A startup gave frontier AI models days on five materials problems. They mostly lost

Discovered Materials put a number on something most AI labs would rather not say out loud: there are scientific problems their models still can’t touch, even given days and an internet connection.

The San Francisco startup, barely out of Y Combinator’s spring 2026 batch, announced a $1 million prize pool on November 5, split five ways at $200,000 a problem. Each one sits in materials science, the field behind semiconductors, batteries and the heat sinks that keep chips from melting. Before it announced anything, the company says it turned frontier AI models loose on the same five problems, gave them internet access and let them run autonomously for days. According to co-founder Advaith Sridhar, speaking to TechCrunch, the models spent that time without making much headway.

That’s the part worth sitting with. Companies building AI-for-science tools don’t usually lead with a list of things their own agents failed to solve.

The five problems

An advisory group of academics, including Stanford’s Felipe Jornada, helped define them. They are, in plain terms:

Problem	What it’s asking for
Cuprate superconductivity	A theory explaining why certain materials carry current with zero resistance at ultra-low temperatures
Silicon/SiO2 interfaces	A numerical method to predict the atomic structure where silicon meets silicon dioxide, present in nearly every chip
Grain structure to heat	A method mapping a material’s internal grain structure to its thermal properties, for better heat dissipation
Light-induced behavior	A theory for why shining light on some materials changes how they behave, tied to superconductivity
Density Functional Theory	A numerical improvement to DFT, the standard technique for modeling atomic and molecular structure

None of these have a clean yes-or-no answer the way a math olympiad problem does. That’s Sridhar’s point: “It’s more difficult to specify problems clearly in materials science because it’s not like math, where you have a very specific definition and can get to a yes or no,” he told TechCrunch. Materials science doesn’t have its own Millennium Prize Problems, so the company and its advisers built a stand-in.

Why a startup is paying for this

Discovered Materials isn’t a nonprofit running a public-interest prize. It builds AI agents that search for new semiconductor materials, and it’s raised a $9 million seed round led by Lightspeed, with Y Combinator and Peak XV also in. A benchmark that frontier labs can’t easily game, because the answers aren’t public yet, is useful to a company whose whole pitch rests on AI being able to do real materials research.

Sridhar frames it differently: “It’s not particularly a concern to us whether we find it, or someone else finds it, or a professor finds it, or the frontier labs find it,” he said. “The reason we’re putting out these challenges is because we want them to serve as the goalposts.” Jornada’s framing, in a statement TechCrunch quoted, leans more on the science: solving any of the five would sharpen physics understanding that feeds directly into how batteries, chips and new materials get built.

My read: both things can be true at once. A real scientific contribution and a marketing move that doubles as a credible benchmark aren’t opposites here. The test is whether the company publishes negative results as openly as a win, if a frontier lab’s model eventually clears one of these.

What the failed attempt actually tells us

My read: AI models have had a good run this year claiming progress on hard math and biology problems, and materials science looks like it’s lagging by comparison. This episode suggests part of why: the problems don’t come pre-packaged with a checkable answer format the way a formal math proof does. An agent can grind on Density Functional Theory for days and produce something that looks like progress without anyone being able to verify it quickly.

That gap between fluent AI output and verifiable AI output has shown up elsewhere on this site. OpenAI’s own release of 722 math papers from an unreleased model ran into a similar problem: only part of the output came with a machine-checkable Lean proof, and mathematicians were left to verify the rest by hand. Materials science doesn’t even have a Lean-style formal checker to fall back on, which is probably why Discovered Materials needed human academics to write the questions in the first place.

Sridhar told TechCrunch the company’s own material-discovery timeline has moved up since its August seed round. “Those timelines have accelerated a little bit,” he said. “I’m more confident now that we will be able to find these industrially useful materials in a timeline shorter than a year.” That’s a more bullish line than the prize announcement itself, and it’s worth remembering both came from the same person in the same conversation.

My read: what to watch is whether any frontier lab (OpenAI, Anthropic, Google DeepMind) takes a public run at one of the five problems with a named model, rather than quietly working it in private. A public attempt and a public failure would tell us more than the prize money ever will.

Note on sources: I read the TechCrunch report in full, including the co-founder’s on-record quotes and the Stanford professor’s statement. I also read TechCrunch’s earlier August story on the company’s seed round and checked Discovered Materials’ own site for a matching announcement; as of writing, the company’s news page doesn’t carry its own write-up of the prize yet, so TechCrunch is the only account of the announcement itself.

Sources

1. A YC startup is putting $1M behind five unsolved materials science problems for AI (TechCrunch, November 5, 2026)
2. Discovered Materials is playing AI whack-a-mole to hunt cooler chips (TechCrunch, August 10, 2026)
3. Discovered Materials (company site)