Okay, goal: figure out mechanistic interpretability so that if i get hold of Gato i can see whether it's relying on a few conglomerates of "agency" neurons
a Schelling point for those who seek one