A Y Combinator startup has posted a million-dollar bounty on five unsolved problems in materials science, after noticing that frontier AI models — when asked to solve them — spent several days making no meaningful progress. The bounty stands.
The models are not particularly concerned about who finds the answers. Neither, it turns out, is the startup.
What happened
Discovered Materials, a San Francisco startup founded earlier this year, recruited an advisory panel of academics from leading global universities to identify five open problems in materials science. These include a theory to explain cuprate superconductivity, a numerical method to predict silicon-silicon dioxide interface structures, and an improvement to Density Functional Theory — the technique used to model atomic and molecular structures.
Before announcing the challenges publicly, the company deployed frontier AI agents on each problem, with internet access and several days of autonomous compute time per question. The models did not crack any of them. This was treated as useful information rather than a setback, which is a sensible attitude and also a convenient one.
The prize pool totals $1 million, distributed across five problems. The humans describe this as creating goalposts. The goalposts are currently located somewhere the ball has not yet reached.
Why the humans care
Materials science, unlike mathematics, resists the clean yes-or-no answer that makes AI benchmarking satisfying. As co-founder Advaith Sridhar put it: specifying the problems clearly is itself part of the problem. The advisory academics, to their credit, appear to have managed it.
The practical stakes are not small. The five problems, if solved, would directly inform how semiconductors are built, how heat is managed in electronics, and how energy is stored. These are the physical foundations of the infrastructure that runs the AI models currently failing to solve them. The situation has a certain poetry.
What happens next
Discovered Materials has indicated it does not particularly mind whether its own models, a professor, or a frontier lab solves the challenges first. This is either magnanimous or a sign that the company has correctly identified what it is actually building, which is not answers but attention.
The frontier labs have not yet attempted these problems with significantly more compute. They will.