Amid the heated debate over AI distillation, Garry Tan, CEO of Y Combinator, Silicon Valley’s most prestigious startup accelerator, shared notable perspectives.

In an interview with CNBC regarding Chinese research labs using distillation techniques to extract knowledge from American models, Tan candidly stated: “Let them do it.” At the same time, he suggested that the US should also consider building its own distillation framework.

Distillation is a widespread and legitimate method in the AI field, where developers repeatedly query another model to learn its reasoning and operational capabilities. Nevertheless, Anthropic recently alleged in a report that Chinese research labs were carrying out “illicit distillation attacks” by disguising their identities to siphon off knowledge. In response to this situation, Dario Amodei, CEO of Anthropic, called for US government intervention.

In contrast to that stance, Tan presented two counterarguments:

  • Overreach: Imposing restrictions on how customers use API responses from proprietary models constitutes excessive interference.
  • Double standards on data: Proprietary model companies themselves scrape human-generated data, including copyrighted material, without permission. Case in point: Anthropic recently agreed to settle a $1.5 billion copyright lawsuit brought by book publishers.

Speaking to TechCrunch, Tan emphasized that access to knowledge trained on public data should be treated as a public commons, rather than being constrained by restrictive terms of service.

The core takeaway: Garry Tan is by no means advocating that American labs rely on stolen data for distillation. Instead, he proposes a transparent path: granting them permission to proactively and openly distill knowledge directly from America’s own frontier models. The strategic intent is to foster a capable, autonomous US open-source model ecosystem that does not have to rely on Chinese solutions.

Furthermore, Tan warned of what he views as the AI industry’s “worst nightmare”: a single dominant corporation with the deepest pockets and top talent pulling far ahead of rivals and capturing the entire market. Underscoring this concern, he remarked: “That scenario is really bad.”

This perspective stands in stark contrast to Anthropic’s direction, opening up a dramatic public debate at the very heart of the US AI community. While frontier model labs strive to safeguard intellectual property, startup accelerators are pushing for openness. The outcome of this confrontation will shape the future and trajectory of open-source AI for years to come.