Dusto comments on AIs at the current capability level may be important for future safety work

[Dusto]

What are your thoughts about the impact of larger more capable models that are also tuned (or capability fenced) to be only task specific? Where the trust range that needs to be factored is considerably smaller.