Despite creating supposedly superintelligent models, leading AI labs still rely on crude access restrictions to prevent 'distillation'—an existential threat where competitors replicate their models. This reveals a critical capability gap: their AI is not yet smart enough to detect and prevent its own theft.

It's a tenuous and difficult argument to suggest that training an AI on the public internet is acceptable, but training an AI on the output of another AI is not. Aaron Levie notes the logical inconsistency, suggesting the ethical line is blurry at best, especially when the original model provider is paid for the API usage during distillation.

There is a profound hypocrisy in the AI industry's stance on intellectual property. Companies that built their foundational models by scraping the entire internet are now seeking regulatory protection to prevent others from distilling or learning from their models—mirroring how the music industry fought Napster after profiting from an open ecosystem.

The debate over stopping AI model distillation reveals a core tension. To effectively police for theft (distillation), AI labs would need to be more restrictive with API access. This directly conflicts with the desire from startups and researchers for broader, more open access to frontier models, creating a strategic dilemma.

As enterprises replace expensive proprietary models with cheaper open-source alternatives, frontier labs like OpenAI and Anthropic face an existential threat. Their strategic response could be to lobby for regulations that effectively make open-source models illegal, creating a protective moat.

API providers like Anthropic struggle to differentiate between users distilling models for competitive purposes and those conducting large-scale evaluations. Both activities generate similar high-volume, repetitive API calls, creating a detection challenge that also raises user privacy concerns.

Frontier AI labs are restricting API access not just for security, but to prevent competitors from using 'distillation' to create cheap copies of their models. This practice makes it impossible to recoup massive R&D investments, forcing a move towards more restrictive, geopolitically motivated access.

Proposals to ban American developers from using Chinese open-source models would backfire badly. It would cut the US off from global innovation, as the rest of the world would continue to build upon these models. The correct approach to stop Chinese distillation is for US AI labs to block access at the source.

Banning open-source model downloads is futile against determined actors. A more effective control point is the specialized, expensive hardware required to run them at a dangerous scale. Implementing "Know Your Customer" (KYC) protocols for data center hardware purchases could mitigate risks more effectively than software controls.