Y Combinator chief executive Garry Tan wants American open-weight AI laboratories to be legally free to distill frontier models through ordinary, paid access. His argument enters a dispute intensified by Anthropic’s accusations that Chinese developers used deceptive accounts and other tactics to collect millions of Claude responses.
Tan does not defend stolen credentials or fraudulent access. Instead, he says US developers should be able to use the front door: pay for a proprietary model, study its outputs and use the resulting examples to improve open models without regulators deciding that the technique itself is forbidden.
A Boundary Between Access and Misuse
Anthropic described the activity it identified as illicit attacks involving fraud and stolen credentials. The company says Chinese laboratories orchestrated large-scale campaigns to obtain responses for training competing systems, and its chief executive, Dario Amodei, has called for stronger restrictions.
Tan draws a narrower line. He rejects bypassing controls but argues that distillation through legitimate access should remain available to American open-weight developers. In his view, a customer who lawfully queries a model should not automatically lose the ability to learn from the response merely because the provider dislikes the downstream use.
That distinction leaves practical questions unanswered. The TechCrunch account does not set out a proposed law, licensing framework or test for separating legitimate learning from prohibited extraction. Providers can impose contractual limits through their terms, while regulators could treat deception, circumvention and ordinary API use differently. Tan offers a principle, not a detailed enforcement regime.
Training Rights Cut in More Than One Direction
Tan challenges the idea that closed-model companies should control what information customers may derive from their APIs. He points to the way proprietary laboratories built their own systems from large collections of publicly available and copyrighted material, often without receiving specific permission from each creator.
That history, he argues, weakens their claim to exclusive control over the outputs of those systems. If frontier companies benefited from broad access to human-created information, an open-weight competitor should not necessarily be blocked from using paid model responses as a training resource.
The comparison is political and ethical rather than a settled legal conclusion. TechCrunch does not establish that every form of model distillation is lawful, nor does it equate all training data collection with credential theft. Copyright, contracts, computer-access rules and technical safeguards can create different obligations in different circumstances.
Tan’s position nevertheless exposes an asymmetry in the market. Closed providers can use their scale and interfaces to distribute advanced capabilities while retaining model weights. Open developers may need examples from more capable systems to narrow the gap, but their access can disappear if providers classify that use as competitive abuse.
Open Weights as a Check on Concentration
Tan frames the issue around the structure of the American AI ecosystem. He wants both frontier laboratories and open-weight projects to remain viable, rather than allowing one dominant provider to determine which applications, researchers or competitors can reach advanced models.
His worst-case scenario is a single monolithic company controlling frontier AI. Open weights can provide an alternative because developers may inspect, adapt and run models without depending continuously on one vendor’s hosted service. Distillation could help those models advance faster, although the article provides no evidence that US labs can match frontier performance through this route alone.
The proposal therefore combines competition policy with technical access. It accepts that providers have grounds to stop fraud and stolen accounts, while resisting a broader rule that would make learning from model outputs inherently illegitimate. The unresolved challenge is turning that balance into rules that protect services without giving incumbents a veto over new open models.
Tan’s intervention does not settle Anthropic’s allegations against Chinese laboratories. It shifts the question toward what American developers should be allowed to do openly. The debate now concerns not only who took which responses, but who gets to define the permitted path from a closed frontier system to a more widely available model.
