I agree with Nick Cammarata and Goodfire that monosemanticity should've been treated as more of a null hypothesis. IMO, the polysemanticity hypothesis has historically achieved a strange degree of memetic status and has been taken far too seriously.
Nevertheless, I want to underscore that neuron-level interpretability is not just another technique. It marks a new paradigm in interpretability research. Unlike previous paradigms, such as the NLA/SAE paradigm of using other neural networks to explain the original one, neuron-level interpretability has the adva... (read more)
I agree with Nick Cammarata and Goodfire that monosemanticity should've been treated as more of a null hypothesis. IMO, the polysemanticity hypothesis has historically achieved a strange degree of memetic status and has been taken far too seriously.
Nevertheless, I want to underscore that neuron-level interpretability is not just another technique. It marks a new paradigm in interpretability research. Unlike previous paradigms, such as the NLA/SAE paradigm of using other neural networks to explain the original one, neuron-level interpretability has the adva... (read more)