The safety switch that doesn't actually work

Sparse autoencoders — the core tool of mechanistic interpretability — can identify and amplify specific concepts inside a neural network, but they cannot reliably suppress unwanted behavior by clampin...