OpenAI CEO Sam Altman argues that a generation of researchers delayed AI progress by underestimating the potential of scaling large language models. He told an audience at Stanford that critics like Yann LeCun, who view these systems as a dead end, are clinging to positions that data has already disproved. Altman noted that even persistent online commentators predicting OpenAI’s failure no longer disturb him, as world models remain essential for robotics and other applications. He highlighted a recent instance where an OpenAI model disproved a long-standing mathematical conjecture, proving these tools can generate new knowledge. However, he conceded that for complex tasks requiring high judgment, human intelligence still outperforms current systems.
The core issue remains the tension between scaling strategies and skepticism regarding model limitations. Altman and Anthropic CEO Dario Amodei believe continued expansion of model size is necessary, while others worry about fundamental architectural flaws. The debate centres on whether current methods will eventually reach diminishing returns or continue to unlock capabilities. Altman insists that betting against scaling at this stage is misguided given the evidence of recent breakthroughs.
* Altman stated that tying one’s identity to a wrong position prevents letting go of outdated views.
* Mathematicians are now questioning the implications of AI solving problems that stumped experts for years.
* Current LLMs remain inferior to humans for long-horizon tasks requiring significant judgment.
