I think a key difference is I do believe the technical alignment/control problem as defined essentially requires no philosophical progress or solving philosophical problems like the hard problem of consciousness, and I believe the reason for this comes down to both a general point and a specific point.
In general, one of the reasons I believe philosophy tends not to be a productive area compared to other branches of science is that usually they either solve essentially proven to be intractable problems nowadays, or they straight up tried to solve a problem in far too much generality without doing any experiments, and that’s when they aren’t straight up solving fictional problems (I believe a whole lot of possible world philosophizing is in that category).
This is generally because philosophers do far too much back-chaining compared to front-chaining on a lot of problems.
For the specific point of alignment/control agendas, it’s because that the problem of AI alignment isn’t a problem about what goals you should assign it, but rather whether you can put in goals into the AI system such that the AI will reliably follow your goals at all.
I think a key difference is I do believe the technical alignment/control problem as defined essentially requires no philosophical progress or solving philosophical problems like the hard problem of consciousness, and I believe the reason for this comes down to both a general point and a specific point.
In general, one of the reasons I believe philosophy tends not to be a productive area compared to other branches of science is that usually they either solve essentially proven to be intractable problems nowadays, or they straight up tried to solve a problem in far too much generality without doing any experiments, and that’s when they aren’t straight up solving fictional problems (I believe a whole lot of possible world philosophizing is in that category).
This is generally because philosophers do far too much back-chaining compared to front-chaining on a lot of problems.
For the specific point of alignment/control agendas, it’s because that the problem of AI alignment isn’t a problem about what goals you should assign it, but rather whether you can put in goals into the AI system such that the AI will reliably follow your goals at all.