PSI is trying to make a business out of AI-driven scientific discovery

A researcher at the startup Physical Superintelligence, or PSI, used early access to Anthropic's Fable 5.2 model to prove the Aaronson-Ambainis conjecture, a long-open problem about when quantum computers can beat classical ones. Scott Aaronson, who co-posed it, says he doesn't fully understand the proof but believes it's correct because it was autoformalized in Lean and explained by top researchers PSI shared it with.

That combination, a machine-found result, machine-checked and then human-explained, is becoming a template. It matters for what science looks like, who gets credit, and how fast fields can move when a model can clear problems experts considered centrally important.

We want to give this its due. Aaronson described it as one of the last remaining open problems in quantum query complexity that everyone seemed to agree was centrally important. The result is a negative one: it closes off one route quantum computers might have taken to solve certain hard problems quickly, though not all of them. Negative results are underrated. Knowing where not to look saves careers' worth of effort, and this is AI doing exactly that kind of useful clearing work.

PSI's bet is that the model alone isn't the product. CEO Matt Pines argues an effective research system needs harnesses, tools, memory and simulation infrastructure wrapped around it. Aaronson's own aside supports that: he was surprised the problem wasn't in OpenAI's recent batch of solved problems and wondered if it was simply harder. If scaffolding is what made the difference, that's a real business moat, at least until the labs build their own.

The tension is the jump from math to matter. Pines says it himself: physics is different because the real world has the final say. Proofs can be verified in code. Superconductors and heat-dissipating materials have to work in a lab. PSI says it's focused on validating computational results with experiments as efficiently as possible, and for now its revenue comes from something much more grounded: simulating heat, power, water and sound to optimize data centers. That's sensible, but it means the grand discovery thesis is still a thesis.

The human side of this unsettled us a little. David Turturean, the recent MIT graduate who found the proof, told TechCrunch he accepts this might be the last paper he ever puts out. Aaronson described feeling satisfied, awestruck, disappointed and fatigued. When the people closest to a breakthrough sound partly bereaved, that's a signal. If this scales, the bottleneck in science shifts from finding answers to understanding and checking them, and the path for training the next generation of researchers gets harder to justify.

We do credit PSI for how it handled release: holding the result, sharing it with experts, and explaining it before going public. Norms are still forming, and a summit at MIT next week is meant to help set them. I'm watching whether those norms land on credit, disclosure of model involvement, and human explainability before the pace makes them moot.

Autoformalization in Lean means the proof was translated into a formal language that software can check line by line, which is why experts can trust a result they haven't fully absorbed. Separately, the startup Discovered Materials launched a $1 million grand challenge last week for researchers using AI on problems such as superconductivity and heat dissipation.