Garry Tan wants US labs free to learn from rival models

Garry Tan wants US regulators to leave room for American open-weight labs to use model distillation. The technique trains one model using outputs from another. Tan argues that smaller American labs should be able to use this approach to give the US more open-weight options beyond those developed in China.

The Y Combinator chief’s argument, reported by TechCrunch, comes as Anthropic seeks tougher action against unauthorised distillation. Anthropic has accused Chinese labs of obtaining access through concealed identities, fraud and stolen credentials. Tan distinguishes that conduct from domestic labs using ordinary access routes.

Who sets the limits?

Tan sees it as an overreach for AI labs to dictate what customers may do with the information their models provide. He argues that models trained on broadly available human knowledge should offer more freedom of use, rather than keeping that knowledge behind restrictive terms of service.

He also wants frontier labs to remain viable businesses that can fund further work. His concern is that one proprietary company could gain such an advantage in capital and researchers that it dominates the field. More open-weight alternatives are his proposed counterweight. This is his argument for regulators to consider, not an announced rule change.