When you ask someone to unpack a concept for you it is counter-productive to repack as you go. Fully unpacking the concept of “good” is basically the ultimate goal of MIRI.
I feel that perhaps you are operating on a different definition of unpack than I am. For me, “can be good at everything” is less evocative than “achieves its value when presented with a wide array of environments” in that the latter immediately suggests quantification whereas the former uses qualitative language, which was the point of the original question as far as I could see. To be specific:
Imagine a set of many different non-trivial agents all of whom are paper clip maximizers. You created copies of each and place them in a variety of non-trivial simulated environments. The ones that average more paperclips across all environments could be said to be more intelligent.
You can use the “can be good at everything” definition to suggest quantification as well. For example, you could take these same agents and make them produce other things, not just paperclips, like microchips, or spaceships, or whatever, and then the agents that are better at making those are the more intelligent ones. So it’s just using more technical terms to mean the same thing.
“Efficient cross-domain optimization” is just fancy words for “can be good at everything”.
Yes. And your point is?
This is the stupid questions thread.
That would be the inefficient cross-domain optimization thread.
Awesome. I need to use this as a swearword sometimes...
“You inefficient cross-domain optimizer, you!”
achieves its value when presented with a wide array of environments.
This is again different words for “can be good at everything”. :-)
When you ask someone to unpack a concept for you it is counter-productive to repack as you go. Fully unpacking the concept of “good” is basically the ultimate goal of MIRI.
I just showed that your redefinition does not actually unpack anything.
I feel that perhaps you are operating on a different definition of unpack than I am. For me, “can be good at everything” is less evocative than “achieves its value when presented with a wide array of environments” in that the latter immediately suggests quantification whereas the former uses qualitative language, which was the point of the original question as far as I could see. To be specific: Imagine a set of many different non-trivial agents all of whom are paper clip maximizers. You created copies of each and place them in a variety of non-trivial simulated environments. The ones that average more paperclips across all environments could be said to be more intelligent.
You can use the “can be good at everything” definition to suggest quantification as well. For example, you could take these same agents and make them produce other things, not just paperclips, like microchips, or spaceships, or whatever, and then the agents that are better at making those are the more intelligent ones. So it’s just using more technical terms to mean the same thing.