Source: What science problem should AI tackle next? A new ‘challenge atlas’ offers answers
Copy:
What science problem should AI tackle next? A new ‘challenge atlas’ offers answers
Four complex physics problems are now within AI’s reach, researchers say
---
Artificial intelligence (AI) models are becoming formidable engines of discovery, as OpenAI’s controversial breakthrough on the famous Navier-Stokes equations made clear this week. The company appeared to best more than 200 years of human attempts with a strategy that boiled down to one key thing: applying massive computational force.
Over the past few months, a group of researchers has been wondering whether a computer-powered approach could advance other fields of science, too. Their answer, released this week, is the Scientific Challenge Atlas: a curated list of scientific problems whose solutions AI could unearth from existing data. The first four challenges, all in physics, range from pinning down complicated quantum calculations to figuring out limits to how complex special geometric spaces can become.
“Based on all the data that we already have, are there new science nuggets just waiting to be discovered?” says Rick Stevens, an associate laboratory director at the Department of Energy’s Argonne National Laboratory.
“We want to immediately advance science,” adds Karthik Duraisamy of the University of Michigan (UM), who co-organized the atlas along with Stevens and Jason Pruet of OpenAI. “Where are problems where AI immediately helps, and what is the path to getting to bigger challenges?”
The first problems will likely take one to two orders of magnitude less effort than OpenAI’s multimillion-dollar investment in attacking Navier-Stokes, according to Duraisamy. So far, the creators estimate solving each problem will require in the range of 1 billion to 10 billion tokens (the basic unit of AI processing) or up to about 1 billion hours of supercomputing time.
The atlas arrives as tech startups promise smarter AI scientists and national governments around the world deploy AI against sprawling grand challenges, in programs such as the United States’s Genesis Mission. But providing the AI with real-world data can be a roadblock. “Many problems are bottlenecked by your ability to build a big telescope or particle accelerator,” Duraisamy says. Other problems, however, already have all the ingredients—such as big sets of data—ready for machines.
The atlas’ creators were originally inspired by AI’s success in tackling a list of problems posed by the celebrated Austro-Hungarian mathematician Paul Erdős. The 1000-plus Erdős problems lie in specific branches of math that are generally friendlier territory to large language models (LLMs). In addition, the problems—which previously had been scattered in the technical literature and thus often largely forgotten—were conveniently collected for the first time on a website in 2023 by University of Manchester mathematician Thomas Bloom. So far, AI models have cracked several Erdős problems in high-profile upsets.
There is also something to learn from how Erdős came up with the problems in the first place, Bloom says. His questions were “the product of decades of talking to people at conferences, writing papers, and thinking about these things,” he says. “You can sort of view these problems as the output of 20 mathematicians talking to each other and exchanging ideas.”
The atlas’ creators wondered whether AI agents (AI systems that can autonomously take actions) could mimic and supercharge that process. So they began with an open-ended experiment: If you give hundreds of agents unlimited tokens, what sorts of science would they identify as most novel and doable?
They began by having the agents search for scientific problems, both by scanning their own knowledge banks and through reading recent papers. But the researchers also wanted to know how hard the problems would be to solve. One way to estimate that was to have the agents reproduce results from published papers. This exercise provided a yardstick: If an agent could replicate an existing result with a certain amount of effort, the researchers could estimate how much more effort it would take to produce a new, publishable result. That calibration let them put a rough price tag on the problems agents proposed.
The result was a prototype map of the scientific problem space, with reasoning measured in tokens along one axis and supercomputing time on the other. “You can see points all over it, where the models think, “Oh, there’s a problem here, and we could make headway if you gave me 10 billion tokens and 10,000 hours on the exascale machine,” Stevens says. “It’s like a star atlas that can help people navigate the space of problems.”
Experts in different fields have been reviewing and refining the problems, and helped pick the first four to be unveiled. They are described in “cards” that list the question, the form an answer would take, why it matters, and how AI could help. “We’re not picking problems that are going to take 5 years,” Duraisamy says. “I mean, it’s our hope that in the next 6 to 12 months we can get answers to these problems.”
For example, one asks whether a theory that can help reconcile quantum mechanics with gravity continues to hold under extreme conditions, by pushing a notoriously difficult calculation to a seventh “loop.” The problem would take more than 3 million integrals to solve, and physicists have been stuck on it for more than 16 years. “It’s not a calculation that I would throw myself into at this point without some completely new way of doing things,” says Henriette Elvang, a physicist at UM who served as a reviewer for the problem. But AI might be able to pursue the strategy at about a cost of 1 billion to 10 billion tokens, the team’s estimates showed—though other reviewers suggested it is difficult to extrapolate such guesses on the frontier.
Looking ahead, Duraisamy wants the atlas to “outgrow us, where each field has its own challenge atlas in their own form.” Eventually, he envisions thousands of cards with challenges across all fields of science. Stevens also plans to make the underlying map available not just as a website, but in some machine-readable form, so that people can consume it with their agents.
The intended audience extends beyond scientists. For policymakers deciding where to put research dollars, the atlas could offer a rough accounting of where AI might make faster advances and the price needed to get there. “Science policymakers have to make decisions about the relative investment of resources against some metric of progress,” Stevens says. “The atlas, if we do a good job, will give us a broader view of what’s possible and roughly what it might cost in different areas to make progress.”
Không có nhận xét nào:
Đăng nhận xét