A team from Sakana AI, a company in Tokyo, along with academic labs in Canada and the UK, have created ‘AI Scientist’. This system goes through the whole research process. It reads what’s already been written about a problem, comes up with new ideas to test, tries out different solutions, and writes a paper. AI scientists even do some of the work of peer reviewers by looking at its results.
The results aren’t groundbreaking yet, and the system can only research machine learning itself. AI scientists lack what most scientists think is key to doing science — the ability to do lab work. “We still need to do a lot to go from AI that makes a hypothesis to putting that into action in a robot scientist,” says Gerbrand Ceder, a materials scientist at Lawrence Berkeley National Laboratory and the University of California Berkeley. Still, Cedar adds, “When you look ahead, I’m sure this is where a big part of science will head.”
AI Scientist has its roots in a large language model (LLM). It uses a paper describing a machine learning algorithm as a template. The process begins with a search of the literature to find similar work. Then, the team put into action a method known as evolutionary computation. This approach draws inspiration from the changes and natural selection seen in Darwin’s theory of evolution. The process moves forward in stages making small random tweaks to an algorithm and picking the ones that boost productivity.
The AI Scientist isn’t the first to try to automate parts of a researcher’s job. The idea to automate scientific discovery goes back to the start of artificial intelligence in the 1950s, according to Tom Hope, a computer scientist at the Allen Institute for AI in Jerusalem.
The outcomes add to a discussion that’s on many researchers’ minds these days, West points out. “My colleagues in different fields are trying to figure out how AI fits into our work. It makes us think about what science is in the 21st century — what it could be, what it is, and what it isn’t,” he says.
