bilalchughtai comments on Activation space interpretability may be doomed

[bilalchughtai]

Also related to the idea that the best linear SAE encoder is not the transpose of the decoder.