LessWrong post disputes a paper's finding that sparse autoencoders recover only 9% of true features

Celene Nightingale argues that the synthetic test in Korznikov et al.'s 'Sanity Checks for Sparse Autoencoders' — 3,200 feature vectors in 100 dimensions, an expansion factor of 32 — is actively hostile to SAEs rather than a forgiving toy setting; it is an unreviewed blog critique of an interpretability result.