🔬Doing Vibe Physics — Alex Lupsasca, OpenAI

Episode Summary

Executive Summary: The episode argues that AI has crossed a milestone in scientific research: it can now solve hard, open problems in theoretical physics and quantum gravity faster than experts, sometimes rediscovering and proving results within minutes or hours. The discussion centers on two papers about single-minus gluon and graviton amplitudes, highlighting AI’s growing ability to do frontier calculations, accelerate discovery, and change how researchers train, verify, and publish science.

Main Topics: AI as a superhuman research tool (Priority: 5/5): Alex describes a shift from AI being useful for routine tasks to handling demanding theoretical physics calculations, including solving problems experts struggled with for a year. Single-minus gluon amplitudes (Priority: 5/5): The first paper shows that single-minus gluon tree amplitudes, previously thought to be zero, are non-zero in a special collinear regime and admit a simple formula discovered with AI assistance. Single-minus graviton amplitudes (Priority: 5/5): A second paper extends the method to gravity, using GPT Pro to derive and verify a similar non-zero result and connect it to symmetry structure in quantum gravity. How AI changes scientific workflow (Priority: 4/5): The conversation emphasizes that models now speed up calculations, reduce confusion, scout multiple solution paths, and can produce near-paper-ready derivations, with humans spending more time verifying than discovering. Training the next generation of physicists (Priority: 4/5): The guests discuss how AI may disrupt traditional graduate education, where students learn via long calculations, and how pedagogy may need to shift toward asking better questions and using AI effectively. Limits, verification, and publication (Priority: 4/5): The group raises concerns about AI-generated scientific output, the need for stronger verification, and the possibility that static papers may eventually be replaced by interactive, AI-native knowledge formats.

Key Arguments: AI has reached a point where it can outperform humans on some frontier scientific tasks, not just routine productivity tasks like email. The single-minus gluon result is scientifically important because a long-assumed zero amplitude was shown to be non-zero in a special kinematic regime. GPT models were not merely summarizing existing knowledge; they helped conjecture a compact general formula and, in the stronger internal model, derive the proof. The graviton extension shows the approach is not confined to gauge theory; it also works in gravity, suggesting broader relevance to quantum gravity research. Most human effort in the graviton project went into verification and write-up, implying a major acceleration in the research cycle. AI is changing the skills needed in physics: the valuable talent is increasingly knowing how to pose the right question and steer a model, not just doing calculations manually. There is a risk of AI-generated low-quality science, so the field will need better verification and possibly more formal methods. Scientific papers may become less central over time, replaced or supplemented by more interactive, explorable knowledge interfaces.

Data Points: Time for ChatGPT to reproduce one of Alex’s best paper-level calculations: ~30 minutes - Alex says GPT-5 reproduced a major paper result that took him a long time to derive. Time for Codex to write a SYK simulation: 10 minutes - A technical quantum mechanics/gravity simulation was reportedly created very quickly by Codex. Time experts spent puzzling over the single-minus gluon problem: Over 1 year - The gluon amplitude result had stumped physicists before AI helped solve it. Terms in the six-point gluon expansion: 32 terms - The transcript highlights how the naive formula becomes extremely complex at six particles. Scaling of the new gluon formula: Linear growth in number of terms - The AI-conjectured formula avoids factorial blowup and scales much better than the brute-force expansion. Time for the internal OpenAI model to derive/prove the gluon result: 12 hours - A stronger internal model rediscovered the formula and produced the proof from scratch. Time for the graviton paper to be turned around after the gluon paper: 3 weeks - This reflects the speed of the follow-up quantum gravity result and verification process. Time the public GPT Pro model thought on key graviton steps: 20 minutes / 31 minutes / 34 minutes - The model was used iteratively to explore and verify the gravity extension. Time to solve the black hole symmetry problem after priming: 18 minutes - Alex cites an earlier example of GPT solving a difficult black hole symmetry problem after a warm-up prompt. Count of the key forces discussed: 4 fundamental forces - Electromagnetism, gravity, weak nuclear, and strong nuclear force are used to frame the physics context.

Pivotal Quotes: "we're at the special time now where, at least in some directions, AI has become superhuman" — Speaker/host introduction: Opening framing of the episode’s central thesis about AI capability in science "the final formula was first conjectured by GPT 5.2 Pro and then proved by an internal open A model because that's what happened" — Alex Lufchaska: Description of how AI contributed to the gluon paper’s main result "I think we now have models that can really churn out papers that are as good as... few but written papers" — Alex Lufchaska: Discussion of how model capability is changing scientific output and publication norms

Implications: AI is shifting from assistant to collaborator in frontier science. For physics, this could accelerate discovery, reshape graduate training, increase the need for verification, and eventually transform papers into more interactive, AI-assisted knowledge systems.