Two Goodfire studies suggest that Transformer neurons are more interpretable than previously thought; pruning the activations of the weakest 92% of MoE neurons increased loss by only 2.9%. Neurons can be broken down into distinct computational functions and assigned natural-language explanations that predict their behavior, but whether those explanations transfer to lab models with undisclosed architectures remains uncertain.
Source:LessWrong · Read the original
Leave a Reply