I would counter that these off-the-wall approaches have become more common with more advanced chess AIs, i.e. AlphaGo Zero and Leela Chess Zero. They frequently

I would counter that these off-the-wall approaches have become more common with more advanced chess AIs, i.e. AlphaGo Zero and Leela Chess Zero. They frequently play gambits and other aggressive lines that previous chess engines didn't. I think these limits may just be the growing pains of applying rationality to a game, and in the long term statistics and optimization can lead us to better games in general.