David Turturean, a recent MIT graduate and member of the technical staff at 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 fully embrace the idea that this may be the last paper I publish,” Tertullian, who postponed plans to earn a master’s degree from MIT’s prestigious computer science and AI lab to join PSI, told TechCrunch. “Why do I need to publish a paper in the next few months?”
The problem in question, the Aaronson-Ambainis conjecture, concerns the ability of quantum computers to solve problems faster than classical computers. Scott Aaronson, one of the two researchers who raised the issue, wrote that although he doesn’t quite understand the proof, he believes it to be correct because it was automatically formalized in Lean, a coding language for mathematics, and explained by a team of leading researchers with whom PSI shared the solution.
This solution appears to close the door on one path quantum computers can take to quickly solve certain difficult problems, but not all of them. “This was a negative result and it allowed people to know, okay, we’re not looking at that anymore,” Tertullian said.
More broadly, it reflects the upheaval in the scientific world caused by AI problem solving.
“Everyone seemed to agree that the proof of the complete[Aaronson-Ambainis conjecture]was one of the last remaining open questions in the complexity of quantum queries, and one of the most important,” wrote Aaronson, who felt satisfied, awed, disappointed, and exhausted by this news. “I was honestly surprised to see it missing from OpenAI’s Mathocalypse dumps a few weeks ago. Sure they tried it, but was it actually that much harder than all the problems they solved?”
This supports PSI’s model of using specialized model harnesses and scaffolds to solve scientific problems.
“An effective research system has to have a lot to do with harnesses, a lot to do with tools, memory systems, and simulation infrastructure,” PSI CEO Matt Pines told TechCrunch. “We want to be able to explore problems broadly and deeply and consider a wide range of potential alternatives.”
For now, the company makes money by using tools to optimize data centers and simulate variables such as heat, power, water, and sound. “The site’s performance and community impact is a direct function of its physics.” But it takes a long time to turn breakthroughs in theoretical physics into real technology.
“Six to nine months ago, when I was talking to investors, you had to stretch your legs a little bit to test those terms, right? They were like, ‘No, no, believe me, very soon theoretical physics will be primarily a mechanical domain,'” Pines said. “And that hasn’t happened yet, but it’s on its way.”
Still, while AI has claimed important solutions in fields like mathematics and computer science where it can be expressed in verified code, it has proven more difficult to achieve in less abstract fields like life sciences and physics. Last week, startup Discovered Materials announced a grand $1 million challenge for researchers to use AI to understand problems like superconductivity and heat dissipation.
“At the end of the day, physics is different from mathematics because physics is the real world…In the real world model, nature has the vote,” Pines said. He said the startup is focused on understanding how to take computational results and verify them through experiments in the most efficient way possible.
“If it’s a real fact of nature, it’s a trick that can be exploited,” Pines says. “You can build devices out of it, right? AI suggests ideas that allow you to do things in the world that you couldn’t do before. So each of them is like a loop that PSI is building.”
While the NeoLab for Science is a path to potentially breakthrough science, participants are still figuring out what the norm is in this new world. PSI needed to hold on to the results until Anthropic’s model was almost ready for release so it could share them with relevant researchers for deeper understanding.
“As academics, we wanted to play a role in doing things the right way,” says Northeastern University professor Jim Halverson, who serves as an advisor to PSI. He is scheduled to attend a summit at the Massachusetts Institute of Technology (MIT) next week, where leading physicists will formulate responses to the new pace of science. “Best practices are still emerging…We thought transparency was a good way to go and reaching out to experts was a good way to go.”
“We had some time to think about it… and it was the first time I had actually printed out and read one of the solutions derived by AI. And as I read it, I thought, how can we express this more naturally to humans?” Tertullian told TechCrunch that the academic community is looking forward to him being able to explain the results. “And I think I realized the importance of what AI can do.”
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
