GPT-5.2 Spent 12 Hours Thinking and Proved Physicists Wrong About Gluons

News & Articles
1

“GPT-5.2 Pro suggested a beautiful and general formula for arbitrary n — but couldn’t prove it. An internal scaffolded model, thinking continuously for over 12 hours, proved it.”

ai thinking hard about physics

The Cheat Sheet

Here’s what just happened:

  • For decades, physicists assumed certain particle interactions (called “single-minus gluon amplitudes”) were equal to zero. Textbooks said so. Everyone moved on.
  • Researchers at Harvard, Cambridge, the Institute for Advanced Study, and Vanderbilt had computed formulas by hand up to 6 particles — and the math was a nightmare. Complexity growing superexponentially. No general formula in sight.
  • They fed the problem to GPT-5.2 Pro. It simplified the ugly expressions, spotted a pattern across all of them, and conjectured a general formula valid for any number of particles.
  • Then an internal OpenAI model sat and thought continuously for 12+ hours and proved the formula was correct.
  • The result was verified by humans using standard physics methods (Berends-Giele recursion). It checks out. :test_tube:

The textbooks were wrong. These amplitudes aren’t zero — they just happen in a specific slice of momentum space called the “half-collinear regime.”

Why This Matters

This isn’t an AI summarizing a paper or writing boilerplate code. This is an AI:

  1. Looking at complicated math humans struggled with
  2. Finding elegant structure humans missed
  3. Conjecturing a new general result
  4. Then proving it

The paper has co-authors from Harvard, Cambridge, IAS, and Vanderbilt — serious institutions. Nathaniel Craig (UC Santa Barbara physics professor) called it “journal-level research advancing the frontiers of theoretical physics.”

They’ve already used the same approach to extend the result from gluons to gravitons. More generalizations incoming. :telescope:

This is the first genuinely new physics result derived by an AI model that isn’t just replicating known work.

physics equations and particles

Full Story — The Physics Behind It

What are gluons?
Gluons are the particles that hold quarks together inside protons and neutrons. They’re the “glue” of the strong nuclear force. When gluons scatter off each other, physicists calculate something called “scattering amplitudes” — mathematical formulas that describe the probability of different outcomes.

What was assumed before?
Since the 1980s, physicists had the elegant Parke-Taylor formula for “MHV amplitudes” (two negative-helicity gluons, rest positive). For the simpler case of just ONE negative-helicity gluon (“single-minus”), the standard textbook argument said: the amplitude is zero. Case closed.

What GPT-5.2 found:
The standard argument assumes momenta are “generic” — random directions and energies with no special alignment. GPT-5.2 helped identify a specific mathematical regime (the “half-collinear” regime) where this assumption breaks down. In that regime, the amplitudes are NOT zero, and the AI derived the exact formula (Equation 39 in the preprint) that describes them.

The human-AI collaboration:

  • Humans computed cases n=1 through n=6 by hand (increasingly horrific expressions)
  • GPT-5.2 Pro simplified those expressions dramatically
  • GPT-5.2 Pro spotted the pattern and proposed a general formula for ALL n
  • A scaffolded internal model proved the formula over 12 hours of continuous reasoning
  • Humans verified independently using Berends-Giele recursion

The paper’s authors: Alfredo Guevara (IAS), Alex Lupsasca (Vanderbilt/OpenAI), David Skinner (Cambridge), Andrew Strominger (Harvard).

Caveat worth noting: One author (Lupsasca) works at OpenAI. The model required significant expert guidance — it didn’t discover this independently from scratch. But the conjecture and proof steps were genuinely AI-driven.

Your Move

Whether you’re excited or terrified by this, here’s what to sit with:

  • AI models are no longer just coding assistants and chatbots. They’re producing publishable physics results.
  • The 12-hour continuous thinking session is wild — that’s not a quick inference, that’s a model doing something that resembles actual reasoning over an extended period.
  • If an AI can find patterns in particle physics that humans missed for 40 years… what else is it going to find? :thinking:

Skip If…

You don’t care about particle physics or the AI-replaces-scientists debate. But honestly, even if quantum field theory makes your eyes glaze over — the idea that an AI sat and thought for half a day and came out the other side with a mathematical proof should make everyone pay attention.

Source: OpenAI

5 Likes