Generative artificial intelligence has seen rapid growth since it entered public view in late 2022. It can write, it can generate images and video, and now, it can be a scientist.
Researchers at the University of British Columbia, working together with the University of Oxford and a startup called Sakana.ai, have created the AI Scientist, designed to completely automate the process of scientific discovery.
According to Arizona State University assistant professor Ben Zhou, this use of AI is a significant advancement that could help scientists avoid blind spots in their research.
"We have to first understand what happened in the past and how different things in the past can be connected with the current thing that we are trying to solve," he said, "and this is particularly hard for a single scientist or a group of researchers because there are just so many things going on in the science domain every day."
Large language models like the ones the AI Scientist uses, Zhou said, have seen an amount of scientific research and material that would be incomprehensible to one or even several human scientists, and it remembers everything it's seen.
Zhou said there are ethical concerns with using an AI program to automate scientific discovery — of course, copyright and plagiarism are among them as with most AI models — but he said he doesn't see AI as a threat to human scientists; he sees it more as a lab partner or assistant than a replacement.
Being able to call on a vast database of scientific knowledge is a feat in and of itself, but the AI Scientist takes it several steps further, according to one of its developers, Cong Lu.
Lu said the AI Scientist comes up with its own experiment ideas, then checks those ideas against existing scientific literature to see if they've already been done.
From there, the Scientist "brainstorms" its experiments, writes the necessary code for the experiment, and then performs the experiment before writing up its results in the format of a standard scientific paper.
The Scientist then submits its results to an AI reviewer also developed by the researchers, which Lu said is often "harsher than a human reviewer."
Lu said the AI Scientist then "reflects" on its experiment and uses the results to inform what it might run in the future.
He said this is different from how ChatGPT "learns" because when ChatGPT receives new information, it updates its own model and responses.
The AI Scientist, said Lu, simply uses the results "in context" to guide its future decisions without updating the core model itself.
The idea of automating the scientific process using computers and artificial intelligence has been around for decades, said Lu, but advances in AI have made it a reality.
"Recent progress in large language models really drove home that now is the time that these things are approaching the level of capabilities where they can begin to discover new scientific knowledge," he said.
The AI Scientist is currently limited to computer coding and machine learning-related tasks, but Lu said the team has already been contacted by other research groups hoping to use the AI Scientist for other fields like biology and chemistry.
AI has already been used to discover new protein structures, and Lu said he thinks the AI Scientist can accelerate this kind of discovery once it's applied in other fields.
In an ideal future, Lu said the AI Scientist would become a sort of "force multiplier" for human progress:
"We [scientists] have way too many ideas for the amount of time we have," he said. "Wouldn't it be great if instead of letting all those ideas rot on the back burner, we could just throw something like the AI Scientist on it and at least get initial tests of how viable things are?"
Lu said he sees the AI Scientist ultimately becoming an aid to scientists around the world, not a replacement for humans.
"I really hope that we can sort of realize human ingenuity much more efficiently with technologies like this. And I think this is going to be our future," he said.
