OpenAI Podcast Episode 10 - How AI Is Accelerating Scientific Discovery Today and What's Ahead
18 snips
Nov 20, 2025 
Kevin Weil, Head of OpenAI for Science, and Alex Lupsasca, an OpenAI research scientist and physicist, explore how GPT-5 accelerates scientific discovery. They discuss its role in literature searches and solving complex physics problems like black hole symmetries. The duo emphasizes GPT-5’s capacity to connect diverse research topics and its application in fusion experiments. They predict significant advancements in science over the next five years, while advising researchers to embrace AI as a collaborator for innovative explorations.
AI Snips
Chapters
Transcript
Episode notes
Fusion Lab Demo Shows Multilevel Expertise
- A Lawrence Livermore physicist demoed GPT handling undergraduate to veteran-level fusion questions and recommending specialized simulation tools.
- The physicist said the model produced expert answers that would have taken days for humans.
Low Pass Rates Hide High Value Problems
- Many frontier problems have low but non-zero AI pass rates; these are the most valuable to pursue.
- Differentiating low pass rate problems from impossibly hard ones is a key research challenge.
Use Warmups To Reach Tough Answers
- Give models warmup or simpler subproblems to improve success on harder tasks.
- Encourage models to self-generate warmups so they can attack complex questions directly.