PSI Is Trying to Turn AI-Driven Scientific Discovery Into a Business
A Recent MIT Graduate Solves an 18-Year-Old Quantum Problem
David Turturean, a recent MIT graduate and technical staff member at the startup lab Physical Superintelligence (PSI), used early access to Anthropic's Fable 5.2 model to solve an 18-year-old problem in quantum computing.
"I'm totally accepting the idea that this might be the last paper I ever put out," Turturean, who deferred plans to pursue a master's at MIT's prestigious Computer Science and Artificial Intelligence Laboratory to join PSI, told TechCrunch. "In a few months, like, why would you even put out a paper?"
The problem in question—the Aaronson-Ambainis conjecture—concerns the ability of quantum computers to solve problems faster than classical ones. Scott Aaronson, one of the two researchers who posed the problem, writes that he doesn't quite understand the proof but believes it is correct because it was autoformalized in Lean, a coding language for mathematics, and explained by a team of top researchers with whom PSI shared the solution.
The solution appears to close the door on one avenue quantum computers could follow to solve certain difficult problems at high speed—but not all of them. "It's a negative result that allows people to know, okay, we're no longer looking there," Turturean said.
A Broader Upheaval in Scientific Problem-Solving
More broadly, the result reflects the upheaval in the scientific world caused by AI-driven problem-solving.
"A proof of the full [Aaronson-Ambainis conjecture] was one of the last remaining open problems in quantum query complexity that everyone seemed to agree was centrally important," Aaronson wrote, describing himself as feeling satisfied, awestruck, disappointed, and fatigued by the news. "I was surprised, honestly, to see it missing from 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?"
That is a point in favor of PSI's model, which is to solve scientific problems with specialized model harnesses and scaffolding.
The Harness-and-Scaffolding Approach
"An effective kind of research system should have lots of things related to harnesses, lots related to tools, memory systems, simulation infrastructure," PSI CEO Matt Pines told TechCrunch. "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."
For now, the company's revenue comes from using its tools to optimize data centers, simulating variables such as heat, power, water, and sound—"where the performance of that site as well as its impact on the community are a direct function of its physics." The long-term play, however, is converting theoretical physics breakthroughs into actual technologies.
"Six to nine months ago, when I was talking to investors, I had to sort of stretch to validate that conditional, right? Like, 'No, no, no, trust me, pretty soon, theoretical physics is going to be mostly the domain of machines,'" Pines said. "And it's not there yet, but it's on its way."
The Limits of AI in Less Abstracted Disciplines
Still, while AI has claimed important solutions in fields like mathematics and computer science—where problems can be expressed in validated code—doing so in less abstracted disciplines such as the life sciences or physics has proven more difficult. Last week, the startup Discovered Materials released a $1 million grand challenge for researchers using AI to understand problems like superconductivity and heat dissipation.
"At the end of the day, physics is different than math because it's the real world," Pines said.