$ cat interpretability/week/2026-10-12.log
the week in review · Oct 12 – Oct 18, 2026
SAELens 6.54.5 fixes silent feature-density skips
By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology
The release ships three bugs that quietly corrupted eval numbers, while nnsight 0.8.0 cuts import time from 7.6s to 1.6s.
fix: honor standalone feature-density evaluation jbloomAus/SAELens
The dispatch condition in run_evals only checked the L2, sparsity and variance flags, so a run that selected feature density as its only activation-based metric silently reported nothing. If you have been trusting feature-density numbers from such runs, they were not measured. Fixed in 6.54.5.
fix: load Gemma Scope 2 affine skip connections in the correct orientation jbloomAus/SAELens
Gemma Scope 2 stores affine_skip_connection as (d_in, d_out) and applies it as x @ A, but SkipTranscoder holds W_skip as (d_out, d_in) and computes x @ W_skip.T. The loader copied the matrix across without transposing, so every Gemma Scope 2 transcoder loaded from that path ran with a wrong skip connection. Reload any transcoders you touched on the affected loader before you record numbers from...
fix: reset sparsity stats after a full feature_sampling_window of steps jbloomAus/SAELens
maybe_reset_sparsity checked (n_training_steps + 1) % feature_sampling_window, so the first window held one step too few and stats were off by one across training. Fixed in the same 6.54.5 release as the feature-density fix. Sparsity curves from earlier runs are not comparable to runs after the upgrade.
Preserve disabled whitening when saving activation scalers jbloomAus/SAELens
Saving an ActivationScaler without whitening into a directory that still held an older activation_whitening.safetensors meant the next load restored the old whitening transform. Disabling whitening is now preserved across save and reload.
v0.8.0 ndif-team/nnsight
Lazy-loading of model classes takes nnsight import from 7.6s to 1.6s, which is the difference between a snappy REPL and a coffee break in any loop that re-imports. The release also carries a left-padding fix and a change to count source-op iterations per-fire rather than per-something-else, so check any tracing loop whose iteration counts you depend on. The release notes for the change say "laz...
Dev ndif-team/nnsight
Alongside: 6.54.5 also folds in the standalone feature-density fix and the sparsity reset fix, and routine evaluation no longer builds a full masked copy of feature activations when density statistics are off, which cuts memory on large activation tensors.
$ ls interpretability/week/ # the briefings behind this review
Keep up with Interpretability in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.
fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.
Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.