If we solve alignment, we still have to make sure that people create aligned AIs and don't create unaligned AIs - at least, don't create unaligned AIs beyond a certain level of "power". This seems hard!
@Nikola and I claim these things constitue almost a complete solution as long as compute continues being the bottleneck and the number of actors with the ability to train super strong models remains very small
@mariopasquato For the purposes of this question, I will use the following definition: an AI is aligned if it does not break any laws, nor does it cause, by act or omission, more than 1,000 biological humans to die, nor does it torture anyone.
(This might sound redundant, but it is not. Consider that an AI might try to brainwash people into passing a law to make torture legal, and then torture people.)
@RobinGreen What about a military AI capable of killing foreigners while respecting the law of the state that developed it? Is it considered aligned or not?
@mariopasquato It depends. It would also have to respect international law like the Geneva Convention.