Looking through thethreads I could find, I didn’t find much in terms of good criticisms. The best two were your point on the AI actively wanting to press the detonator (which, with all due respect, was incorrect) and Paul Christiano’s long exchange, which concluded when he retracted his objection.
Utility indifference does have challenges of course: you need to be able to modify/pre-modify the AI’s utility in this way, and you need to identify the event X. These are necessary, and sufficient, for utility indifference to be implemented.
Didn’t we criticize the hell out of utility indifference the last time it came up? Did I miss the defense that proved it could work?
Looking through the threads I could find, I didn’t find much in terms of good criticisms. The best two were your point on the AI actively wanting to press the detonator (which, with all due respect, was incorrect) and Paul Christiano’s long exchange, which concluded when he retracted his objection.
Utility indifference does have challenges of course: you need to be able to modify/pre-modify the AI’s utility in this way, and you need to identify the event X. These are necessary, and sufficient, for utility indifference to be implemented.