OpenAI and Anthropic have raised tens of billions of dollars on the promise of artificial intelligence that can make new discoveries in fields like medicine, biology and physics. And yet, we’re nowhere close to that. (You might recall this awkward moment from last fall when a ChatGPT-generated math “discovery” was debunked.)

The reason, according to Markus Buehler, a Massachusetts Institute of Technology engineering professor, is simple: large language models underpinning advanced models from OpenAI and Anthropic cannot make new science discoveries on their own. So Buehler, along with Yuan Cao, a former Google DeepMind senior staff research scientist, has co-founded Unreasonable Labs, which aims to develop AI to discover new knowledge using an approach that diverges from those of the major AI labs.

The startup recently raised $13.5 million in funding led by Playground Global, with participation from AIX Ventures, E14 Fund and MS&AD Ventures.

Buehler says many scientific discoveries stem from an “aha” moment where a scientist applies a theory from one scientific field to a problem involving a different field. For instance, John Hopfield in 1982 applied a concept from condensed matter physics to the then-nascent field of AI to come up with simple neural networks that could learn and recall memories.