ChatGPT Co-creator Liam Fedus: Building an AI Scientist Requires Doing Real Experiments

机器之心 · wechat · 2026-10-11

In a 50-minute Frontier podcast, former OpenAI post-training VP and ChatGPT co-creator Liam Fedus recounts ChatGPT's origin and explains his new startup Periodic Labs. Before launch, the team considered writing bots, coding bots, even a meeting-note bot, until John Schulman insisted on a general chatbot; what went viral was actually the long-available GPT-3.5—the real variable was switching from completion to conversation. Internal user predictions topped out at 1M.

On why chat and coding broke through first: ChatGPT rode strong pretraining plus RLHF, while coding unlocked via test-time compute and verifiable rewards, with GitHub providing natural unit tests. Materials science, by contrast, lacks good data—noisy results, omitted details, almost no negative results. That's why Periodic Labs ($300M seed from a16z, DST, Bezos, etc.) builds its own high-throughput lab combining LLM reasoning, physics simulation, and real experiments; it already released Periodic Neon, a 1T-parameter model for interpreting X-ray diffraction. He also thinks practical fusion is 'very plausible within 7 years' and that fixing materials bottlenecks could extend Moore's Law and sustain AI progress.

Original post →

More from Companies & People

Companies & People channel →