🔬Doing Vibe Physics — Alex Lupsasca, OpenAI

What It Covers

Vanderbilt physicist and OpenAI fellow Alex Lupsasca describes how GPT models solved two open problems in theoretical physics — single-minus gluon and graviton tree amplitudes — that stumped expert researchers for over a year. The episode traces AI's progression from email assistant to quantum field theory collaborator, covering methodology, implications for scientific training, and the verification bottleneck now facing researchers.

Key Questions Answered

• •AI Research Inflection Points: Three distinct capability jumps mark AI's entry into frontier physics: o3 solved a calculation in 11 minutes that would have taken days; GPT-5 reproduced a published paper's hardest derivation in 30 minutes; and an internal OpenAI model spent 12 hours independently rediscovering and proving a formula that three expert physicists could not crack in over a year of sustained effort.
• •Vibe Physics Workflow: The gluon amplitude paper was produced by feeding known formulas into GPT-5.2 Pro, asking it to simplify, then requesting a general-case conjecture. The model ran Python across 5,000 cases autonomously, reduced 32-term expressions to 4-term products, and proposed a linearly-scaling formula — replacing factorial growth. Researchers then used a separate internal model in a fresh session to independently verify and prove the conjecture without being given the answer.
• •Graviton Paper in Days, Not Months: The graviton amplitude result — mathematically distinct from the gluon case — was produced in roughly three days using only publicly available GPT-5.2 Pro. Researchers provided the gluon paper as context, wrote two paragraphs of steering instructions, and the model applied the directed matrix tree theorem unprompted. The three-week publication delay was spent on human verification and writeup, not on derivation.
• •Two Concrete Research Accelerators: AI compresses two specific bottlenecks in physics research. First, confusion time — the days spent reconciling contradictory results — drops sharply when a model can immediately identify overlooked assumptions. Second, researchers can now launch parallel "scout" sessions across 10 different approaches simultaneously, getting rapid signal on which directions are viable before committing to the full calculation, replacing the sequential trial-and-error that previously defined theoretical work.
• •The Verification Bottleneck Replaces the Derivation Bottleneck: As models handle derivations, human effort shifts almost entirely to checking outputs. This creates a new constraint: models don't consistently signal confidence levels on individual steps, making it hard to know where to focus scrutiny. Lupsasca identifies two near-term model improvements needed — better calibration of expressed uncertainty on specific steps, and integration of formal verification tools like Lean to automate output checking at scale.
• •Graduate Training Has No Clear Answer Yet: The traditional PhD model relies on professors giving students "safe" problems — questions with known solutions — to build technical confidence over six-month cycles. Models can now solve many of those training problems in under 30 minutes. Lupsasca states academia has no established replacement framework. The skill that transfers most directly to effective AI collaboration is the same skill developed by advising humans: knowing how to frame a question at the right level of specificity for a given collaborator.
• •Raising the Bar Rather Than Increasing Volume: Models can now produce a publishable physics paper per day on incremental questions. Lupsasca argues the correct response is not to maximize output but to target harder problems — specifically, questions that have blocked entire research communities for decades rather than individual groups for months. The single-minus amplitude results open a line of attack on quantum gravity questions, and the goal is to use AI to reach problems that previously had no viable computational pathway.