A theoretical physicist asked GPT-5 to reproduce one of his hardest papers. It took 30 minutes. Then GPT-5.2 went further and derived entirely new results in quantum gravity, producing 110 pages of novel physics calculations in under a day. The results were verified by experts and published as a preprint with researchers from Harvard, Cambridge, Vanderbilt, and the Institute for Advanced Study.
This isn’t AI helping someone write a better email. This is AI extending the frontier of human knowledge. And if you run any kind of service business that touches complex analysis, you should be paying very close attention.
What happened
- Alex Lupsasca, a theoretical physicist who won the 2024 New Horizons in Fundamental Physics Breakthrough Prize (often called the “Oscar for physics”), joined OpenAI’s new Science team in October 2025.
- When GPT-5 launched, Lupsasca asked it to solve a problem from a just-published paper. It initially said no answer. But after “priming” the model with a related textbook warmup problem (a tip from OpenAI CRO Mark Chen), GPT-5 reproduced the full result in 11 minutes.
- Lupsasca then brought a harder problem to GPT: a formula involving 32 terms that had stumped his former PhD advisor (Prof. Andrew Strominger at Harvard) and collaborators for over a year. ChatGPT fully solved it in one week, before Strominger’s plane even landed for a planned visit to OpenAI.
- The model didn’t just brute-force it. It found a clever limiting case (the “half-collinear regime”) that collapsed the ugly formula into something simple and intuitive. Then it proved the result using a technique the authors didn’t even know about.
- The team then asked ChatGPT to generate entirely new physics from scratch, this time for gravitons (particles related to gravity and quantum mechanics). GPT produced 110 pages of novel calculations and techniques in less than a day. The team spent three weeks verifying everything. All correct.
- OpenAI published the graviton result as a preprint on February 13, 2026, with co-authors from five institutions. The tweet announcing it got 4.52 million views.
The numbers
- 11 minutes for GPT-5 to reproduce a peer-reviewed physics result (after priming)
- 30 minutes to reproduce one of Lupsasca’s best papers, which originally took months
- 110 pages of novel physics output from the graviton research session
- Less than 3 days total to generate the new quantum gravity results (plus 3 weeks of human verification)
- 1+ year that human experts had been stuck on the gluon problem before AI solved it in a week
- 4.52 million views on OpenAI’s announcement tweet
- 9,539 likes and 1,489 reposts on the announcement
5 things every service business operator should take from this
- Priming is a real technique, not a gimmick. The breakthrough only happened after they gave the model a simpler warmup problem first. Think of it like stretching before a sprint. If you’re throwing hard problems at AI and getting bad results, try walking it through an easier version of the same type of problem first.
- The “jagged frontier” is real. Lupsasca noted that most people found GPT-5 underwhelming because they tested it on things like writing emails. But at the frontier of difficulty, capabilities had “really taken off.” The lesson: you’re probably underestimating what these models can do on your hardest problems while overestimating what they add to your easiest ones.
- AI doesn’t just compute. It finds approaches humans miss. ChatGPT used proof techniques that the expert physicists didn’t know about. It imported methods from adjacent fields. This is the part that should make you rethink your R&D workflows entirely.
- Speed changes the game, not just the score. When you can test a hypothesis in 11 minutes instead of 11 months, you don’t just do the same work faster. You explore 100x more possibilities. Lupsasca said AI lets him “map out much more ambitious problems to tackle.”
- Human taste still matters (for now). The AI didn’t wake up and decide to study gravitons. Lupsasca and his colleagues chose which problems to tackle, how to prime the model, and spent three weeks verifying results. The human role shifted from doing the math to directing the inquiry and validating the output.
The hot take
This isn’t just a physics story. It’s a preview of what’s about to happen in every domain that involves complex analytical reasoning. Legal research. Financial modeling. Engineering simulations. Actuarial science. If a model can derive novel quantum gravity results that stumped Harvard professors for a year, it can almost certainly find patterns in your contracts, your financial data, or your building specs that your team is missing. The businesses that figure out “priming” and structured AI workflows for their specific domain problems will have an absurd advantage over competitors still using AI to rewrite marketing copy.
The Agency OS play
Pick your hardest analytical problem this week. Not the routine stuff. The thing that’s been sitting in a spreadsheet or a document or someone’s head for months because nobody has time to crack it. Feed it to the best model you have access to. But don’t just dump it in cold. Use the priming trick: give the model a simpler version of the same type of problem first, let it solve that, then escalate to the real thing.
Start building a library of “warmup prompts” specific to your industry. If you’re in finance, that might be a textbook DCF problem before asking for a complex valuation. If you’re in legal, a straightforward contract interpretation before a gnarly multi-jurisdictional question. If you’re in construction, a basic load calculation before a complex structural analysis. The point is to get the model thinking in the right domain before you ask for the hard stuff.
Finally, rethink your team’s time allocation. The physicists in this story spent three weeks verifying three days of AI output. That ratio (roughly 7:1 verification to generation) is probably close to right for high-stakes work. Stop thinking about AI as a tool that saves time on generation. Start thinking about it as a tool that lets you attempt problems you never would have tried before, with your team’s time shifting toward verification and direction-setting. That’s where the real value sits.