PSI wants to turn AI-driven scientific discoveries into a business

Startup Physical Superintelligence (PSI) used early access to Anthropic’s Fable 5.2 model to solve an 18-year-old problem in quantum computing. Recent MIT graduate David Turturian, a member of PSI’s technical team, helped prove the Aaronson–Ambainis conjecture. The company makes money by optimizing data centers, but hopes to turn theoretical physics discoveries into working technologies in the future.

What the proof establishes

The Aaronson–Ambainis conjecture concerns how much faster quantum computers can solve problems than classical computers. Scott Aaronson, one of the two researchers who formulated the problem, acknowledged that he does not yet fully understand the proof. Still, he believes it is correct: it was automatically formalized in Lean, a language designed for mathematical proofs, and then explained by a group of leading researchers with whom PSI shared the result.

The proof appears to close off one possible way of using quantum computers to solve certain hard problems quickly, but it does not rule out other approaches. Turturian called the result negative: researchers now know they no longer need to investigate this path.

He put off plans to enter MIT’s Computer Science and Artificial Intelligence Laboratory graduate program to join PSI. Speaking to TechCrunch, Turturian said this might be his last scientific paper: in a few months, he said, publications may no longer serve the same purpose.

Aaronson called the proof one of the last unresolved problems in quantum query complexity, whose importance many specialists recognized. He wrote that the news left him feeling satisfied, amazed, disappointed and tired at the same time. Aaronson was also surprised that the problem was not included in OpenAI’s Mathocalypse set, published a few weeks ago. He assumed the problem had been attempted and wondered why it had proved harder than the problems that made the set.

How PSI is building a research system

For PSI, the result supports an approach in which specialized models are paired with tools and other resources for tackling scientific problems. CEO Matt Pines believes an effective research system needs tools for working on problems, memory systems and infrastructure for simulations. It should make it possible to explore a problem broadly and deeply, considering many possible solutions.

For now, PSI earns revenue by using its tools to optimize data centers. The company models heat output, energy consumption, water use and noise — factors that affect both a data center’s operations and its impact on the local community.

Over time, PSI wants to turn theoretical physics results into technologies. Pines said that just six to nine months ago, he had to convince investors that theoretical physics would soon become primarily a machine-driven field. In his view, that future has not arrived yet, but it is approaching.

AI has already helped produce important results in mathematics and computer science — fields where proofs can be expressed in code and verified. That is harder in less abstract disciplines such as physics and the life sciences. Last week, startup Discovered Materials announced a competition with a $1 million prize fund, challenging participants to apply AI to problems such as superconductivity and heat dissipation.

Pines explains the difference by pointing out that physics deals with the real world: nature can confirm or disprove a model. That is why PSI is studying how best to verify computational results through experiments. If AI identifies a valid pattern in nature, it can be used to build a device; the company is developing workflows that connect a model’s ideas to actions in the world.

How the rules of scientific work are changing

AI-powered research labs are becoming a possible route to new discoveries, but the rules for this kind of work are still taking shape. PSI had to delay publishing its result until Anthropic’s model was nearly ready for release. In the meantime, the team was able to show the solution to researchers in the field and develop a better understanding of it.

Jim Halverson, a professor at Northeastern University and an adviser to PSI, said the company wanted to follow academic norms. Next week, he plans to attend a summit at MIT where leading physicists will discuss how to respond to the acceleration of scientific work. Halverson said best practices are still emerging; the PSI team found openness and consulting experts useful.

Turturian said that during the wait, he printed out one of the solutions proposed by AI for the first time and read it, trying to express the result in more natural language. He understands that the scientific community will expect him to be able to explain the proof, and acknowledged that the work helped him grasp the scale of AI’s capabilities.

Dataist’s take: PSI’s result shows that AI can already contribute to solving problems that can be checked using mathematical methods. But applying this approach to physics requires experiments: computation alone does not confirm that a proposed pattern holds in nature.