I'd love to see someone explore this - the idea of a trained model unexpectedly exhibiting "good" behavior due to its internal workings is wild to me. https://www.reddit.com/user/Objective_River_5218