh insights.

 While utilizing chatbots to solve challenging math problems, researchers made a groundbreaking discovery.

Using a large language model (LLM), Google DeepMind AI researchers have made the first scientific discovery in history, suggesting that ChatGPT and related software can produce information that is beyond human comprehension.

The finding implies that these massive language models are capable of producing fresh insights in addition to repurposing training data.

Head of AI for science at DeepMind Pushmeet Kohli stated, "When we started the project there was no indication that it would produce something that's genuinely new."

"To the best of our knowledge, this is the first time a large language model has produced a true, novel scientific finding."

Large-scale text and data sets are used by LLMs, which are robust neural networks, to identify language patterns. According to The Guardian, ChatGPT, which was released a year ago, has gained popularity for both producing different kinds of content and debugging software.

Nevertheless, chatbots are unable to produce original knowledge and are prone to confabulation, which results in well-written but inaccurate responses.

"FunSearch"—short for "searching in the function space"—was developed by DeepMind using an LLM by writing computer programs to solve issues. Alongside the LLM is a "evaluator" who assigns a program's performance rating.

The most effective programs are merged and fed back to the LLM for enhancement, gradually transforming weak programs into strong ones that are able to uncover new information.

FunSearch was deployed on two puzzles by the researchers.

For the first, FunSearch tackled the venerable problem of determining the largest set of points in space where no three points form a straight line by creating programs that could generate large cap sets beyond the capabilities of the best mathematicians currently working.

The second one was the bin packing problem, a mathematical puzzle that deals with the efficient packing of various-sized objects into containers. It can be applied to the scheduling of computing tasks in data centers and physical objects such as shipping containers.

Usually, packing items into the first available bin or the bin with the least amount of space available is the solution.

Results published in Nature show that FunSearch discovered a better method that avoided leaving tiny gaps that were unlikely to ever be filled.

"There have been some exciting examples of human mathematicians working with AI to obtain advances on unsolved problems in the last two or three years," said Cambridge University mathematics professor Sir Tim Gowers, who was not involved in the study.

This work can provide us with another fascinating tool for these kinds of collaborations, allowing mathematicians to effectively look for novel and creative constructions. Even better, these constructions can be understood by humans.