Eh the free energy principle isn’t that magical and that comes from someone who likes it quite a bit. At the end of the day extremizing the free energy is a relaxation of constrained optimization. There are ways it’s used I find very appealing but not so appealing I’d expect it to become the dominant paradigm. Moreover, you can write a lot of standard ML problems free energy optimization in one way or another, but I don’t think standard ML will quite get us there. We need at least one more conceptual step forward.
The explicit world modelling is probably the thing that makes me more hopeful about their approach. I think integrating ML and their approach might end up being the new paradigm.
