Former OpenAI VP Liam Fedus: Real Experiments Are Irreplaceable for AI Science

In October 2026, Liam Fedus, former OpenAI vice president and co-creator of ChatGPT, joined Brendan Foody on The Frontier podcast for a roughly 50-minute conversation about AI and scientific discovery.

Fedus recalled details around ChatGPT's creation: the team discussed formats including writing bots, coding bots, and meeting bots, before John Schulman insisted on keeping the product general and building a chatbot. The ChatGPT that took off globally was actually powered by GPT-3.5, and the key shift was moving from a text-completion interface to a conversation interface. Early internal signals were mixed -- about 90% of testers showed little interest, but 5% to 10% were highly engaged, which team members later saw as a positive signal.

When explaining why ChatGPT and AI coding broke through first, Fedus said ChatGPT relied on a strong pretrained base and RLHF, though RLHF is not sensitive to correctness. Coding, by contrast, benefits from long-chain reasoning, test-time compute, and verifiable rewards, plus large knowledge bases like GitHub full of unit tests. Material science, in his view, lacks high-quality data: papers omit key details, and negative results are rarely published.

That is why his startup Periodic Labs is building physical laboratories. Fedus co-founded the company with Ekin Dogus Cubuk; it emerged from stealth in 2025 with a $300 million seed round backed by a16z, DST, Nvidia's NVentures, Accel, and others. In mid-September 2026, Periodic released its first model, Periodic Neon, a roughly 1-trillion-parameter model for interpreting X-ray diffraction patterns. Fedus stressed that mathematics has Lean as a referee, while conclusions about the physical world must ultimately be settled by real experiments. Periodic aims to combine language-model reasoning, computational physics simulation, and high-throughput labs, with a long-term goal of "intentionally understanding and designing matter."