AivexaNewsSearch
AI news for builders and product teamsChecked every hour

PSI is trying to make a business out of AI-driven scientific discovery

Collected Nov 10, 2026

David Turturean, a recent MIT graduate and technical staff member at the startup lab Physical Superintelligence (PSI), used early access to an unreleased Anthropic model called Fable 5.2 to solve the Aaronson-Ambainis conjecture, an 18-year-old open problem in quantum computing. The result was shared with outside researchers before publication, and Scott Aaronson, one of the two researchers who originally posed the problem, wrote that he does not fully understand the proof but believes it is correct because it was autoformalized in Lean, a programming language used for formal mathematics, and explained by a team of top researchers with whom PSI shared the solution.

The conjecture concerns whether quantum computers can solve certain problems faster than classical ones. The solution appears to rule out one avenue quantum computers might have used to achieve high-speed solutions for difficult problems, though not all of them. Turturean described it as a negative result that lets people know they no longer need to look in that direction. Aaronson wrote that a full proof was one of the last remaining open problems in quantum query complexity that everyone seemed to agree was centrally important. He said he felt satisfied, awestruck, disappointed and fatigued by the news, and expressed surprise that the problem had been missing from OpenAI's Mathocalypse dump a few weeks earlier, wondering whether it had been attempted and found harder than the problems that were solved.

Turturean told TechCrunch that he had deferred plans to pursue a master's at MIT's Computer Science and Artificial Intelligence Laboratory to join PSI. He said he is fully accepting the idea that this might be the last paper he ever publishes, and wondered why anyone would put out a paper in a few months. He also described taking one of his AI-elicited solutions to print for the first time and reading through it to make it present more naturally for a human audience, realizing the weight of what AI can do. The academic community will expect him to be able to explain the result.

The result supports PSI's model of attacking scientific problems with specialized model harnesses and scaffolding. CEO Matt Pines told TechCrunch that an effective research system should have many components related to harnesses, tools, memory systems and simulation infrastructure, so that it can pursue a problem at breadth and depth and consider a wide range of alternatives. In practice, a harness is the surrounding software that feeds a model context, tools and feedback loops; scaffolding is the structured workflow that lets a model plan, execute and check multi-step work. This suggests PSI is betting that raw model capability is not enough and that the engineering around the model is where scientific discovery becomes repeatable.

PSI's current revenue comes from using its tools to optimize data centers by simulating variables such as heat, power, water and sound, where site performance and community impact are a direct function of physics. The long-term play is converting theoretical physics breakthroughs into actual technologies. Pines said that six to nine months ago he had to stretch to validate to investors the idea that theoretical physics would soon be mostly the domain of machines, and that while it is not there yet, it is on its way.

AI has claimed important solutions in fields such as mathematics and computer science where results can be expressed in validated code, but doing so in less abstracted disciplines such as the life sciences or physics has proven more difficult. Pines said physics is different from math because nature gets a vote in models of reality, and that PSI is focused on understanding how to take computational results and validate them with experiments as efficiently as possible. He added that if something is a true fact of nature, it can be exploited as a trick to build a device, and that each AI-suggested idea that enables something new in the world is a loop PSI is building. Last week, the startup Discovered Materials released a $1 million grand challenge for researchers using AI to understand problems such as superconductivity and heat dissipation, a sign that others are also trying to organize work around AI-driven physical science.

As neolabs for science become a pathway for potentially groundbreaking research, participants are still working out standards. Because PSI had to hold its result until Anthropic's model was nearly ready for release, the team had time to share it with relevant researchers and better understand it. Jim Halverson, a professor at Northeastern University who advises PSI, said that as an academic he wanted to play a role in doing things the right way and that best practices are still emerging, with transparency and contacting experts seen as good approaches. He will attend a summit at MIT next week where leading physicists will try to hash out a response to the new pace of science. A likely trade-off is that holding results for a model release, while allowing expert review, can delay public scrutiny and credit allocation; this is my own inference from the stated facts.

Why it matters: Developers and product teams building AI systems for research should note that the result came from a specialized harness and scaffolding around an unreleased Anthropic model, not from a standalone prompt, which suggests investment in tooling, memory and validation loops is where differentiation may lie. The Lean autoformalization also points to formal verification as a practical way to make AI-generated proofs checkable, a pattern that could extend to other domains where results can be encoded. For product teams, PSI's path from data center optimization to physics-based technology shows a near-term revenue bridge while the longer scientific bet matures, though whether AI can reliably produce validated discoveries in physics and the life sciences remains an open question. It is my inference that the arrival of AI-assisted proofs will push research organizations to define review and credit norms sooner rather than later.

Read at TechCrunch · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

PSI used an unreleased Anthropic model to answer a fundamental question about quantum computing that scientists have pursued for nearly two decades.