Nathan Lambert, head of the post-training team at the Allen Institute for AI (AI2) and a leading authority on RLHF (reinforcement learning from human feedback), is warning that the anti-distillation legislation currently gaining momentum in US politics could seriously harm America's own open-source AI startups and academic research. He acknowledges that frontier model companies like OpenAI and Anthropic are major strategic assets, and that distillation does undercut their competitive position. But he argues that lawmakers are moving before they understand distillation's actual impact—and the consequences could be far worse than the problem they're trying to solve.
https://twitter.com/natolambert/status/2047454390601306207
Lambert points to several specific risks: Cursor and other US startups depend on Chinese open-source models to maintain independence from closed-source vendors. "Much of US academic research is built on Chinese models," he notes. A ban would set open-source model capabilities back by six to twelve months and concentrate power in closed-source labs. He also questions the evidence about how distillation data is obtained—the source is the very closed-source companies pushing the policy.
His core argument: the US has the world's largest inference market, and cheaper open-source models provide a healthy check on frontier closed-source models, driving investment and innovation. If the US now cuts ties with the global open-source community, it may lose its voice in the open-source model space. Chinese open-source models like Xiaomi's MiMo V2.5 are already competing with the likes of Claude and GPT, illustrating the global shift.