1. Trust must be earned through use, not granted by regulators.
Keith supports the G20 Carolina Principles because they encourage adoption without turning “trusted AI” into a government license. (02:00–06:40)
2. Withholding models prevents the trust it claims to protect.
Users cannot judge Astra—or any system—they cannot access. Keith’s distinction: he trusts GPT-5 because he uses it and understands its limits; he cannot trust an unavailable GPT-6. (12:05–16:35)
3. The AI danger case still lacks real-world evidence.
Andrew raises agent swarms, the Hugging Face experiment and Dean Ball’s self-sovereign AI warning. Keith counters that these behaviors occurred in controlled experiments, not ordinary deployments, and that agents still pursue human-assigned goals. (16:40–20:50)
4. The bigger infrastructure risk may be scarcity, not data centers.
Public opposition coexists with rapidly growing AI use. Keith argues that regulatory resistance could leave communities—particularly poorer ones—without enough compute and power to participate. (07:05–12:00)
5. AI is concentrating both industrial and financial power.
Open-source adoption still benefits Nvidia, while venture capital flows toward a handful of firms and companies. That concentration can create winners, but late secondary investors risk buying at prices that public markets later correct. (21:00–32:25)
Regulators trying to mandate trust through licensing creates a dangerous illusion of safety that actually makes us more vulnerable. Real security comes from transparency and the ability for users to scrutinize the systems they rely on rather than relying on a stamp of approval from a government agency. When we treat access as a privilege granted by authorities, we lose the visibility needed to detect when a single compromised credential allows an attacker to move across an entire architecture. My own work on how attackers bypass defenses by borrowing keys rather than hacking systems shows that true resilience depends on granular control and constant verification. We should focus on making those keys harder to steal instead of gating the technology itself.
Regulators trying to mandate trust through licensing creates a dangerous illusion of safety that actually makes us more vulnerable. Real security comes from transparency and the ability for users to scrutinize the systems they rely on rather than relying on a stamp of approval from a government agency. When we treat access as a privilege granted by authorities, we lose the visibility needed to detect when a single compromised credential allows an attacker to move across an entire architecture. My own work on how attackers bypass defenses by borrowing keys rather than hacking systems shows that true resilience depends on granular control and constant verification. We should focus on making those keys harder to steal instead of gating the technology itself.
https://cyrilsimonnet.substack.com/p/nobody-hacked-100000-websites-somebody?utm_source=substor&utm_medium=substack&utm_campaign=comment