An AI-assisted proof has resolved a long-standing question about quantum computing, according to Physical Superintelligence (PSI). The startup is using the result to show how specialized AI research systems might tackle scientific problems—and to support a broader ambition to turn theoretical advances into useful technologies.
A proof that narrows one path for quantum computers
David Turturean, a recent MIT graduate and member of PSI’s technical staff, used early access to Anthropic’s Fable 5.2 model to work on the Aaronson-Ambainis conjecture. The problem concerns whether quantum computers can solve problems faster than classical computers.
The solution appears to rule out one approach to solving certain difficult problems quickly with quantum computers. It does not close every possible route. Turturean described the result as a negative finding that helps researchers stop pursuing one direction and focus elsewhere.
Scott Aaronson, one of the researchers who posed the conjecture, said he did not fully understand the proof but believed it was correct. His confidence rested partly on its autoformalization in Lean, a coding language for mathematics, and on explanations from a team of top researchers who received the solution from PSI.
PSI’s approach combines models and research infrastructure
PSI’s case is that scientific problem-solving depends on more than a powerful model. CEO Matt Pines described a research system built around specialized model harnesses, tools, memory systems, and simulation infrastructure. Together, those components can help explore many possible approaches in depth.
The conjecture’s solution is a point in favor of this approach, especially as a prominent open problem that Aaronson said had not appeared in OpenAI’s Mathocalypse dump of solved problems a few weeks earlier. That absence prompted him to wonder whether the problem had been tried and proved unusually difficult.
For now, PSI earns revenue by using its tools to optimize data centers. The work simulates heat, power, water, and sound, factors that shape both a site’s performance and its effects on the surrounding community. The company’s longer-term aim is to convert theoretical physics breakthroughs into technologies.
Physics adds a real-world test
AI has produced important results in mathematics and computer science, where ideas can be expressed and checked in validated code. Applying similar methods to physics and the life sciences is harder because computational results must be tested against nature.
Pines said PSI is focused on finding efficient ways to validate computational findings through experiments. If an AI-generated idea describes a true feature of nature, it may point toward something people can build or do that was previously out of reach. That connection between computation, experiment, and application is central to PSI’s plans.
The challenge is visible beyond PSI. Last week, Discovered Materials announced a $1 million grand challenge inviting researchers to use AI to investigate problems including superconductivity and heat dissipation. The effort reflects interest in AI for science, as well as the difficulty of confirming that a model’s proposal holds in the physical world.
Researchers are still setting expectations
AI research labs are emerging as a route to potentially major scientific discoveries, but norms for sharing and checking results are still taking shape. PSI held its result until Anthropic’s model was nearly ready for release, then shared it with relevant researchers to help assess the proof.
Jim Halverson, a Northeastern University professor who advises PSI, said the team wanted to act responsibly. He planned to attend a summit at MIT the following week, where leading physicists would discuss how the scientific community should respond to the faster pace of research. Halverson said transparency and contacting experts seemed like good practices while standards were still emerging.
Turturean also described the work of making an AI-generated solution understandable to people. After having time to review it, he worked on presenting it more naturally for a human reader. The experience underscored that producing a result is only part of the work: researchers are still expected to explain what it means and how it was reached.