I worry more about unintelligent optimization like evolution and search than intelligent systems. Only a few people historically put focused study in, during a time when computers were less powerful; if somebody revisits them now with a billion times as much compute and a recursion schema just slightly improved, people would be surprised at how much such a system could do even being as stupid as it would be.

@Mira are you talking about genetic algorithms?

@firstuserhere Classic genetic algorithms have similar flavor, but are thoroughly harmless. People should be using the weaker variant "randomized tree search" for lots more actually: Even just as automated tests for computer code, and I would be surprised if there wasn't a paper for your interpretability work in "automatically finding expressions equivalent to vertices in a neural network" if only you can find a suitable base language.