News · 2026-09-13

Garry Tan says American open-weight labs should be free to distill closed frontier models

Garry Tan, the president and chief executive of startup accelerator Y Combinator, said he “would do nothing” about AI model distillation and wants American open-weight labs to be free to learn from closed US frontier models. His comments, made to CNBC at Y Combinator's Demo Day and expanded in a statement to TechCrunch on 11 September 2026, cut against a federal push to treat distillation as a national-security threat.

Key facts

• The quote: “I would do nothing” about distillation, Tan told CNBC, adding, “We could argue that there should be an American distillation regime.”
• When: CNBC published on 10 September (US Eastern time); TechCrunch followed on 11 September.
• Context: 149 of the 196 startups presenting at Demo Day were machine learning and AI companies, according to CNBC.
• Primary sources: CNBC's interview and TechCrunch's report with Tan's statement.

What distillation is and why it became a security issue

Distillation trains a “student” model on the answers of a stronger “teacher” model. It is like a student learning a subject by studying a top expert's worked solutions rather than rediscovering everything from scratch: much faster and cheaper, and the student can get surprisingly close.

The technique is standard inside labs. The fight is about doing it to someone else's model. Frontier companies forbid it in their terms, and this month Washington escalated. An advisory from the NSA, CISA and FBI named six Chinese AI firms and cast distillation as central to their strategy, and Treasury Secretary Scott Bessent floated sanctions. Anthropic has been the loudest lab on the issue, though it has said it never asked to ban open-weight models.

What Tan actually said

Tan's position is easy to misread in both directions. He is not arguing for Chinese labs' behaviour, and he is not telling American startups to break rules covertly. He is arguing that the rules themselves should change so that smaller US labs can learn from the big US labs.

To CNBC, he framed it as good for competition: “This is actually the ideal case. You want open weight models to give people freedom and access.” In his statement to TechCrunch he went further: “Controlling what users and customers do with API calls to closed weight models feels constraining, and there's a role government can play here to normalize the fact that access to intelligence that was trained on broad public access data should itself also be more a form of a public good than something locked away behind restrictive terms of service.”

TechCrunch reports that Tan is not advocating stolen credentials; he wants labs to be free to “come in the front door.” The subtext is concentration. TechCrunch quotes him warning that “the nightmare scenario, the doomer scenario for AI is that there's just one company.”

Why it matters

The week's AI policy debate has been about slowing the frontier. Anthropic's chief executive proposed pacing, and critics argue that any arrangement letting two leading labs set the speed also protects their lead. Tan's remarks put a concrete policy on that critique: if the frontier labs' outputs can be legally distilled, their advantage erodes faster and open-weight models stay closer behind.

For developers, the practical stake is large. Training on a stronger model's outputs is one of the cheapest ways to make a capable downloadable model, and the legal status of that practice decides whether American startups can do openly what Chinese labs are accused of doing quietly.

The counter-argument

Critics say Tan blurs a line that matters. Startup Fortune noted that the Chinese firms at the centre of the advisory are accused of violating terms of service and relying on hidden infrastructure, and wrote: “That's not the same thing.” There is also a commercial objection Tan does not answer: frontier labs spend billions on training and argue that letting competitors copy the results for the cost of API calls would undercut the incentive to build frontier models at all.

The story drew one of the day's busier AI threads on Hacker News.

The caveat

Tan runs an investment firm whose portfolio is dominated by AI startups that would benefit from cheaper access to frontier capabilities, and no lab or government official has publicly responded to his proposal. It is a position, not a policy change.

Primary source, verified: read the paper →

Ground Truth. (2026, September 13). Garry Tan says American open-weight labs should be free to distill closed frontier models. Ground Truth. https://groundtruth.day/news/garry-tan-says-american-open-weight-labs-should-be-free-to-distill-closed-models.html

@misc{groundtruth:garry-tan-says-american-open-weight-labs-should-be-free-to-distill-closed-models,
  title  = {Garry Tan says American open-weight labs should be free to distill closed frontier models},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/garry-tan-says-american-open-weight-labs-should-be-free-to-distill-closed-models.html}
}

Topics: open-weights · distillation · ai-policy · y-combinator · anthropic · terms-of-service