Skip to content
OpenAIOpenAI

How AI is accelerating scientific discovery today and what's ahead — the OpenAI Podcast Ep. 10

AI is beginning to change how science gets done. Head of OpenAI for Science Kevin Weil and OpenAI research scientist Alex Lupsasca talk about the early signs of acceleration researchers are seeing with GPT-5—from surfacing literature across fields and languages, to speeding up complex calculations, to designing follow-up experiments. They unpack what’s possible today, what doesn’t work yet, and why the next few years could reshape the trajectory of scientific progress across physics, math, biology and beyond. Chapters: - 00:00:40 — OpenAI for Science mission - 00:06:00 — Literature search and intersections across fields - 00:11:19 — A fusion physicist shows what GPT-5 can do - 00:15:08 — GPT-5 Pro and black hole symmetries - 00:19:02 — Getting the most out of the models - 00:24:33 — OpenAI’s new research paper (https://openai.com/index/accelerating-science-gpt-5/) - 00:29:59 — Looking ahead to the next 5 years - 00:32:05 — Will predictions outpace experiments? - 00:36:43 — The pace of model improvement - 00:40:31 — What do scientific benchmarks look like? - 00:44:16 — Fusion and the promise of abundant energy - 00:48:07 — Closing: Science 2.0 moment

Andrew MaynehostKevin WeilguestAlex Lupsascaguest
Nov 20, 202548mWatch on YouTube ↗

At a glance

WHAT IT’S REALLY ABOUT

AI as a 24/7 collaborator accelerates frontier science and discovery

  1. Kevin Weil and Alex Lupsasca describe OpenAI for Science as an initiative aimed at compressing decades of discovery into just a few years by putting frontier AI models into scientists’ hands.
  2. They share concrete stories where models helped with hard theoretical physics and math tasks, plus “conceptual” literature search that finds relevant work across disciplines and languages (often missed by keyword search).
  3. A key nuance is that frontier problems often have low pass rates: effective use requires iterative back-and-forth, warm-up problems, verification, and human judgment to separate “too hard” from “rarely right but solvable.”
  4. They preview an upcoming multi-author paper documenting what works/doesn’t work today, including shared chat logs and several new non-trivial math results, and they forecast rapid near-term change in how science is conducted—especially as models get more “thinking time” and adoption widens.

IDEAS WORTH REMEMBERING

5 ideas

The biggest felt impact of AGI may come through science.

Weil argues that breakthroughs in personalized medicine, scalable fusion, and other scientific advances will affect everyday life more profoundly than “AI inside a chatbot,” making science a central pathway for societal AGI benefits.

Acceleration isn’t only ‘better answers’—it’s parallel exploration.

Even when a scientist could eventually solve a problem, getting to test many approaches quickly (e.g., 10 paths in an hour vs. 2 paths in a week) changes what’s feasible and increases the chance of breakthroughs.

AI’s most immediate superpower in research is conceptual literature search.

Models can link ideas across disciplines, terminologies, and languages—finding obscure, relevant prior work (e.g., a German PhD thesis) that keyword search would likely miss, reducing reinvention and speeding progress.

Frontier performance is ‘jagged’ for both humans and models.

Lupsasca notes simple-sounding questions can be unsolved while highly technical predictions can be extremely precise; similarly, models can fail on basics yet succeed on niche, hard tasks—often in ways complementary to human strengths.

Low pass-rate tasks are where the most valuable science help is today.

Weil explains that some problems are solvable by the model only occasionally (e.g., 5% pass rate). The practical challenge is knowing when to keep iterating versus when the task is truly beyond capability—an area OpenAI wants to make easier.

WORDS WORTH SAVING

5 quotes

“Maybe the most profound way that people are going to feel AGI in their lives is through science.”

— Kevin Weil

“The acceleration that is going to come from these tools is going to change science.”

— Kevin Weil

“It had to go and find this special identity… published in one paper from the 1950s in a Norwegian Journal of Mathematics.”

— Alex Lupsasca

“These are the worst AI models that we will ever use for the rest of our lives.”

— Kevin Weil

“Human and AI together are much more powerful than human alone or AI alone.”

— Kevin Weil

OpenAI for Science mission: compressing timelines of discoveryAI as collaborator: breadth, depth, and patiencePhysics case studies: pulsars, black holes, symmetry-findingConceptual literature search across fields and languagesLow pass-rate frontier problems and iterative workflowsThinking time/compute as a lever for better resultsBenchmarks moving from GPQA to frontier-science evalsLife sciences: pruning search spaces and regulatory accelerationFive-year outlook: bottlenecks, experimentation capacity, automationAdoption gap: model capability vs. community usage

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.

Add to Chrome