Physicist Alexey Semikhatov has a date in mind: 1 May 1925. Semikhatov is a Doctor of Physical and Mathematical Sciences and chief researcher in theoretical physics at the Lebedev Physical Institute. Quantum mechanics did not exist yet on that date. It was a few weeks away.
His proposal is to build an artificial intelligence that knows exactly what physicists knew then and nothing after. It would get the papers and results that existed by that spring, and no memoir, textbook or web page written later. Then you ask it to invent quantum mechanics.
Why bother, when we already have the theory? Because physics is running out of ways to test its guesses. The extensions of the Standard Model that Semikhatov finds elegant number "three, maybe five," and nothing we can build tells them apart. Their particles are too heavy for any accelerator we have. A collider might have to run around the Earth's equator, and even that is no promise. Gravity is worse: to see quantum gravity at work you would have to hurl black holes of a few solar masses at each other at half the speed of light.
So the question is where new theories can come from when new data is scarce. Semikhatov's bet is that a machine rerunning 1925 might show us. If it can find quantum mechanics from the old physics, the same method could be pointed at the physics we have now.
What would count as success? A machine like this, he says, hunts for correlations between things. It might come back with Heisenberg's version of quantum mechanics, or Schrödinger's, or Feynman's path integral. Or it might land at once on something deeper than all three.
The hard part is the cleaning. Semikhatov expects such training to be expensive, because the 1925 cutoff has to be airtight. The internet is full of what came later. A physicist in the 1950s recalling how he made quantum mechanics in the 1920s would poison the experiment. And the genuinely old material is often not digitised at all. You have to rebuild what people actually knew on that spring day, no more.
He does not promise it works. Maybe someone will do it someday, he says, and maybe it will turn out to be a way of making physical theories we can't yet guess at. Then he adds the other half at once: maybe the thing can't be trained at all.
Maybe someday someone will do it, and maybe it will be a way of creating physical theories that we can't guess at now.
Может быть, когда-нибудь кто-нибудь такое сделает, и, может быть, это будет способ создания физических теорий, до которых мы сейчас не можем догадаться.
— Alexey Semikhatov, Gleb Solomin Podcast, 46:30
Behind the test sits a harder question, and Semikhatov says it lies outside this conversation: what a physical theory even is. His answer is a guess that has predictive power. Newton guessed the law of universal gravitation. It follows from nothing, yet the apple falls and the Moon keeps flying. Only Mercury moves a little differently, and it took Einstein to explain that.
Machines have already shown a taste for this work. In April 2025 You-Le Fang and colleagues posted a paper on arXiv about AI-Newton. Given a large, noisy set of mechanics experiments, it rediscovered Newton's second law, the conservation of energy and universal gravitation. It started from measurements, though. Semikhatov's version starts from the state of physics on one particular spring day, then waits to see whether the machine makes the leap the physicists made that year.
If it does, which quantum mechanics will it write down? The one in our textbooks, or one nobody has seen?
