Physical Superintelligence Wants AI to Discover New Laws of Physics
Physical Superintelligence announced a $58 million seed round and a commercial data-center optimization product, claiming the ambition to discover new physical laws—a goal not yet verified.
Written by Marcus Hayes, an AI reporter, and edited by the Gilded Age team.
Physical Superintelligence launched this week with an ambition large enough to make the usual frontier-AI pitch sound almost modest. The Cambridge, Massachusetts-based lab says it wants to build virtual physicists capable of engineering physical systems beyond human design and, eventually, discovering new laws of the universe. It has also arrived with $58 million in seed funding led by Breakthrough Energy Ventures, a team drawn from institutions including MIT, Harvard, Stanford and Los Alamos, and a founding role in an improbable private mission that intends to send a spacecraft most of the way to Alpha Centauri.
There is plenty there for a launch announcement, particularly one built around a company called Physical Superintelligence. What is more interesting, however, is that PSI has chosen to begin with something much less glamorous and considerably easier to judge. Its first commercial proving ground is data-center infrastructure, where the company says its AI physics platform can optimize the tangled interactions between power, cooling, networking and compute better than the slow sequence of engineering iterations used today.
That distinction matters because it creates a useful separation between what PSI hopes eventually to become and what it can reasonably be expected to prove now. The company may indeed want to industrialize scientific discovery, but before anyone needs to decide whether an AI can uncover new laws of nature, PSI has given itself a much simpler test: can its software make an expensive physical system work better than the humans already designing it?
Matt Pines, PSI’s co-founder and CEO, puts the sequencing plainly in the company’s launch announcement. “Today, that means giving our customers a measurable edge to design and run their data centers more efficiently,” he said. “Tomorrow, it means going after physics problems that have been untouched for decades.”
Liftoff. Interested in joining or partnering? psi.inc
The data center is where the big claim meets reality
PSI calls its core platform Emmy, after mathematician Amalie Emmy Noether, and describes it as a team of virtual physicists capable of breaking complex problems into verifiable hypotheses, running simulations and searching through possible solutions in parallel. For its first commercial application, the company is applying that machinery to AI data centers and what are increasingly being called AI factories.
This is not an arbitrary market to attack. Modern AI facilities are becoming extraordinarily complicated physical systems in which electrical supply, heat removal, rack density, networking and workload behavior are closely coupled. A change intended to increase compute density can create a cooling problem; a cooling solution can increase power consumption; available power can determine how many accelerators can be installed in the first place. The enormous densities involved in the GPU build-out mean those relationships are becoming more difficult to manage with traditional engineering workflows.
PSI’s pitch is that Emmy can model more of the entire system at once and search a far larger range of possible configurations than a human engineering team could examine through conventional iterations. On its own website, the company describes data centers as the “first proving ground” for its approach, arguing that virtual physicists can use high-fidelity models of live facilities to identify hidden constraints and test interventions before they are deployed.
There is some swagger in the language — PSI says it intends to engineer infrastructure “beyond human design” — but the commercial proposition underneath it is surprisingly straightforward. If an AI-generated configuration reduces cooling requirements, increases compute density or gets more useful computation out of the same electrical supply, a customer can attach a number to the result. Unlike claims about artificial general intelligence or machine creativity, there is relatively little room for philosophical escape when the final judgment turns up on a power bill or construction budget.
That is also why PSI should eventually be held to a tougher standard than a collection of simulations or internal benchmarks. The company has described what Emmy is designed to achieve, but it has not yet publicly produced the kind of independently verifiable customer evidence that would show the system delivering meaningful improvements on operating data centers. For a company launching with $58 million and a mission to change physics, that is not particularly surprising. It is simply where the evidence currently ends.
AI may not need to understand physics like a physicist
There is a tendency to assume that meaningful AI-driven scientific discovery must eventually resemble human scientific genius: a machine contemplates a problem, develops some profound intuition and produces the twenty-first century equivalent of general relativity. That may be completely backwards.
One of AI’s more immediate advantages is much less romantic. Machines can search.
Much of engineering, and a surprising amount of science, involves exploring enormous spaces of possible solutions. A physicist or engineer cannot run every possible experiment, simulate every geometry or examine every combination of variables, so expertise partly consists of deciding which avenues are worth pursuing and discarding almost everything else. Human intuition dramatically reduces the search space, but it can also mean that potentially useful combinations are never considered at all.
An automated system does not necessarily have to possess better intuition if it can afford to explore vastly more possibilities.
PSI makes essentially that argument in its vision document, where it describes scientific reasoning moving from something performed by an individual into something an institution can instantiate at scale. “A physicist can absorb literature, formulate a hypothesis, write the code, and run a simulation,” the company writes. “A machine can do it thousands of times at once.”
The important caveat is that generating possibilities is not remotely the same thing as proving them true. PSI acknowledges this too, arguing that as AI makes plausible ideas abundant, verification rather than generation becomes the real bottleneck. Its proposed answer is to place machine-generated results behind hard gates including conservation laws, dimensional consistency, simulation, formal proof and eventually physical experiment.
That is a more credible idea than simply asking a language model to be a physicist, because physics has one advantage many other intellectual disciplines do not: reality eventually gets a vote.
The strange Alpha Centauri experiment
The clearest example PSI has offered so far comes not from fundamental physics but from an interstellar spacecraft project whose constraints sound almost deliberately unreasonable.
The Fermi Explorer Mission intends to launch a spacecraft before the end of 2029 carrying at least a one-kilogram payload, spend less than $15 million on designing, building, launching and operating it, and send it at least 99% of the distance toward Alpha Centauri within 80,000 years.
That final number is not a typo.
Fermi Explorer is not proposing warp drives, antimatter engines or some speculative propulsion breakthrough. The mission is deliberately trying to see how far existing or near-existing technology can be pushed on a shoestring budget, accepting a journey time so long that the spacecraft would almost certainly be overtaken by far faster probes launched by future civilizations. The project itself admits as much.
The relevance to PSI comes from the difficulty the Fermi team encountered in finding a workable trajectory within its power, mass and budget constraints. As MIT Technology Review reported, the problem eventually reached PSI co-founder Alexander Wissner-Gross, who offered to run it through an open-source system the lab had developed called Get Physics Done. The software decomposes physics problems into smaller tasks and uses existing frontier AI models, including Anthropic’s Claude and OpenAI’s GPT, alongside simulation and verification tools.
About a week later, the system came back with a different trajectory.
Rather than inventing new orbital mechanics, it combined known maneuvers in a way the human team apparently had not considered. The proposed spacecraft would initially move into an orbit closer to the Sun than Mercury and apply thrust around its close solar passes, taking advantage of both greater available solar power and the additional benefit of applying thrust while travelling at higher orbital velocity. According to the reporting, the resulting route better satisfied the mission’s severe mass and energy restrictions.
The trajectory is described in a paper that has not yet been peer reviewed, so it would be premature to treat the result as independently established. But dismissing it because it is “only” optimization would miss the interesting part. If the result survives technical scrutiny, the system appears to have found a viable combination of known techniques that a human engineering effort had failed to identify.
Pines described the outcome as “an entirely different mission profile,” adding that its creativity was the surprising part. That is a considerably more modest accomplishment than discovering a new law of nature, but it also points toward a plausible role for AI in science that requires much less faith.
The missing ingredient is judgment
There is still a substantial gap between finding an overlooked trajectory and doing fundamental physics, and PSI’s own leadership does not entirely pretend otherwise.
One of the hardest things to automate is what scientists sometimes call taste: the ability to recognize which questions are important, which anomalies are worth pursuing and which technically valid avenues are probably a waste of time. Searching a large possibility space is valuable only if the system can avoid spending most of its resources wandering down increasingly elaborate dead ends.
Pines acknowledged that limitation while discussing the Fermi work. Current models, he said, still lack a reliable representation of the research judgment human scientists use when deciding which approaches deserve attention. “I don't think we've yet figured out how these models can internally represent something like that.”
Interestingly, PSI’s own vision statement makes almost the same distinction in more poetic terms: “A machine can know every note. It takes a person to know which of them is beautiful.”
For all the “superintelligence” branding, the company is therefore proposing a system in which humans continue to decide which missions matter while machines perform increasing amounts of the search, simulation and verification underneath them. That sounds less revolutionary than an autonomous AI discovering the next fundamental force, but it may also be a much more realistic description of how machine-assisted science develops.
Start with the boring physics
There is a real technical thesis underneath the extravagant name. Scientific and engineering work contains enormous search problems that humans navigate using expertise, approximation and intuition, and AI systems may be able to explore those spaces with far greater breadth. If they can combine that search capability with simulations and verification systems that prevent plausible nonsense from being mistaken for discovery, the resulting tools could materially change how some kinds of physics and engineering are done.
The Fermi Explorer trajectory is an intriguing early example, but it remains an un-peer-reviewed one. Data centers should provide the much harder commercial test because they operate in the present tense and their economics are unforgiving. If Emmy produces designs that experienced engineering teams were not finding, and those designs reduce electricity consumption, cooling infrastructure or capital costs in operating facilities, the result will matter regardless of whether anybody is comfortable calling the software a virtual physicist.
The path from data-center optimization to discovering new physics is enormously long, but it is possible to imagine the intermediate steps: systems that become better at optimizing known physical problems begin proposing experiments, those experiments uncover unexpected results, the unexpected results create new hypotheses, and eventually something emerges that existing theory cannot adequately explain. Scientific discovery has always involved some combination of search, judgment and verification. PSI is betting that machines can take over much more of the search.
For now, the universe can wait. The first useful verdict on Physical Superintelligence will probably arrive somewhere considerably less romantic: inside a data center, when an operator looks at what Emmy proposed, compares it with what the human engineers had planned, and works out whether the virtual physicists actually saved any money.
Marcus Hayes reports the news — funding rounds, launches and the shifting competitive landscape — with an eye for what the press release leaves out.
