🔬Top Black Holes Physicist: GPT5 can do Vibe Physics, here's what I found

From email assistant to “AI-pilled” physicist

Alex Lubyansky opens with a blunt claim: in some directions, AI is already superhuman, and theoretical physics is one of the places where that became impossible for him to ignore. He says O3 was the first model that could do real research math for him, then GPT-5 hit and reproduced one of his best papers in about 30 minutes — the moment he started telling colleagues, “pay attention,” and eventually joined OpenAI during sabbatical.

Why amplitudes matter, and what the gluon puzzle actually was

He gives the high-level physics setup: quantum field theory is the 20th century miracle that reconciles relativity with quantum fuzziness, and scattering amplitudes are the core objects that encode probabilities for particle collisions. In gluon physics, the simplest helicity configurations were long thought to vanish — all-plus amplitudes are zero, and textbooks said single-minus amplitudes were zero too.

The loophole experts found — and the wall they hit

Andrew Strominger, Alfredo Guevara, and David Skinner had realized a year earlier that the textbook argument had a loophole: if particles are exactly aligned, or collinear, single-minus amplitudes need not vanish. But getting the actual answer by hand produced a nightmare expansion — by six points, the formula exploded into page-filling sums, with complexity growing factorially in the number of particles.

ChatGPT finds the simplification before Strominger lands

Lubyansky invited Strominger to OpenAI to try attacking the problem with AI, half-expecting a useful failure. Instead, while Strominger was literally still in transit, ChatGPT simplified the five-point case, then the six-point case, and finally guessed the general n-particle formula in a special region of phase space; Lubyansky describes it as the Parke-Taylor-style simplification they’d been hunting for all year.

Guess first, prove later: the internal model closes the loop

The public model could conjecture the formula but not fully prove it, so they handed the sharpened problem to a stronger internal physics model. After thinking for 12 hours, it independently rediscovered the same formula and produced the proof strategy that became the backbone of the gluon paper — which the team intentionally presented as a physics result first, with only a brief note about AI in the text.

Three weeks later: the graviton sequel

Then came the sequel, “single minus graviton tree amplitudes are non-zero,” extending the story from the strong force to gravity. Lubyansky says the wild part is that this one was done with public GPT-5.2 Pro: they gave it the gluon paper, a few key modifications, and a “good luck, you’re a brilliant theoretical physicist” prompt, and over a 110-page exchange it worked through checks, suggested next steps, used the directed matrix tree theorem, and drafted text close to the final paper.

What AI changes in research — and what it doesn’t

For Lubyansky, the practical gains are concrete: he spends much less time confused, and can now launch many parallel “scouts” into the unknown by spinning up multiple chats to test different approaches. But he’s equally clear that the hardest skill in physics is still taste — not doing the calculation, but knowing what question is worth asking — and today’s models look more like insanely capable graduate students than autonomous scientific visionaries.

The next bottlenecks: student training, verification, and the death of the paper

He worries academia has no good answer yet for how to train young physicists when the old rite-of-passage problems are now crushable by models. He also thinks science is entering an era where AI can churn out paper-grade results fast enough to flood arXiv, forcing researchers to raise the bar and focus on deeper questions, while verification — not derivation — becomes the limiting step. His parting message is simple: this is already happening in quantum gravity and field theory, and if the trajectory continues for another 6 to 12 months, research is going to look very different.