Demis Hassabis didn’t know he was getting the Nobel Prize in chemistry from the Royal Swedish Academy of Sciences until his wife started being bombarded with calls from a Swedish number on Skype.
“She would put it down several times, and then they kept persisting,” Hassabis said today in a press conference convened to celebrate the awarding of the prize, alongside John Jumper, his colleague at Google DeepMind. “Then I think she realized it was a Swedish number, and they asked for my number.”
That he won the prize—the most prestigious in science—may not have been all that much of a shock: A day earlier, Geoffrey Hinton, often called one of the “godfathers of AI,” and Princeton University’s John Hopfield were awarded the Nobel Prize in physics for their work on machine learning. “Obviously the committee decided to kind of make a statement, I guess, when having the two together,” said Hassabis in a press conference organized after his win.
In case it wasn’t clear: AI is here, and it’s now possible to win a Nobel Prize by studying it and contributing to other fields—whether physics in the case of Hinton and Hopfield or chemistry in the case of Hassabis and Jumper, who won alongside David Baker, a University of Washington genome scientist.
“It’s no doubt a huge ‘AI in science’ moment,” says Eleanor Drage, senior research fellow at the University of Cambridge’s Leverhulme Center for the Future of Intelligence. “Going by highly accomplished and illustrious computer scientists winning a chemistry prize and a physics prize, we’re all bracing for who will be awarded a peace prize,” she says, explaining that colleagues in her office were joking about xAI owner Elon Musk being tipped for that award.
Drage calls the awarding of physics and chemistry prizes to AI researchers “a major polemic, not only within those disciplines, but looking in from the outside.” She suggests the awards could be for one of two reasons: either a notable shift in disciplinary boundaries enabled by the ubiquity of AI in academic research, or because “we’re so obsessed with computer scientists that we’re willing to slot them in anywhere.”
She isn’t sure which route this week’s decisions signify. But she and others are sure that it’ll make a meaningful difference to the future of research.
“Winning a Nobel by using AI may be a ship that’s sailed, but it will influence research directions,” says Matt Hodgkinson, an independent scientific research integrity specialist and former research integrity manager at the UK Research Integrity Office. The question is whether it’ll influence them in the right way.
Baker, one of this year’s winners of the Nobel Prize for chemistry, has long been one of the leading researchers in the use of AI for protein-structure prediction. He had been laboring away for decades at the problem, making incremental gains, recognizing that the well-defined problem and format of protein structure made it a useful test bed for AI algorithms. This wasn’t a fly-by-night success story—Baker has published more than 600 papers in his career—and neither was AlphaFold2, the Google DeepMind project that was awarded the prize by the committee.
Yet Hodgkinson worries that researchers in the field will pay attention to the technique, rather than the science, when trying to reverse engineer why the trio won the prize this year. “What I hope this doesn’t do is make researchers inappropriately use chatbots, by wrongly thinking that all AI tools are equivalent,” he says.
The fear that this could happen is founded in the explosion of interest around other supposedly transformative technologies. “There’s always hype cycles, recent ones being blockchain and graphene,” says Hodgkinson. Following graphene’s discovery in 2004, 45,000 academic papers mentioning the material were published between 2005 and 2009, according to Google Scholar. But after Andre Geim and Konstantin Novoselov’s Nobel Prize win for their discovery of the material, the number of papers published then shot up, to 454,000 between 2010 and 2014, and more than a million between 2015 and 2020. This surge in research has arguably had only a modest real-world impact so far.
Hodgkinson believes the energizing power of multiple researchers being recognized by the Nobel Prize panel for their work in AI could cause others to start congregating around the field—which could result in science of a changeable quality. “Whether there’s substance to the proposals and applications [of AI] is another matter,” he says.
We’ve already seen the impact of media and public attention toward AI on the academic community. The number of publications around AI has tripled between 2010 and 2022, according to research by Stanford University, with nearly a quarter of a million papers published in 2022 alone: more than 660 new publications a day. That’s before the November 2022 release of ChatGPT kickstarted the generative AI revolution.
The extent to which academics are likely to follow the media attention, money, and Nobel Prize committee plaudits is a question that vexes Julian Togelius, an associate professor of computer science at New York University’s Tandon School of Engineering who works on AI. “Scientists in general follow some combination of path of least resistance and most bang for their buck,” he says. And given the competitive nature of academia, where funding is increasingly scarce and directly linked to researchers’ job prospects, it seems likely that the combination of a trendy topic that—as of this week—has the potential to earn high-achievers a Nobel Prize could be too tempting to resist.
The risk is this could stymie innovative new thinking. “Getting more fundamental data out of nature, and coming up with new theories that humans can understand, are hard things to do,” says Togelius. But that requires deep thought. It’s far more productive for researchers instead to carry out simulations enabled by AI that support existing theories and involve existing data—producing small hops forward in understanding, rather than giant leaps. Togelius foresees that a new generation of scientists will end up doing exactly that, because it’s easier.
There’s also the risk that overconfident computer scientists, who have helped advance the field of AI, start to see AI work being awarded Nobel Prizes in unrelated scientific fields—in this instance, physics and chemistry—and decide to follow in their footsteps, encroaching on other people’s turf. “Computer scientists have a well-deserved reputation for sticking their noses into fields they know nothing about, injecting some algorithms, and calling it an advance, for better and/or worse,” says Togelius, who admits to having previously been tempted to add deep learning to another field of science and “advance” it, before thinking better of it, because he doesn’t know much about physics, biology, or geology.
Hassabis is an example of using AI well in order to advance science. He was a neuroscientist by training, gaining a PhD in the subject in 2009, and has credited that background to helping advance AI via Google DeepMind. But even he acknowledged a change in how the sector ekes out efficiencies. “Today, [AI] has become more engineering-heavy,” he said in his Nobel Prize press conference. “We have a lot of techniques now that we’re improving just algorithmically, without reference to the brain anymore.”
That too could have an impact on what kind of research gets done—and who does it, their level of knowledge of the field, and the incentives behind them entering it. Rather than researchers who have devoted their lives to a specialism, we could see more research by computer scientists, detached from the reality of what they’re looking at.
But that’s likely to take a backseat to the celebrations for Hassabis, Jumper, and the colleagues they both thanked for helping them win the Nobel Prize this week. “We’re very close to cleaning up the [AlphaFold3] code to release it for the academic community to freely use,” he said earlier today. “Then we’ll keep progressing from there.”
Source : Wired
