Physical Superintelligence (PSI), a startup research lab, says early access to Anthropic’s Fable 5.2 helped it solve the Aaronson-Ambainis conjecture. That’s an open problem in quantum computing that had gone unsolved for 18 years. TechCrunch AI reports that the proof was formalized in Lean and checked by top researchers before PSI went public. The company is now trying to build a business on AI-driven discovery, starting with data centers and aiming much higher.
The Short Version
- What happened: David Turturean, a recent MIT grad on PSI’s technical staff, used Fable 5.2 to prove a central conjecture in quantum query complexity.
- Who checked it: Scott Aaronson, who co-posed the problem, says he doesn’t fully follow the proof. He still believes it’s correct because it was autoformalized in Lean and explained by a team of experts PSI shared it with.
- What it means: It’s a negative result. It closes off one route quantum computers might take to beat classical machines on certain hard problems, though not every route.
- The business angle: PSI makes money today by simulating data center physics. The bigger bet is turning AI-generated physics breakthroughs into real devices.
Why This One Matters
Aaronson called the full conjecture “one of the last remaining open problems in quantum query complexity that everyone seemed to agree was centrally important.” He also said he felt satisfied, awestruck, disappointed and fatigued all at once. That mix of feelings tells you a lot about where academic science is right now.
What stands out to me is a detail Aaronson pointed out himself. The problem was missing from what he called OpenAI’s “Mathocalypse dump” a few weeks ago. “Surely they tried it, so was it really so much harder than all the problems that were solved?” he wrote.
That gap supports PSI’s core argument. A raw frontier model isn’t enough on its own. The edge comes from the system you build around it.
The Harness Is the Product
PSI CEO Matt Pines describes the company’s approach as specialized scaffolding wrapped around frontier models. According to Pines, a good research system needs:
- Harnesses that steer the model through long, structured problem-solving
- Tools the model can call during its work
- Memory systems so progress builds up instead of resetting
- Simulation infrastructure to test ideas against physical reality
“You want to be able to really go after a problem at breadth and depth, and be able to consider a wide range of potential alternatives,” Pines told TechCrunch.
Practitioners should take note. The model is becoming a commodity input. The orchestration layer on top is where the differentiation happens.
Physics Is Harder Than Math
AI has racked up wins in math and computer science because results there can be checked as code. Physics and the life sciences don’t work that way. As Pines put it, physics “is different than math because it’s the real world … nature gets a vote in your models of reality.”
So PSI is working on the loop that turns computation into experiments. The AI suggests an idea, experiments confirm it, and a confirmed fact of nature becomes “a trick that you can exploit.” In Pines’ words: “You could build a device out of it.”
PSI isn’t alone here. Last week, the startup Discovered Materials launched a $1 million grand challenge for researchers using AI on problems like superconductivity and heat dissipation.
Pines admits the investor pitch has gotten easier. Six to nine months ago he had to argue that theoretical physics would soon be “mostly the domain of machines.” His take now: “it’s not there yet, but it’s on its way.”
The Norms Are Still Being Written
Because PSI had to wait for Fable 5.2 to near public release, it had time to share the proof with experts before going public. Northeastern professor Jim Halverson, who advises PSI, said “transparency is a good way to go, and contacting experts is a good way to go.” He’ll be at an MIT summit next week where leading physicists will try to agree on a response to the new pace of science.
Turturean, for his part, spent that waiting period rewriting the AI’s proof so a human could actually follow it. He knows academics will expect him to explain it. He’s also blunt about where this is heading: “I’m totally accepting the idea that this might be the last paper I ever put out.”
What to Watch
- The MIT summit: look for early standards on how to verify, disclose and credit AI-generated results.
- Harness competition: expect more “neolabs for science” built on frontier models, competing on scaffolding rather than model weights.
- Physical validation: the real test is whether any lab can close the loop from AI insight to a working device.
If PSI can do in physics what Fable just did for quantum complexity theory, the pitch stops being a bet and becomes a business. Full details are in the original TechCrunch AI report.