A Sputnik Moment in AI?

Recent debates have focused on whether DeepSeek’s achievement in generating a new model with comparatively limited resources qualifies as a “Sputnik moment” for Western AI companies. This discussion requires distinguishing between developing an entirely new model and adapting pre-existing models. DeepSeek’s approach exemplifies the latter; their methodology involves refining established large language models rather than constructing a new system from scratch.

DeepSeek asserts that its models are “distilled from DeepSeek-R1” and “based on Llama and Qwen” (as seen in the above screenshot, taken from here). In this context, “distilled” refers to the process of transferring the reasoning patterns from a larger, complex “teacher” model to a smaller “student” model. Given that Llama and Qwen were originally developed by Meta and Alibaba Cloud respectively, the adaptation process employed by DeepSeek requires fewer resources than building an entirely new model: a task typically associated with large technology companies.

If DeepSeek had developed an entirely new model, the achievement would align with current achievements and would not necessarily represent a breakthrough. However, by developing a new model, based on existing models and distilled from a third model, this may be a “Sputnik moment.” Microsoft’s CEO Satya Nadella commented that DeepSeek “has had some real innovations”.

Further empirical analysis is required as additional technical details emerge, particularly from independent investigations by entities such as researchers at Hugging Face.

This case illustrates that organizations can, through adaptation and distillation, develop specialized models optimized for targeted applications at a fraction of the cost of larger, more complex models.