Startup PSI Closes 18-Year-Old Aaronson–Ambainis Conjecture and Starts Selling AI Science

David Turturean had just graduated from MIT, put off a master's program at the famous computer science and AI lab, and joined the startup Physical Superintelligence (PSI). There he got early access to Anthropic's new model Fable 5.2 and used it to close the Aaronson–Ambainis conjecture, which had hung over quantum complexity theory for 18 years. "I fully admit this could be my last paper," he says. "In a couple of months, why even publish them?"

What exactly was solved, and why it can be trusted

The conjecture concerns where quantum computers can outperform classical ones. The result is negative: it closes one of the directions along which quantum machines might have quickly solved hard problems, but not all of them. According to Turturean, researchers now know there is no need to look in this direction any more. For science that is valuable, because it saves others years of attempts.

The main question is a different one: how do you verify a proof written by a model? Scott Aaronson, one of the authors of the conjecture, admits he does not fully understand it. He believes it for two reasons. First, the proof was autoformalized in Lean, meaning a compiler checks it, the way tests check code. Second, a team of strong researchers to whom PSI showed the solution went through it. For a developer the analogy is simple: you don't read ten thousand lines of generated code, but you have a test suite that is hard to fool.

Aaronson described his reaction as a mix of satisfaction, awe, disappointment and exhaustion. He was also surprised that the problem was not in OpenAI's "Mathocalypse" dump, published a few weeks earlier. Surely it had been tried, which means it turned out to be noticeably harder than the others.

A bet on the harness, not only the model

The model in this story is someone else's, from Anthropic. PSI's business is built around specialized harnesses and scaffolding for scientific tasks. CEO Matt Pines says a good research system contains a lot besides the model itself: tools, memory systems, simulation infrastructure. You need to be able to dig into a problem both broadly and deeply and to try a wide range of alternatives.

There is no money in theoretical physics yet. PSI's revenue comes from data center optimization: the company simulates heat, power consumption, water and noise. A site's efficiency and its impact on its neighbors depend directly on physics. The long-term goal is to turn theoretical physics breakthroughs into real technologies. Six to nine months ago, Pines says, he had to "stretch" this thesis in conversations with investors. Now he puts it this way: "Not there yet, but we're moving."

Physics is not mathematics

AI is already taking over important work where the result can be expressed in verifiable code: in mathematics and computer science. In the life sciences and physics it is harder. Pines explains: "Nature also gets a vote in your models of reality." That is why PSI works on how to most efficiently check computational results with experiments.

The logic is this: AI proposes an idea, and if it is a real fact of nature, it can be exploited, up to building a device on its basis. PSI is building each such "hypothesis, verification, application" cycle. Others are working nearby: last week the startup Discovered Materials announced a $1 million competition for researchers applying AI to superconductivity and heat dissipation.

The standards haven't been invented yet

PSI held the result back until Anthropic's model was almost ready for release. In that time the solution was shown to specialist researchers. Jim Halverson of Northeastern University, a PSI advisor, says best practices are still taking shape, and that transparency and contact with experts seem a sensible path. Next week MIT will host a summit where leading physicists will discuss how to respond to the new pace of science.

After this, Turturean for the first time took an AI-generated solution and rewrote it so that it reads more naturally to a human: the scientific community expects him to be able to explain the result. In his words, that was when he truly felt the weight of what AI can do.

Takeaways

The limitations are worth keeping in mind. Lean confirms that a statement is true, but not that it is important. The result itself is narrow and negative. The information comes from a TechCrunch interview and Aaronson's blog, and we did not review the full proof. PSI's revenue so far comes not from discoveries but from applied simulations.

Three practical points follow. If you are building agents for science or engineering, the main value lies in the verifier and the harness, not only in the choice of model. Where there is a formal checker such as Lean or tests, AI is already winning; where there is none, the experiment becomes the bottleneck. And for founders, PSI shows a working scheme: earn money from applied simulation today while the big discoveries mature.