Sam Altman: A Generation of AI Researchers Were Wrong About Scaling, and They Knew It
Sam Altman has never been shy about picking fights, and his latest target is the cohort of AI researchers who spent years confidently declaring that scaling large language models would hit a wall. Speaking at Stanford, the OpenAI CEO argued that those researchers didn't just get it wrong — they actively slowed the field down.
His remarks were partly aimed at critics like Yann LeCun, who has repeatedly called LLMs a dead end and insisted the path to real intelligence runs elsewhere. Altman's diagnosis is blunt: some people get so attached to a position that contradictory evidence stops registering. Identity gets tangled up with the thesis, and the thesis never gets updated.
He's equally unbothered by the doom-posters on social media who've been predicting OpenAI's collapse for years. "Betting against LLMs scaling at this point feels quite misguided to me," he said. Anthropic's Dario Amodei made remarkably similar noises recently, so there's at least a consensus forming among the people who've built the biggest models.
Altman did concede that world models matter — particularly for robotics, where you actually need to understand physical space rather than just predict the next token. But he thinks the evidence for continued scaling is too strong to dismiss.
On the capability side, he pointed to an OpenAI model that recently disproved a mathematical conjecture that had resisted human effort for years. Mathematicians are apparently now having something of an existential moment about what that means for their profession. "So clearly, LLMs are capable of figuring out new knowledge," Altman said. Which is a fairly significant thing to be able to say.
He was more measured about the limits. Tasks requiring sustained judgment over long time horizons are still firmly in the "worse than humans" category. So the picture is uneven — superhuman in some narrow domains, noticeably subhuman in others. Not exactly the profile of a dead end, but not omniscience either.