Kavli Affiliate: Max Tegmark | Summary:Sparse autoencoders (SAEs) are designed to extract interpretable features from language models by enforcing a sparsity constraint. Ideally, training an SAE would yield latents that are both sparse and semantically meaningful. However, many SAE latents activate frequently (i.e., are emphdense), raising concerns that they may be undesirable artifacts of the […]
Continue.. Dense SAE Latents Are Features, Not Bugs