New AI Tool Could Help Scientists Make ‘Rare’ Cross-Disciplinary Leaps More Common

Some major scientific advances involve taking a technique developed in one field and applying it to another, but interdisciplinary research is tricky and time-consuming. Now a new AI research system has automated the process, generating 36 papers in three months.

Most researchers work within a narrow specialty, and few have time to comb through unrelated fields for tools that might help them. As a result, researchers can end up banging their heads against a problem scientists in another discipline have already solved.

But Harvard physicist Matthew Schwartz stumbled across a potential solution while working as a visiting researcher at AI firm Anthropic. After getting the company’s AI, Claude, to develop a suite of tools for mathematical physics, Schwartz discovered they could help solve quantitative problems in a host of other areas.

Called BootLoops, the tools, which are model agnostic, help AI search scientific literature, write code, and solve mathematical problems. Working with 19 coauthors, Schwartz used BootLoops to produce 36 manuscripts on topics ranging from ecology and linguistics to health insurance and sunspots.

“Once Claude had BootLoops, it kept noticing the same pattern: many fields have problems that a technique from mathematics, physics, or computer science would solve outright if anyone knew it existed,” Schwartz wrote in an Anthropic blog post about the system.

Schwartz first started using Claude as a general-purpose research assistant last December, and while he was able to do some useful work with it, he says it required constant handholding. So he decided to instead point it at problems that were a good match for the model’s strengths—what he calls “Claude-shaped problems.”

He started with the equations physicists use to interpret particle collisions at the Large Hadron Collider. The best methods were scattered across software in several programming languages or existed only on paper, so he got Claude to combine them into a single framework. Using the tools it had built itself, the model reproduced the results of one of his papers in about 20 minutes.

Schwartz then moved onto a harder class of these equations, and within a few weeks, the model had solved 30 of them, including 15 that had never been calculated before. It didn’t take long for Schwartz to realize that many of the methods the AI had developed could be used in neighboring areas of physics such as cosmology and string theory.

But it was Claude itself that made the bigger leap, pointing out that the same kinds of equations turn up in population genetics and evolutionary biology. The problem for Schwartz, however, was that he found it harder to tell when Claude was overselling its findings in fields he was less familiar with. So he brought in experts from each field to judge which of Claude’s findings were worth pursuing, and from some 400 candidate problems, 36 ended up as papers.

One project examined the star ratings used to grade private health plans under Medicare, which determine the bonuses insurers receive. The rules say the boundaries between ratings levels should be drawn so each plan’s score sits as close as possible to the average of its group, but the algorithm Medicare actually uses is a shortcut that doesn’t guarantee this.

Using a statistical method dating back to 1958 to calculate the exact answer, BootLoops found drawing precise boundaries would push 34 plan ratings across the bonus line, 16 up and 18 down. This would shift an estimated $0.9 to $1.4 billion in bonus payments between 2025 and 2027.

Another project tackled a 150-year-old puzzle about what kills sunspots. Because the sun’s rotation hides sunspots from view for much of their lives, astronomers have never been able to agree on how long they last or why they disappear.

Borrowing a tool doctors use to measure how the risk of dying changes with age and applying it to every sunspot logged since 1874, the team found that sunspots age. At a given size, their risk of “dying” rises about 20 percent a day, something none of the published theories could account for.

For each new problem, Claude built new tools, so the BootLoops toolkit kept growing, and it now spans 22 fields, according to the project’s website. Other problems it’s tackled include how Earth’s atmosphere first filled with oxygen, how quickly tree species come and go in a Panamanian rainforest, and how natural selection weeds out harmful mutations in humans.

The kinds of cross-disciplinary leaps BootLoops can make are extremely rare, Georgia Channing, artificial intelligence for science lead at Hugging Face, told Science. “A physics researcher wouldn’t spend six months reading linguistics research in order to figure out how to repurpose linguistics methods,” Channing said.

That’s not to say the system isn’t without flaws. Schwartz says many of Claude’s early findings were technically correct but scientifically unremarkable until experts steered them, and he warns that the model “loves to declare victory” too early.

But the results suggest AI may be a powerful way to pollinate ideas across research fields, even if we still need humans to decide which ones are really worth pursuing.

The post New AI Tool Could Help Scientists Make ‘Rare’ Cross-Disciplinary Leaps More Common appeared first on SingularityHub.