What a Distilled Model Inherits From Its Teacher
SMRTR summary
A new study tested whether censorship from Chinese AI models transfers to American models trained on their outputs. Researchers trained GPT-OSS-120B on data from DeepSeek V4 Flash, a heavily censored Chinese model, to improve financial reasoning. The distilled model gained strong finance performance, scoring 83.61% on FinanceReasoning, but showed zero inherited censorship, at 62 times lower cost than competing models.
SMRTR provides this summary for quick context. The original article belongs to Daily.dev.
Read the original article