What's this about?
A research write-up showing that distilling a Chinese model's (DeepSeek's) capabilities into an open model (GPT-OSS) doesn't carry over its built-in content restrictions — the skills transfer, the censorship doesn't.
What exactly is "distillation"?
Distillation trains a smaller/other "student" model to imitate a "teacher" model's outputs, copying its capabilities cheaply. The finding here is that what transfers is the teacher's skill, not its guardrails — the political content filters are a separate layer that doesn't come along for the ride.
What's the community saying?
HN 89 pts, 63 comments
- Interest that capability and censorship are separable — you can inherit one without the other.
- Questions about how robust the result is and how it was measured.
- Broader point about what "open" really inherits when models are distilled across borders.