Altman Criticizes Scaling Skeptics

Sam Altman, CEO of OpenAI, responded to researchers who have taken a skeptical view of scaling—the large-scale expansion of AI research—during a lecture at Stanford University. "An entire generation of researchers delayed AI progress by underestimating the potential of scaling," he stated, citing as evidence OpenAI's recent achievement of refuting an unsolved mathematical conjecture.

Scaling refers to an approach to improving model performance by increasing the volume of data used for training and computational resources. While OpenAI and Google have pursued this technique in developing large language models (LLMs), the AI research community has seen considerable skepticism, particularly from the late 2010s through the early 2020s, with many researchers expressing doubts that "scaling alone has limits" and "increasing data will not lead to true intelligence." Altman's remarks represent a direct rebuttal to such criticism.

The evidence Altman cited is OpenAI's recent achievement of refuting an unsolved mathematical conjecture. Advanced mathematical reasoning was historically considered one of the areas where AI performed most poorly. The concrete results in this domain were presented as evidence supporting the effectiveness of scaling.

The significance of this statement extends beyond technical debate. As OpenAI's CEO, Altman reiterated the company's strategic direction to the research community, justifying investment in scaling. For companies continuing to invest in massive computational resources, scaling skepticism can influence research priorities and resource allocation, so such discussions carry meaning beyond pure academic interest.

However, concerns about an exclusive focus on scaling have not been entirely dispelled. Questions about rising training costs and energy consumption, as well as whether "the scaling law might eventually hit a plateau," continue to be debated among researchers. Altman's remarks represent only one perspective on the matter, and whether a single example—the refutation of a mathematical conjecture—proves the general effectiveness of scaling remains open to dispute.

Future attention will focus on how scaling evolves in combination with more efficient architectures and inference methods. Research aimed at achieving high performance with fewer computational resources, independent of large-scale expansion alone, is also proceeding in parallel. As achievements from both approaches accumulate, discussions about the upper limits of AI capabilities are expected to deepen further.