OpenAI's GPT-5.2 Derives New Physics
OpenAI's GPT-5.2 derived a new theoretical physics result for 'single-minus gluon tree amplitudes,' a finding previously thought impossible. This demonstrates a shift from LLMs regurgitating training data to performing novel scientific reasoning. Physicist Alex Lupsasca found that while GPT-5's general skills seemed stagnant, its frontier capabilities exploded, reproducing a complex paper in 11 minutes. This suggests expert 'priming' can unlock high-level reasoning in foundation models for compl
Why it matters
While public reception to GPT-5 was lukewarm, focusing on marginal gains in tasks like email writing, specialists are discovering massive leaps in scientific reasoning at the model's "jagged frontier." OpenAI's GPT-5.2 has moved beyond reproducing known information to deriving novel results in theoretical physics, acting as a reasoning partner rather than a search engine. This demonstrates that a model's most profound capabilities may be invisible to casual users and only accessible through expert-level prompting.
This shift is critical for engineers in R&D and complex problem-solving domains. It suggests that with the right domain knowledge and prompting strategies, foundation models can accelerate discovery and solve problems previously thought to be exclusively in the human domain. The ability to reproduce a post-training-data paper in minutes and then derive a completely new result signals a new era for AI-assisted science and engineering.
What changed
• Novel Physics Discovery: GPT-5.2 derived a new result showing that a specific "single-minus gluon tree amplitude" interaction, which physicists thought was impossible, can occur under certain conditions. The finding was co-published as a preprint with researchers from Harvard, Cambridge, and other institutions. • Expert-Led AI: Theoretical physicist Alex Lupsasca, winner of the 2024 New Horizons in Physics Prize, joined OpenAI to form a new "AI for Science" team after discovering GPT-5's potential. • "Priming" Technique: A key method involved "priming" the model by having it solve a textbook warmup problem. This unlocked its ability to tackle the more complex, novel problem successfully. • Rapid Replication: In an earlier test, a primed GPT-5 reproduced the full result of a complex physics paper in just 11 minutes, despite the paper being published after the model's training cutoff.
What to watch
• Cross-Domain Application: Watch for how this "AI for Science" methodology is applied to other complex fields like materials science, formal verification, or chip design, where AI could explore vast solution spaces. • Prompting Frameworks: Expect the development of more sophisticated prompting frameworks and interfaces designed specifically for scientific discovery, moving far beyond the standard chat interface.
Interview question
What is the most significant takeaway from GPT-5.2's novel physics discovery?
- a.It proves that LLMs can now independently generate groundbreaking scientific theories.
- b.It highlights that advanced LLMs can perform novel scientific reasoning when expertly guided.Correct
- c.It shows that LLMs' general performance has dramatically improved for everyday tasks.
- d.It confirms that LLMs have access to a hidden repository of future scientific knowledge.
Why? this is the answer
The card emphasizes that GPT-5.2's discovery demonstrates a shift to novel scientific reasoning, but this capability was unlocked through expert 'priming' and 'expert-led AI'. Option A is incorrect because the model did not act independently; it required expert guidance. Option C is incorrect as the card states public reception focused on marginal gains for everyday tasks, not dramatic improvement.
Just read this? Test yourself on what you have been reading.
Read the original → latent.space
Put your scrolling time to good use
Learn one idea, try a quiz and save useful cards for revision. Tezvyn makes it easy to learn and stay current in your tech field, a few minutes at a time.
The iPhone app is on the way
We are building it. Until it lands, nothing here is held back from you: every interview card, your saved cards, streaks and the job board all work in Safari, plus hundreds of free practice quizzes of thirty questions each. Sign in and it all carries over to the app the day it arrives.
Want it as an icon? Tap Share at the bottom of Safari, then Add to Home Screen. It opens full screen and the cards you have read stay available offline.
We are hiring for this. Open roles that interview on ai — each one lists the topics its interview covers.
See open roles