Anthropic wants regulators to put an end to "illegal distillation," but Y Combinator president Garry Tan's response is: do nothing.
In September, Anthropic released its second threat intelligence report, accusing Chinese labs of conducting "illegal distillation attacks" — hiding identities and using fraud and stolen credentials to extract the capabilities of cutting-edge models. CEO Dario Amodei had previously called on U.S. regulators to crack down on distillation. Tan, however, took the opposite stance in a CNBC interview: "I would do nothing. We can argue; there should be a U.S.-specific distillation mechanism."
Two arguments: API control is overstepping, and authors were never asked during training
Tan's position is straightforward: instead of banning distillation, let U.S. open-source weight labs distill from cutting-edge labs through the "front door" — openly and legally — not by stealing credentials, but through legitimate authorization.
This view is based on two arguments. First, closed model labs' control over what users "can do with API outputs" is overstepping. Tan's exact words were: "Controlling what users and customers can do with outputs from closed-source model APIs feels suffocating. The government has a role here — normalizing the idea that intelligence trained on widely available data should be a public good, not locked behind restrictive service terms."
Second, proprietary labs also used copyrighted materials during training without asking anyone's permission — the recent $1.5 billion copyright settlement Anthropic reached in July is a clear example.
The real driver is fear of monopoly
"The nightmare scenario for AI, the apocalyptic scenario, is when only one company remains. It has the best capital channels, the best AI researchers, and suddenly becomes a monopolistic giant," Tan said. In his view, open-source weights are a counterbalance to this monopoly: "They are at the forefront and drive progress. We want this to be fundable, a good business. You also need to give people freedom and access through open-source weight models."
This statement carries weight because of the timing: Anthropic's threat report had just been released, and the U.S. government was considering whether to take action against AI distillation. YC, the biggest supporter of the U.S. open-source startup ecosystem, saw Tan step forward and say "don't ban it," effectively speaking on behalf of the entire open-source community.
However, he did not answer a sharp question: What if U.S. labs truly opened the "front door," and the distilled models were then sold or leaked? How to open the "front door," how wide to open it, and whether there are walls behind it — Tan gave no answers — which is why "do nothing" sounds casual but is extremely vague in policy terms.
