A once-obscure piece of machine learning jargon is turning into a geopolitical fight. Reuters has published an explainer asking what AI model distillation is and why it is becoming a flashpoint between the United States and China — a sign that a technique long treated as a lab-bench efficiency trick is now being argued over in policy terms.
Distillation, in plain language, is the practice of using one AI model as a "teacher" to train another, usually smaller, "student" model. Instead of building a system from scratch at enormous cost, developers can transfer much of a bigger model's capability into something cheaper to run. That is exactly what makes it contentious: it can compress the advantage of an expensive frontier model into a package that is far easier and cheaper to reproduce.
A concrete example surfaced this week on Hacker News, where developers at CTGT posted a Show HN describing how they used DeepSeek V4 Flash as a teacher to train GPT-OSS-120B on finance tasks. According to their post, the approach works well: at a constrained 8,000-token budget, their self-distilled 120B model scored 83.61% on the FinanceReasoning benchmark, ahead of Kimi K3 at 81.93% and Inkling at 65.13%. They also released 20B open weights.
Their headline claim goes to the heart of the policy anxiety. The post asserts that distilling DeepSeek into GPT-OSS does not transfer censorship — in other words, that the student model picks up the teacher's capability without inheriting its content restrictions. That claim comes from the developers themselves and invites independent testing.
Why it matters: if capability can be copied across borders cheaply while the political constraints baked into a model do not travel with it, then export controls and model-level safeguards are far leakier instruments than policymakers on either side may assume.