cousin_it comments on Beware of black boxes in AI alignment research

[cousin_it]

I think what you’re describing is possible, but very hard. Any progress in that direction would be much appreciated, of course.