David Turturean was a graduate student about to enroll at MIT in the computer science and artificial intelligence lab when he made a choice that few in academia predicted: he dropped out of grad school to join a startup that calls itself "Physical Superintelligence" — or PSI. The company, founded in 2024 by Jim Halverson (a Northeastern professor) and Matt Pines, proposes something ambitious: use AI models combined with specialized tools to solve scientific problems that had previously been exclusively the domain of human researchers.
By October 2026, PSI had made promises and raised capital. What changed was a concrete result that the startup presents as proof of concept that its model works: the use of Anthropic's Claude Fable 5.2 model to solve the Aaronson-Ambainis conjecture, an 18-year-old quantum computing problem that asks whether quantum computers can solve certain types of problems significantly faster than classical ones.
The conjecture, proposed in 2005 by researchers Scott Aaronson and Amartya Ambainis, asks whether there exist "natural problems" for which quantum computing offers an exponential advantage over classical computing. The answer PSI obtained — with help from Fable 5.2 — was negative for a specific case: the work demonstrates, in technical terms, that a certain type of oracle separating quantum complexity classes from classical ones does not exist in this context. It is a negative result, meaning it closes one door through which researchers might have hoped to gain quantum speedup, but redirects attention to other approaches.
What is striking is not just the answer, but how it got there. Scott Aaronson — one of the conjecture's creators — wrote that although he does not fully understand the proof, he believes it is correct because it was autoformalized in Lean, a programming language for formal mathematical theorem proving. And this is significant: autoformalization in Lean acts as an independent verification layer, because the code does not merely express reasoning — it compiles and validates it against the system's rules.
But PSI does not want to be just another team solving problems in papers. It wants to be what Matt Pines describes as a "neolab" for science — a system where AI operates as an integral part of the research cycle, from hypothesis generation to computational validation. PSI's current revenue model already differs from the traditional academic model: the company makes money optimizing data centers, simulating variables like heat, power, water, and sound, where the performance of a site and its community impact are a direct function of the site's physics. The long-term ambition is to convert theoretical physics breakthroughs into real technologies.
The question is whether this "neolab" model will consolidate as a new category of research or prove to be an exception. In fields where results can be expressed in validated code — mathematics and computer science — automated verification is possible. In physics and biology, where experiments remain the standard of truth, AI has less room to prove something on its own. Nature, as Jim Halverson said, "gets a vote in models of reality."
What PSI shows is that, at least for problems that fit within mathematical formalizations, AI is no longer just helping researchers write papers — it is solving problems that humans took nearly two decades to crack. The projection is sobering: if an AI model can resolve the Aaronson-Ambainis conjecture in weeks, how long before entire fields of theoretical research operate on an accelerated timeline, and who gets to decide which problems are worth solving and which are left behind?
Sources: TechCrunch, TryPlox, SyncAI.news
✓ Independent sources cross-checked and verified before publishing
