Interconnects AI · 2026-05-04 · notable
Nathan Lambert: 'The Distillation Panic' — Why the New U.S. Crackdown on Distillation Risks Killing Open Research
Nathan Lambert argues that calling Chinese API jailbreaks 'distillation attacks' is sloppy framing that will end up banning the most important technique for diffusing AI capabilities into the open ecosystem.
Lambert says lumping API jailbreaking together with legitimate distillation will hurt U.S. open labs more than it slows China.
What is it?
A 1,500-word essay on Interconnects pushing back against the term 'distillation attack' that has dominated coverage since Anthropic's February disclosure on DeepSeek, Moonshot and MiniMax. Lambert separates two different things: API terms-of-service abuse (the actual problem) and distillation as a training technique (used by every frontier lab).
How does it work?
Lambert walks through the technique step by step: legitimate distillation uses outputs of one model to train another and is core to releasing smaller open models. He documents that xAI, Nvidia and AI2 all distill from competitors. He warns the proposed reforms (banning open weights trained on closed-model outputs) would hit Western researchers hardest because Chinese labs can simply self-publish.
Why does it matter?
With the White House drafting a frontier-model EO and Anthropic publicly accusing three Chinese labs of distillation, this is the first prominent open-camp pushback. It frames the policy choice as 'fix the API auth' vs 'kill open distillation' and is likely to be cited in the next round of U.S. open-weights regulation.
Who is it for?
ML researchers, policy teams, open-source leads