LLM scaling: Sam Altman says researchers underestimated it

Sam Altman spoke at Stanford on Jun 21, 2026 and defended continued scaling of large language models, saying a generation of researchers held the field back by underestimating what scaling could do. He added, "Betting against LLMs scaling at this point feels quite misguided to me." Altman pointed to an OpenAI model that recently disproved a mathematical conjecture as evidence that these systems can produce new knowledge, and he said mathematicians are now asking what that means for their field.

What did Altman say at Stanford?

At Stanford on Jun 21, 2026, Altman framed the debate as empirical: he said many researchers were too confident about what scaling could not achieve, argued the data supports continued scaling, and cited an OpenAI model disproving a mathematical conjecture as a concrete result prompting questions in mathematics. He described that outcome as proof that LLMs can generate novel insights in at least some domains, while also noting limits on very long-horizon tasks that require high judgment.

Altman contrasted domains. He said world models matter for robotics but insisted the empirical record favors more scaling. He placed the mathematical example at the center of his case, using it to push back on claims that large models have reached a dead end.

How did he describe critics and the limits of LLMs?

Altman named critics including Yann LeCun and said some people tie their identity to a position and cannot let go, and he dismissed persistent online negativity as not affecting him. He also acknowledged clear gaps: for very long-horizon tasks that demand deep human judgment, he said LLMs remain much worse than people, even as they outperform humans in specific areas.

He noted that Anthropic CEO Dario Amodei has recently made similar remarks in defense of scaling. Altman painted a mixed picture: demonstrable new capabilities in areas like mathematics and other tasks, alongside important failure modes when problems require extended planning or specialized judgment.

Why it matters

Altman shifted the argument from abstract limits to measurable outcomes, using a dated and named example to argue that scaling produces results worth testing. Framing scaling as an empirical strategy forces researchers, funders, and critics to reconcile model-driven discoveries with prior theoretical skepticism, and it raises practical questions about where to apply large models and where human judgment must remain central.

If LLMs can regularly produce new domain knowledge, academic fields and product teams will need to decide whether to incorporate model-led discovery into their workflows, and whether current evaluation methods capture the most consequential capabilities.

What to watch

Watch for follow-up publications or demonstrations that expand on the OpenAI model's mathematical result and for responses from the mathematics community, which the talk said has already begun asking questions. Also watch statements from named skeptics such as Yann LeCun and parallel comments from Dario Amodei to see whether the debate shifts from theoretical skepticism to engagement with empirical examples.