What is the research that would be most helpful for AI safety while still having value if it turns out not to be useful for safety research?
Conversation
Replying to
@The_Lagrangian safety wrt what? threat models are important.
(who attacks? what is attacked? how can it be attacked?)
2
1
Replying to
@The_Lagrangian safety/security is a tradeoff between getting what you want and not getting screwed over. less #1 => less opportunity for #2
1
Replying to
@The_Lagrangian that said, assuming worst case, a theoretical framework for this could be interesting and short:
Quote Tweet
assume malevolent omniscient boxed god: what's the upper bound on the bits we can read from the box without taking [quantity] damage?
Replying to
have revisited the idea: halting problem with a derivate of us (nature unknown yet) as executing machines