AI access programs that allow third parties working on safety/security/resilience to access the most capable AIs (without their work being blocked by safeguards) seem very important going forward. Robustness to misuse should be doable with strong KYC and monitoring.
By "most capable", I really mean "not significantly behind the best internally deployed AI".
Broader access programs that avoid the AI company having as much discretion seem more robust. Like, ideally the company just focused on avoiding misuse.
AI access programs that allow third parties working on safety/security/resilience to access the most capable AIs (without their work being blocked by safeguards) seem very important going forward. Robustness to misuse should be doable with strong KYC and monitoring.
AI access programs that allow third parties working on safety/security/resilience to access the most capable AIs (without their work being blocked by safeguards) seem very important going forward. Robustness to misuse should be doable with strong KYC and monitoring.