YC President Pushes Back on Calls to Crack Down on AI Distillation

Introduction

The debate over large-model distillation is moving beyond engineering circles and into regulation and industrial policy. Anthropic wants U.S. regulators to curb what it describes as “illegal distillation attacks.” Garry Tan, president of Y Combinator, took the opposite view in a CNBC interview, arguing that the United States should not respond with new restrictions simply because such practices exist.

Key points

• In its second threat intelligence report, released in September, Anthropic accused some Chinese laboratories of using hidden identities, deception, and stolen credentials to obtain the capabilities of frontier models.
• Anthropic CEO Dario Amodei had already publicly called for regulators in the United States to address the alleged activity.
• Tan argued that U.S. AI laboratories should also be able to use distillation and opposed broadly suppressing the technique through regulation.
• The dispute involves more than whether distillation is acceptable. It also raises questions about authorization, access controls, trade secrets, competition, and national security.

Why distillation is controversial

Model distillation generally means using the outputs of a more capable model to help train another model, allowing the second system to reproduce some capabilities at a lower cost. For model developers, this can weaken the competitive advantage created by expensive training. When the process involves bypassing access controls, impersonating users, or using stolen credentials, however, the issue moves from technical imitation into security and compliance.

Distillation itself is not automatically unlawful. It can be used for research, model compression, deployment optimization, and transferring capabilities to smaller systems. The harder questions concern which data and outputs may be used legitimately, whether terms of service provide a clear basis for enforcement, and how responsibility should be established across borders. The available material does not include the full report or supporting evidence, so it does not allow a definitive judgment about every allegation or its legal status.

What the disagreement signals

Anthropic’s position emphasizes protecting frontier-model capabilities and limiting risks created by model access. Tan’s position places more weight on diffusion and competition, with concern that broad restrictions could reduce room for U.S. companies to research and deploy AI. These are two different policy priorities: protecting valuable capabilities versus avoiding rules that block innovation.

A blanket ban on distillation could affect legitimate research and engineering work. At the same time, a completely unregulated environment could intensify disputes involving credential abuse, automated extraction, and commercial secrets. A more precise framework would distinguish authorized compression and research from deceptive access, while clarifying rules for permission, auditing, logging, and accountability. The debate has no settled answer yet, but it shows that AI regulation is expanding from how models are released to how their capabilities are accessed, transferred, and reused.

Source: OSChina