Lots. Been infatuated with deep learning for ~10 years. RL is unfortunately data-inefficient, but so far the best algorithm we have for some things. Its data inefficiency isn't really endemic though, it's just that we don't yet have scalable mechanisms for learning abstractions, only distributions.
I'll copy my last slime mold-related note:
Meaningful human existence is to expand into the unknown. Through exploration, technology, children, and philosophy.
This is grounded in the fact that humans are evolutionary beings, and evolution is an algorithm for finding ever more usable energy and transforming it into evolutionary information.
-this means happiness is an insidiously misleading goal
-evolution exploring niches, neurons growing, breadth-first design, engineering, and all mental effort. They fractally all are the same algorithm. Expand, reinforce, and prune. Lightning, slime molds, it appears everywhere. Greedy graph-search is evolution, it is brains, it is intelligence, and it underlies all kinds of physics and life. I think the most fundamental articulation is that it's the optimal algorithm for expanding with only local knowledge (or feedback). It's an information-theoretic reality.
Oh wow, honored to have your insights.
Never throught interns of a graph search. Tendrals are everywhere in life.
I think I get it, when the action space gets too big a bredth then depth graph search approach is the only viable option.
