Distillation has emerged as one of the most discussed issues in the larger narrative around US-China relations and the spread of AI technology. The term itself has multiple meanings, but its common usage today refers to training a weaker model using the outputs of a more powerful AI system. The term originates from a more precise technical concept known as knowledge distillation (Hinton, Vinyals, & Dean 2015), which refers to a particular method of training a model to replicate a teacher model’s probability distribution. Today’s distillation is more accurately characterized in broader terms as the use of synthetic data. You use outputs from a more powerful model—typically accessed via an API—and train your own model to replicate or predict them. Knowledge distillation isn’t technically feasible with API-based models, as they don’t provide the necessary internal details to users. Meanwhile, synthetic data stands out as arguably the most valuable tool that AI researchers rely on daily to enhance their models. Yes, architecture matters, and certain data will always require exclusive human input. Emerging ideas like reinforcement learning with verifiable rewards at scale could reshape the industry. Yet much of the practical, day-to-day work in advancing today’s models boils down to how to effectively capture and scale synthetic data.
To expand on the point raised at the beginning of this piece: the repeated claim is that top Chinese labs are leveraging distillation to extract capabilities from the leading American API-based models. The most high-profile incident so far involved the release of DeepSeek R1 in 2025, when OpenAI publicly accused DeepSeek of stealing its reasoning traces by jailbreaking their API (these traces are not exposed by default). For context, a “reasoning trace” is the industry term for the model’s internal chain-of-thought process—the same stream of intermediate reasoning that open-weight reasoning models normally show to users. Fear of distillation is also likely why Gemini quickly flipped from exposing the reasoning traces to users to hiding them. There was even notable early reasoning research that was built on Gemini!
Interconnects AI