Vibe physics needs a referee

618 words

For Travis Kalanick. Everyone else can read it too.

Travis, I don't think vibe physics is crazy. I think it's missing a referee.

Back in July 2025 you went on All-In and said you'd been pushing GPT and Grok out to the edge of quantum physics, "vibe physics." The internet had a good laugh. I didn't, because I do basically the same thing with math every week, and the laughing kind of misses what's actually broken.

Amateurs asking questions at the frontier is fine, that's how half of science started. What's broken is that the chatbot will tell you you're onto something whether you are or not. It can't tell its own insight from its own hallucination. You can't either, and the physicist who could isn't picking up at 2am.

In math this is close to solved. I write proofs in Lean, a programming language where the compiler only accepts a theorem if every single step checks. So when a model hands me an argument I don't have to trust it or go find a referee, I just try to make it compile. Compiles, it's proved. Doesn't, and the error tells you which step is bogus. The only thing left to argue about is whether the statement says what you meant, and that's a way shorter argument. The model vibes, Lean referees.

What I'd build

Physics is harder because it's about the world and not just definitions. But a lot of what feels like a breakthrough at the edge of QM is secretly a math claim. This limit exists. This approximation holds. This thing is conserved. Those you can check.

What I'm actually working on: get the physics at one scale out of the physics underneath it, mechanically. Start from a microscopic model (my test case is getting approximate Maxwell out of effective QED, with the cutoffs left as parameters), let the AI make the judgment calls about what's small enough to throw away, and let Lean keep the books. The right language for "small enough to throw away" is nonstandard analysis, actual infinitesimals done rigorously, which is my home turf. I don't know if it works yet. It needs a few people and a lot of GPUs.

Also robots

Atoms puts robots in mines. A bug in a robot hauling tons of rock isn't a stack trace, it's a crater. The same tools that check a proof can check a controller: the arithmetic never overflows, the planner never asks for a pose outside the safe set. Lean for robotics is wide open right now, the formal methods people are all busy with compilers. And geometric algebra, which I gave a talk on, is the cleanest language I know for rigid motion, which is exactly the kind of thing a type checker can hold you to.

Me

Mathematician, Berkeley. I work in Lean 4 and nonstandard analysis, and on AI safety. I wrote FormalWeaver with Hantao Lou and Max Tegmark, about having language models translate entire formal repositories. I'm one of the authors of the vericoding benchmark (formally verified program synthesis in Lean, Dafny and Verus). I took a robot dog to the UN with Max to scare some politicians about what's coming. I help run the SF Lean Meetup. More on the papers page and in the archive.

The offer

Send me the vibe-physics thread you were most sure about. I'll turn the central claim into a Lean statement and we'll see which parts hold up. If some of it does, cool, that's the start of something. If none of it does, you find out in a week instead of never.

alokbeniwal@gmail.com, or the note box at the bottom of this page.

Alok