Is distilling other labs' models a legitimate strategy, and should labs prohibit it?

Steeve Morin · Feb 24, 2025

Distillation of other labs' models is fair game

There is no free lunch — the frontier models were themselves trained on other people's data, as shown by an image model reproducing a Star Wars screenshot on request; you took it from the beginning too

53:09 Distillation is fair game since frontier models themselves trained on others data no free lunch

Mike Krieger · Mar 3, 2025

Distillation within a lab is valuable, but no nation should be able to distill models from another nation's labs

Internally, distilling your highest-end model into lower-latency, cheaper versions is useful; across nations there is national security value in being thoughtful as AI capabilities grow

30:25 Cross national distillation should be restricted though intra lab distillation is fine

Alex Atallah · Aug 10, 2026 · hedged

American labs could get pretty far by distilling Chinese open-weight models, and doing so is safer than feared because the outputs are inspectable

Most Chinese open-weight models permit distillation, letting you do RL on their outputs; and because you see the outputs during RL rollouts, you can catch misalignment with your model's voice or constitution

46:51 Distilling chinese open weights is viable and inspectable

Alex Atallah · Aug 10, 2026

Labs have a legitimate right to prohibit distillation in their terms of service and cut off competitors, and a market will split between permissive and restrictive providers

A company can cut off access to someone building a competitive model, though most frontier labs don't prohibit building small non-competitive task-specific models

49:05 Labs may legitimately prohibit distillation and the market splits

Jonathan Ross · Feb 17, 2025

China's edge is a greater willingness to use methods Western labs treat as off-limits — DeepSeek distilled the OpenAI model, which most model providers considered a red line

Many argue OpenAI scraped the internet so distillation is fair game, but the established model providers had treated distilling a rival's model as a line they wouldn't cross

53:51 Chinas willingness to cross distillation red lines gives it an edge