Okay, I don't understand how legitimate access is granted then. Surely, Dario isn't saying to ban Sonnet, because I can definitely make it do cyber harm, as most exploits that I've seen in the wild with my own eyes were trivial.
So the only way I see his proposal working is:
- No open weights, AI is centralized in the hands of few
- We get AI-FAA that sets the rules and monitors
- If I want to do a security scan of my codebase, I get a time and scope limited license from AI-FAA that I upload to claude that will allow it to run the security scan in cloud with their models - Claude Mythos Scanner(TM).
Dario's proposal ultimately requires that people lose direct access to inference via API. Is this why they've been building SaaS clones with AI bolted on?