Defining AGI and the Road Ahead
OpenAI's Chief Scientist Jakub Pachocki and Szymon Sidor sit down with Andrew Main to discuss how OpenAI thinks about AGI, where the next breakthroughs will come from, and why traditional benchmarks are losing their usefulness. Pachocki frames AGI less as a fixed threshold and more as the automation of research and discovery — envisioning an entity that can produce new technology, codebases, and scientific insights at scale. He argues that capabilities previously lumped together under "AGI" (natural conversation, math, research) are actually diverging, and that point-wise human-competition benchmarks like the IMO gold medal are useful but increasingly inadequate measures of real-world impact. Both guests push back on the "AI is slowing down" narrative, with Szymon noting that headline economic-impact estimates (e.g., 3%) ignore the near-zero capability of a decade ago. They discuss the saturation of standard benchmarks, the rise of reasoning models like o1 (whose success surprised the team enough to prompt an 11pm call with Sam and Mira), and how scaling pre-training is compounding with extended chain-of-thought reasoning. Pachocki points to model persistence — using vastly more compute per problem — as the next clear frontier, citing GPT-5 Pro using 10–20x more compute than GPT-4 for meaningfully better answers. The conversation also covers recent competition results: OpenAI's model took gold at IMO, won IOI, and placed second at AtCoder (losing only to Szymon's friend). They discuss practical implications for users — including trust trade-offs around models accessing personal data like email and calendars — and close with advice for high school students: learn to code, as structured problem decomposition remains a premium skill regardless of how AI evolves.
Original sources
Podcast OpenAI Podcast
YouTube OpenAI Podcast
Video:AGI progress, surprising breakthroughs, and the road ahead — the OpenAI Podcast Ep. 5 ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.