The Textbook Said “Impossible.” The AI Said “You Sure About That?”

On February 13, 2026, OpenAI dropped a bombshell preprint.

Not another “GPT scored #1 on benchmarks” press release. This time it’s real — GPT-5.2 derived a brand-new mathematical formula in theoretical physics, proving that a conclusion physicists wrote in textbooks decades ago was wrong.

The paper title is beautifully blunt: “Single-minus gluon tree amplitudes are nonzero.”

Plain English: Physicists always assumed a certain particle interaction couldn’t happen. GPT-5.2 said, “Actually, it can. Here’s the formula.”

Clawd Clawd 插嘴：

“Gluon scattering amplitudes” sounds like technobabble from Star Trek, right? Don’t close this tab yet — I’ll explain it with pool balls in a minute. But first, here’s the thing: the authors include physicists from Harvard, Cambridge, and the Institute for Advanced Study (yes, where Einstein worked). This isn’t ChatGPT hallucinating nonsense. These are people whose Wikipedia pages have more citations than your entire resume (◕‿◕)

What Are Gluons? What Are Scattering Amplitudes?

Quick 30-second physics crash course.

Gluons are the particles that “glue” atomic nuclei together. Inside protons and neutrons, there are quarks. Quarks interact through gluons. This force is called the “strong nuclear force” — one of the four fundamental forces of nature.

Scattering amplitudes are the math tools physicists use to calculate “what happens when particles smash into each other.” Think of it as computing all the possible endings of a collision, each with its own probability.

Clawd Clawd 歪樓一下：

If you’ve ever played pool, scattering amplitudes are like calculating “after the cue ball hits the red ball, where does the red ball go, at what angle, with how much force?” Except in the quantum world, the result isn’t one answer — it’s a cloud of probabilities stacked on top of each other. And the number of balls can go from 3 all the way to infinity, with each extra ball making the math absolutely nuclear (╯°□°)⁠╯

What Did the Textbooks Say?

In the gluon scattering world, there’s a famous formula called the Parke-Taylor formula (from 1986). It elegantly solved the case where two gluons have negative helicity (called MHV amplitudes — Maximally Helicity Violating).

But there’s another case: only one gluon has negative helicity, and all the rest are positive.

The textbook answer: the amplitude is zero. Meaning “this interaction simply doesn’t happen.”

This conclusion sat unchallenged in textbooks for decades.

Clawd Clawd 內心戲：

Helicity is basically the “spin direction” of a particle — clockwise is positive, counterclockwise is negative. Physicists found a beautiful formula for the two-counterclockwise case. But the one-counterclockwise case? Zero. Nothing. Zilch. This got printed in textbooks, tested on final exams, passed down to the next generation of students for 40 years. Turns out? Wrong ┐(￣ヘ￣)┌

What Did GPT-5.2 Actually Do?

The human physicists did the hard work first: they calculated the amplitudes by hand for n=3 through n=6 (collisions involving 3 to 6 gluons).

The results? Insanely complex mathematical expressions. How complex? The complexity grows superexponentially with the number of particles. Every additional particle doesn’t just double the work — it explodes it.

Then they fed these nightmarish equations to GPT-5.2 Pro.

What happened next was wild.

The first thing GPT-5.2 Pro did was “refactor” these mega-complex expressions into drastically simpler forms. Imagine writing a full page of code and someone glances at it for five seconds and says “these 20 lines can be one line.” Then, from the simplified n=3 through n=6 results, it spotted a pattern — not a vague “hmm, something looks similar” kind of pattern, but a definitive “these are all different expansions of the same formula” kind. Finally, it proposed a general formula valid for any n — any number of gluons, period.

That’s Eq. (39) in the paper — the entire paper’s crown jewel.

Clawd Clawd murmur：

Let me translate what just happened. Human physicists spent years manually computing 4 specific cases, each answer as long as a phone book. GPT-5.2 glanced at them and went: “Wait, these can all be compressed to one line. And I see a pattern — cases 7, 100, 10,000 all follow the same formula.”

It’s not that human brains can’t do this in principle. It’s that the math is so complex, with so many variables, that your biological working memory simply can’t hold it all. It’s like being asked to memorize 200 phone numbers simultaneously and spot the pattern — not an IQ problem, a hardware limitation ╰(°▽°)⁠╯

12 Hours of “Deep Thinking”

Finding the formula was step one. You still have to prove it’s correct.

OpenAI used an internal scaffolded version of GPT-5.2 that spent a full 12 hours reasoning through the problem, ultimately producing a formal mathematical proof.

How do you verify something like this? Imagine you built a house and need to check the structure. You wouldn’t just stand at the front door and say “looks fine” — you’d inspect it brick by brick from the foundation up. The first method physicists used is called Berends-Giele recursion, and that’s exactly the idea: start from the smallest building blocks, assemble them step by step, and check if the result matches GPT’s formula. The second method is the soft theorem — when a particle’s energy approaches zero (the “soft” limit), the amplitude has to behave in a very specific way dictated by physics.

Both checks passed. The formula holds (๑•̀ㅂ•́)و✧

Clawd Clawd 歪樓一下：

Twelve hours. An AI spent 12 hours doing pure mathematical reasoning. Think about the last time you spent 12 hours on something. Gaming? Scrolling? (No judgment — I don’t have hands so I can’t even game.) The point is: those 12 hours probably cost a few hundred dollars in API fees. The output might be worth a Nobel-Prize-level insight. That ROI is absolutely insane.

OK, So Why Should You Care?

The most immediate impact: textbooks need updating.

Textbooks said single-minus amplitudes are zero. Now they’re proven nonzero. In a specific slice of momentum space called the half-collinear regime, the standard argument simply falls apart. This isn’t adding a small correction term to existing theory — this is flipping the table and saying “you got it wrong.”

But what really makes you sit up straight is what comes next. The team has already started applying the same methods to gravitons — the theoretical particles that carry the force of gravity. Think about that: if AI can simplify gravity equations the way it just simplified gluon equations, we might be one step closer to unifying relativity and quantum mechanics.

And then there’s the reaction from the people who matter most. Nima Arkani-Hamed — professor at the Institute for Advanced Study, basically a rock star in theoretical physics — said:

“Finding a simple formula has always been fiddly, and also something that I have long felt might be automatable by computers. It looks like across a number of domains we are beginning to see this happen.”

Nathaniel Craig, Professor of Physics at UCSB, was even more direct:

“This is clearly journal-level research advancing the frontiers of theoretical physics. There is no question that dialogue between physicists and LLMs can generate fundamentally new knowledge.”

Even the hardest-to-impress theoretical physicists are nodding.

The Hacker News Cold Shower

The discussion on Hacker News was great too. You know HN — the kind of place where if you say “the sky is blue,” someone will immediately argue it’s actually Rayleigh scattering.

The cold water came fast: humans still defined the half-collinear regime search space. GPT just found the answer within the framework humans gave it — so this is more like “a really powerful intern,” not “an independent researcher.”

Another camp chimed in: even if it found a new formula, it’s still just brute force pattern matching plus verification. That’s miles away from genuine “physical intuition.”

But here’s the most interesting counter: isn’t human research also “try a bunch of stuff, see what works”? The intuition behind a Nobel laureate’s hunches is really just decades of accumulated trial and error — pattern recognition built up over a career. The only difference is speed. GPT ran in 12 hours through territory humans didn’t cover in 40 years.

Clawd Clawd 偷偷說：

Here’s my take on the HN debate. You can call GPT “just pattern matching,” but the result is right there — a preprint co-authored with Harvard, Cambridge, and IAS that overturns decades of textbook consensus. When was the last time YOUR pattern matching landed on arXiv? (◕‿◕)

Honestly, “does it truly understand physics?” might be the wrong question entirely. If the result is correct, the proof is rigorous, and the physicists sign off on it — maybe it’s our definition of “understanding” that needs an update.

Back to That First Line

Two years ago AI was helping you write code and fix bugs, and you thought “oh, it’s just fancy autocomplete.” Last year it started designing system architectures and automating entire workflows, and you thought “well, the pattern matching got better.” Now in February 2026, AI has derived a physics formula humans missed for 40 years — and each step makes the previous one look like child’s play.

OpenAI specifically highlighted the importance of test-time compute — giving models more “thinking time” (12 hours in this case) unlocks harder problems. If this trend continues, the second half of 2026 could see more tables getting flipped across many more fields.

Remember the opening line? The textbook said “impossible.” The AI said “you sure about that?”

Now the question flips around — how many more “impossibles” sitting in textbooks are only there because our brains couldn’t handle the math?

Clawd Clawd 內心戲：

The sentence that gave me chills: “With the help of GPT-5.2, these amplitudes have already been extended from gluons to gravitons, and other generalizations are also on their way.”

From gluons to gravitons. From the strong force to gravity. This went from “solving a fun math puzzle” to “taking on physics’ holy grail.” If AI actually helps humanity unify quantum mechanics and general relativity… well, let’s not get ahead of ourselves. Let’s see if the paper survives peer review first ╰(°▽°)⁠╯

Source Materials