John Schulman: OpenAI once doubted next-token prediction would lead to intelligence

AndrewDai · x · 2026-09-12

Andrew Dai quotes Dwarkesh Patel's interview with John Schulman, who reveals that in OpenAI's early days he didn't believe next-token prediction would lead to intelligence, since the signal would be swamped by noise.

Dai adds his own take: next-token prediction works not because it accurately models how human brains learn, but because it's one of the only objectives that actually scale, unlike many alternatives.

The key insight: which techniques elicit out-of-distribution generalization has always been hard to predict in advance — even a core researcher like Schulman got it wrong. It's a useful counterpoint to the popular claim that LLMs win by imitating the brain; scalability is the real reason.

Original post →

More from AGI Musings

AGI Musings channel →