“If you took an LLM and gave it a lobotomy, took everything about the physical world, trained it only on stuff before 1900, could it come up with quantum mechanics and relativity? We tried giving it just the orbital history of the planet Mercury from JPL. It couldn't derive the Riemann curvature or the Ricci tensor. It matched the planet, but it was just copying. It was just mimicking.”

I'm trying to do what I call the Einstein test. If you take LLM now, if you take Claude or OpenAI or whatever, and you say, tell me about the multiverse or tell me about this and that, it knows so much, not just from the physics, but also from the culture. In other words, there's movies about wormholes like we just talked about. There's movies about even honking and honking radiation, right? So you can't even trust the training data, which is a large fraction of what makes these things unique. So we tried asking the following quote. What if you gave it an LLM lobotomy? What if you took everything about the physical world, just trained it on stuff before 1900? Could it come up with quantum mechanics and relativity? And of course, there are other people working on this and there are people that have actually made these LLMs and actually denatured them, denuded them of any pop culture references after 1930. But that's even too late for quantum mechanics, right? And relativity. And really, we try to do it with just giving it data from the planet Mercury and just this orbital history over thousands of years that JPL has. And it gave us better. Back, you know, it couldn't…