What a crazy link:
So we train a second copy of Claude to work backwards—reconstruct the original activation from the text explanation. We consider an explanation to be good if it leads to an accurate reconstruction. We then train Claude to produce better explanations according to this definition using standard AI training techniques.
What are you even talking about? If there was a "baseline" for how to decode activations into text we wouldn't be having this conversation.
The choice is between "this reconstruction sucks" and "no reconstruction", and we're only now beginning to learn how to make those reconstructions suck less.