How AI Is Accelerating Scientific Discovery
Kevin Weil (head of OpenAI for Science) and Alex Lupsasca (OpenAI research scientist and Vanderbilt physicist) discuss how AI is beginning to push the frontier of scientific research. OpenAI for Science's stated mission is to compress roughly 25 years of scientific progress into 5 by putting frontier models in the hands of leading researchers. Weil argues GPT-5 is crossing into territory where it can produce novel results—small breakthroughs and existence proofs that mark the transition from 'the model can't do this' to 'I can't imagine doing this without AI.' Lupsasca, a self-described former AI skeptic, recounts concrete breakthroughs: GPT o3-pro deriving a closed form for an infinite sum involving Gegenbauer polynomials tied to pulsar magnetospheres (drawing on a 1950s Norwegian math paper), and GPT-5 Pro finding conformal symmetries of black hole equations after a warmup prompt—an answer at the edge of his own abilities. Both guests emphasize that current frontier use requires back-and-forth iteration, low pass rates, and even 'warmup' prompts, and that a major OpenAI for Science goal is reducing this cognitive load for researchers. They preview an upcoming paper (12 sections, ~8–9 external academic collaborators across math, physics, astronomy, CS, biology, materials science) that catalogues working use cases and includes 4–5 new non-trivial mathematical results. Weil repeatedly stresses that today's models are the worst we'll ever use, that pass rates climb sharply with more thinking time, and that giving expert scientists large amounts of compute is itself an underleveraged lever. Looking ahead 5 years, they highlight personalized medicine, scalable fusion, and AI-driven experimental design—while framing OpenAI for Science as a general-purpose push rather than a top-down agenda. Weil acknowledges parallel work at Google (Demis Hassabis, AlphaFold) and notes GPQA has gone from GPT-4 at ~39% to recent models near 90%, requiring harder frontier-science evaluations.
Original sources
Podcast OpenAI Podcast
Episode:How AI Is Accelerating Scientific Discovery Today and What's Ahead - Episode 10 ↗
YouTube OpenAI Podcast
Video:How AI is accelerating scientific discovery today and what's ahead — the OpenAI Podcast Ep. 10 ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.