Garry Tan to Frontier Labs: Let Open-Weight Rivals Distill Your Models
Y Combinator's CEO says US open-weight labs should distill frontier models to prevent a single AI monopoly.
Policy · Source: Hacker News
What happened
Y Combinator CEO Garry Tan is pushing back against calls to regulate AI distillation. Anthropic recently claimed Chinese labs are using stolen credentials to secretly distill its frontier models. Anthropic CEO Dario Amodei wants the US government to crack down on the practice. Tan strongly disagrees. He thinks regulators should stay entirely out of it.
Tan wants American open-weight labs to do the exact same thing. He just wants them to use the front door instead of stolen credentials. He told CNBC the US needs an American distillation regime. He argues smaller labs should freely prompt frontier models to learn how they reason. This would create a robust ecosystem of US open-weight options to counter foreign threats.
Tan points out the massive hypocrisy of frontier labs. Companies vacuumed up copyrighted human knowledge without permission to build their products. Now they want to use restrictive terms of service to stop others from learning from their outputs. Tan believes intelligence trained on public data should act as a public good. He calls a single monolithic AI monopoly the true doomer scenario.
Key facts
- Anthropic — Released a second report alleging Chinese labs use illicit distillation attacks.
- Dario Amodei — Anthropic CEO who called for US regulators to crack down on distillation.
- Garry Tan — YC CEO advocating for an American distillation regime.
Why it matters
This is a direct clash between incumbent frontier labs and the startup ecosystem. Frontier labs want to build massive moats. They use restrictive API terms to prevent competitors from using their outputs for training. If Tan wins this argument, startups get a massive shortcut. They can use cheap API calls to train highly capable open-weight models. They avoid spending billions on base training.
The second-order effect is a total shift in AI economics. If distillation becomes legally and culturally protected, the value of massive base models drops. The moat shifts away from raw intelligence. It moves toward application workflows and proprietary data. It also prevents the nightmare scenario Tan fears most. A single company cannot control all the capital and research if open-weight models constantly distill their best features.
For builders
Cheaper training through legal distillation
If distillation becomes normalized, startups can train specialized models at a fraction of the cost. You pay frontier labs for API calls, but you save millions on raw compute and data acquisition.
Terms of service risks remain high
Frontier labs still explicitly ban using their outputs to train competing models. If you distill today, you risk losing API access and facing potential lawsuits from well-funded incumbents.
Open-weight models get a massive boost
Expect open-weight models to close the gap with frontier models faster. Builders relying on local or open-source infrastructure will benefit from a constant trickle-down of frontier intelligence.
My take
Tan is absolutely right. You cannot scrape the entire internet without permission and then cry foul when someone scrapes your API. Frontier labs want regulatory capture to protect their massive compute investments. Builders need open weights to survive, and distillation is the great equalizer.
Original reporting: Hacker News. This is my rewrite and opinion.