Okay, but I’m not at all sure this applies to communicating AGI x-risks, which is what I most want to communicate. Making these too specific (which seems necessary) seems to have backfire effects.
And studies like this are easy to overgeneralize.
I certainly agree that the strawberry example is not a good argument. We need to move away from the literal genie examples since current AI does tend to understand intentions reasonably well. There are plenty of other arguments for existential risks from takeover.
Okay, but I’m not at all sure this applies to communicating AGI x-risks, which is what I most want to communicate. Making these too specific (which seems necessary) seems to have backfire effects.
And studies like this are easy to overgeneralize.
I certainly agree that the strawberry example is not a good argument. We need to move away from the literal genie examples since current AI does tend to understand intentions reasonably well. There are plenty of other arguments for existential risks from takeover.