Was it toxicity though as understood by the model, or just a cluster of concepts that you've chosen to label as toxic?

I.e., is this something that could (and therefore, will) be turned towards identifying toxic concepts as understood by the chinese or us government, or to identify (say) pro-union concepts so they can be down-weighted in a released model, etc?

My understanding coming from mechanistic interpretability is that models are typically (or always) in superposition, meaning that most or all neurons are forced to encode semantically unrelated concepts because there are more concepts than neurons in a typical LM. We train SAEs (where we apply L1 reg and a sparsity penalty to “encourage” the encoder output latents to yield sparse representations of the originating raw activations), to hopefully disentangle these features, or make them more monosemantic.This allows us to use the SAE as a sort of microscope to see what’s going on in the LM, and apply techniques like activation patching to localize features of interest, which sounds similar to what you’ve described. I’m curious what this work means for mech interp. Is this a novel alternative to mitigating polysemanticity? Or perhaps neurons are still encoding multiple features, but the features tend to have greater semantical overlap? Fascinating stuff!