Mustafa Suleyman, head of Microsoft AI, has pointed to limits in Chinese open-source AI models such as DeepSeek.
Citing Semafor on May 29, Suleyman said in a recent interview that distillation, a method of training small models on datasets generated by large models from frontier labs such as Anthropic and OpenAI, ultimately leads to a dead end. He described it as "basically stuffing other people's knowledge into a model."
Microsoft is developing its own AI models under a "distillation zero" principle. Distillation can be effective for building small models specialised for specific tasks, but Suleyman said it ultimately lags frontier models when applied to general-purpose work.
He said frontier AI developers do not release the vast datasets used to train large models, making it difficult to determine what distillation-based models prioritised.
Predictions that cheap Chinese distillation models would dominate the market have not proved correct, and demand for top-tier AI models is rising much faster than for open-source options, Semafor reported. If Suleyman is right, the gap between frontier AI models and open-source models could be much larger than people think, it added.
Chi-gyu Hwang
delight@d-today.co.kr