Brief IA
›

Garry Tan Advocates for Authorized Distillation by Open US Labs

Research·Tom Levy·

Garry Tan Advocates for Authorized Distillation by Open US Labs

Garry Tan Advocates for Authorized Distillation by Open US Labs
Key Takeaways
1Garry Tan, CEO of Y Combinator, wants American open-weight labs to be able to distill cutting-edge models with authorized access
2He opposes regulatory tightening, unlike Anthropic, which denounces illicit distillations by Chinese labs
3Tan warns against the concentration of cutting-edge AI in the hands of a single company
Why it matters : The debate on distillation pits openness against control in AI development, with implications for technological sovereignty and model diversity.

The head of Y Combinator argues that American open-weight labs should be able to distill cutting-edge American models with authorized access. He claims that intelligence derived from public data should be more of a common good than subject to restrictive terms of use. This stance contrasts with Anthropic's warnings about illicit distillations by Chinese actors and its call for stricter regulations.

Avoiding a Single Player in Cutting-Edge AI, According to Garry Tan

Garry Tan considers it an apocalyptic scenario for all the power of cutting-edge AI to be concentrated in the hands of a single proprietary company. He advocates for a balance between open-weight labs and cutting-edge labs. According to him, cutting-edge labs drive the industry forward and need funding to ensure a viable long-term business model. He also believes that open-weight models offer more freedom and access to users.

Authorized Distillation to Strengthen American Options

Garry Tan finds it excessive that AI labs impose restrictions on their clients regarding the use of information derived from their models. He believes that the government could play a role in making access to intelligence trained on public data a public good, rather than a service locked behind restrictive terms. He points out that proprietary labs did not seek permission to use human knowledge, including copyrighted content, when training their models. He hopes that small American open-weight labs can use the same training techniques on cutting-edge American models to provide more American open-weight options rather than Chinese ones.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Distillation is a Common Training Method

Distillation involves having one model query another so that it learns its functioning and reasoning. This method is commonly used and considered valid for training new models. This week, Anthropic released a second report stating that Chinese labs are allegedly conducting illegal distillation attacks, masking their identity and resorting to fraud as well as stolen credentials to carry out unauthorized distillation operations. Dario Amodei, CEO of Anthropic, has publicly expressed support for tightening American regulations against distillation.

No Stolen Credentials and a Non-Repressive Framework

Garry Tan clarifies that he does not encourage the use of stolen credentials for distillation and wishes for it to be done with authorized access. He hopes that American regulators will not intervene regarding distillation by Chinese labs. He stated that he would do nothing and mentioned the possibility of an American distillation regime. Garry Tan, who has previously described himself as having a "cybernetic psychosis" as a passionate AI user, holds this position as CEO of Y Combinator. His stance contrasts with Dario Amodei's call for a crackdown on distillation.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.