Image: OpenAI

OpenAI CEO Sam Altman is still betting that large language models will keep improving with more scale, and he thinks many AI researchers underestimated the idea for too long.

Speaking at Stanford, Altman pushed back against critics who argue that LLMs are reaching a dead end.

Altman said a whole generation of researchers held the field back because they were too confident about what scaling could not do.

“Betting against LLMs scaling at this point feels quite misguided to me,” Altman said.

The comment lands directly in the long-running argument around whether AI progress will keep coming from larger models, more compute, better training data and stronger post-training, or whether the field needs a different architecture to move beyond today’s limitations.

Altman is pushing back at the LLM dead-end argument

Meta’s Yann LeCun has been one of the best-known critics of the current LLM path. LeCun has argued that language models lack the kind of world understanding needed for real intelligence, especially in areas such as robotics, planning and physical reasoning.

Altman does not fully dismiss that point. He said world models matter for robotics and similar use cases, but he also argued that the evidence still supports continued scaling. In his view, some researchers tied their identity too closely to the belief that LLMs could not keep improving, and then struggled to update that belief when the models got better.

Anthropic CEO Dario Amodei has made a similar argument in favor of scaling, which shows how much of the frontier AI industry still believes bigger and better-trained models have room to run.

Altman says LLMs are already creating new knowledge

Altman pointed to an OpenAI model that he said recently disproved a mathematical conjecture that had remained unsolved for a long time. His argument is simple: if a model can help produce a new mathematical result, it is no longer accurate to say LLMs only remix existing text.

That does not make the scaling debate settled. Altman also admitted that LLMs still perform worse than people on very long-horizon tasks that require judgment over time. That remains one of the biggest gaps between impressive demos and dependable AI agents.

The tension is already visible in OpenAI’s product roadmap. The company is preparing broader access to GPT-5.6, a model family pitched around stronger reasoning, coding, scientific work and long-horizon agentic tasks.

The strongest argument against endless scaling may be practical rather than philosophical. Bigger models need more chips, more power, more data centers and more money. That is why the AI race is increasingly tied to AI compute spending, not only research breakthroughs.

Altman wants you to believe that people who bet against LLM scaling have been wrong so far, and OpenAI does not plan to slow down. The harder question is how much more scaling can deliver before cost, energy, data quality and reliability become the real limits.

Note: Links to online stores in articles can be so-called affiliate links. If you buy through this link, MIXED receives a commission from the provider. For you the price does not change.