PSI claims early-access Anthropic model helped solve 18-year-old quantum computing conjecture

An 18-year-old quantum question gets an answer

The startup Physical Superintelligence, or PSI, is trying to turn AI-driven scientific discovery into a business, and it now has a headline result to point to. According to TechCrunch, David Turturean, a recent MIT graduate on PSI's technical staff, used early access to Anthropic's Fable 5.2 model to produce a proof of the Aaronson-Ambainis conjecture, a quantum computing problem that had stood open for 18 years.

The conjecture, posed by Scott Aaronson and Andris Ambainis, concerns when quantum computers can outperform classical machines. As TechCrunch reports, Aaronson wrote that he cannot fully follow the argument himself, but that he accepts it because it was autoformalized in Lean, a programming language for expressing mathematics that computers can check, and because a group of top researchers PSI shared the proof with were able to explain it to him.

The outcome is a negative result: it apparently rules out one route quantum computers might have taken to certain hard problems quickly, while leaving others open. Turturean told TechCrunch the value is in telling the field where to stop searching.

Harnesses, not just models

PSI's core claim is that raw model capability is not the whole story. CEO Matt Pines told TechCrunch that a serious research system needs layers of harnesses, tools, memory systems and simulation infrastructure so it can attack a problem at both breadth and depth.

Aaronson, for his part, said a full proof of the conjecture was among the last open problems in quantum query complexity that researchers widely agreed was centrally important, and that he was surprised to see it missing from OpenAI's recent mass release of solved mathematics problems, referred to as the "Mathocalypse" dump. He reasoned that OpenAI must have tried it, and wondered whether it was simply harder than everything else that fell. TechCrunch reads that omission as indirect support for PSI's approach: specialized scaffolding aimed at one problem may beat a general-purpose system run at scale.

What pays the bills today

PSI's current revenue, per TechCrunch, comes from a more mundane application: using its simulation tools to optimize data centers, juggling heat, power, water and noise, since a site's performance and its effect on the surrounding community come down to physics. The long-term play is converting theoretical physics results into actual technologies.

That second step is where it gets hard. AI has notched notable wins in mathematics and computer science, where answers can be checked in validated code; physics and the life sciences have resisted. Pines told TechCrunch that physics differs from math because reality can overrule your model — experiments, not just proofs, decide what is true. He said PSI is working on the most efficient ways to test computational results in the physical world, since a genuine fact of nature is something you can build a device around. TechCrunch also notes that the startup Discovered Materials recently put up a $1 million challenge for researchers using AI on problems such as superconductivity, part of a wider wave of so-called neolabs trying to industrialize discovery.

Norms are still being written

Because PSI had to hold the result until Anthropic's model was nearly ready for release, it used the waiting period to share the proof with domain experts. Jim Halverson, a Northeastern professor who advises PSI, told TechCrunch that the team settled on transparency and early contact with specialists as the responsible route, and that he will attend an MIT summit where leading physicists will discuss how to respond to the new pace of science.

Turturean, who deferred a master's at MIT's computer science and AI lab to join PSI, said this was the first time he took an AI-produced solution into print and worked out how to present it so humans could follow it, and that the experience made the scale of what AI can do sink in. He also told TechCrunch he has fully accepted that this may be the last paper he ever publishes, reasoning that within months there may be little point in publishing one at all.

Why it matters

If the proof holds, one of the last major open problems in quantum query complexity is settled, giving researchers a clearer map of where quantum speedups are and are not possible. The story is also a data point in a larger argument: frontier models aimed at open problems with purpose-built tooling may open scientific territory that general systems miss. And it lands amid unresolved questions — how to verify machine-generated results, how to credit them, who gets early model access and on what terms, and whether a durable business can be built on turning proofs into machines.