by Artificial intelligence researchers claim to have made the world’s first scientific discovery using a large language model, suggesting that the technology behind ChatGPT and similar programs can generate intelligence beyond human knowledge.
The result came from Google DeepMind, where scientists are investigating whether large language models, which support modern conversations such as ChatGPT OpenAI and Google’s Bard, can do more than repackage information learned in training and come on new insights.
“When we started the project there was no indication that it would create something truly new,” said Pushmeet Kohli, head of AI for science at DeepMind. “As far as we know, this is the first time that a major language model has made a truly new scientific discovery.”
Large language models, or LLMs, are powerful neural networks that learn patterns of language, including computer code, from vast amounts of text and other data. Since ChatGPT’s introduction last year, the technology has debugged faulty software and extracted everything from college essays and itineraries to Shakespeare-style climate change poems.
But while chatbots are popular, they don’t generate new knowledge and are often counterproductive, giving answers that are fluent and believable, but very flawed on par with pub crawls the best.
Related: ChatGPT to summarize Politico and Business Insider articles in ‘first of its kind’ deal
To build “FunSearch”, short for “search in function space”, DeepMind used LLM to write solutions to problems in the form of computer programs. The LLM is paired with an “evaluator” that automatically ranks the programs according to how well they perform. The best programs are then compiled and fed back to the LLM for improvement. This drives the system to constantly evolve poor programs into more powerful programs that can acquire new information.
The FunSearch researchers set loose on two puzzles. The first was a well-established challenge in pure mathematics known as the capping problem. It is about finding the largest set of points in space where no three points are in a straight line. FunSearch has released programs that generate new sets of large caps that go beyond the best mathematicians have done.
The second answer was the bin packing problem, which seeks the best ways to pack items of different sizes into containers. While it applies to physical objects, such as the most efficient way to arrange boxes in a shipping container, the same math applies in other areas, such as scheduling computer jobs in data centers. The problem is usually solved by packing items into the first bin that has room, or into the bin with the least space available where the item will still fit. FunSearch found a better approach that avoided leaving small gaps that are unlikely to ever be filled, according to results published in Nature.
“Over the past two or three years there have been some exciting examples of human mathematicians collaborating with AI to make progress on unsolved problems,” said Sir Tim Gowers, professor of mathematics at Cambridge University, who was not involved in the research. . “This work could give us another very interesting tool for such collaborations, enabling mathematicians to efficiently search for unexpected intelligent constructs. Even better, these constructs are humanly interpretable.”
Researchers are now exploring the range of scientific problems FunSearch can handle. A major limiting factor is that the solutions to the problems must be automatically verifiable, which precludes many questions in biology, where hypotheses often need to be tested with laboratory experiments.
The more immediate impact may be on computer programmers. Over the past 50 years, coding has improved greatly as people have created more specialized algorithms. “This is truly going to be transformative in how people approach computer science and algorithmic discovery,” Kohli said. “For the first time, we are seeing LLMs that are not taking over, but certainly helping to push the limits of what is possible in algorithms.”
Jordan Ellenberg, professor of mathematics at the University of Wisconsin-Madison, and a co-author on the paper, said: “What’s really exciting to me, even more than the specific results we found, are the prospects it suggests for the future of human interaction – engine in mathematics.
“Instead of generating a solution, FunSearch generates a program that finds the solution. Solving a particular problem might not give me any insight into how to solve other related problems. But a program that finds the solution, that’s something that a human can read and interpret and hopefully generate ideas for the next problem and the next problem and the next problem.”
