RepoJournal
Interpretability Interpretability
11 wires and counting

$ follow Interpretability

Keep up with Interpretability in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

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.

$ status

wire 2026-10-13
stories 16

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

Gemma Scope 2 affine transcoders loaded wrongly, SAELens sparsity stats off by one

By RepoJournal · Filed · About Interpretability · Composed from the cited sources · methodology

Two correctness fixes land in SAELens: every Gemma Scope 2 `*_affine` transcoder was loading with its skip connection transposed and producing wrong output silently, and the sparsity-stat reset was firing a step early.

fix: load Gemma Scope 2 affine skip connections in the correct orientation jbloomAus/SAELens

by chanind

Gemma Scope 2 stores `affine_skip_connection` as (d_in, d_out) and applies it as `x @ A`, while `SkipTranscoder` holds `W_skip` as (d_out, d_in), so the loader copied the matrix over untransposed and the skip connection ran backwards. Because the matrices are square nothing raised; every `*_affine` transcoder loaded cleanly and returned wrong numbers, so re-run any analysis you did on those SAEs.

fix: reset sparsity stats after a full feature_sampling_window of steps jbloomAus/SAELens

by jackiectl2

`SAETrainer.maybe_reset_sparsity()` checked `(n_training_steps + 1) % feature_sampling_window == 0`, so the first window only ever held `feature_sampling_window - 1` steps and with a default window of 2000 and 30000 total steps the running stats were wiped immediately before the final step. Sparsity figures logged from affected runs are computed over the wrong window.

Preserve disabled whitening when saving activation scalers jbloomAus/SAELens

by AaravMehta-07

Saving a scaler with whitening disabled into a directory that already held an older `activation_whitening.safetensors` would silently restore the stale whitening transform on the next load, so the reloaded scaler no longer matched what you saved. The scaler JSON now persists whether whitening is enabled and loading honors that flag; older JSON files without it keep the existing behavior.

fix: reset sparsity stats after a full feature_sampling_window of steps (#757) jbloomAus/SAELens

by Tianlang (Jackie) Chen

The sparsity-reset correction is merged as the commit behind PR 757, so it is in the tree rather than queued.

Quick answers

What shipped in Interpretability on October 13, 2026?
Two correctness fixes land in SAELens: every Gemma Scope 2 `*_affine` transcoder was loading with its skip connection transposed and producing wrong output silently, and the sparsity-stat reset was firing a step early. In total, 8 commits, 4 pull requests, and 4 releases landed.
Who contributed to Interpretability on October 13, 2026?
3 developers shipped this update, including jackiectl2, AaravMehta-07, and chanind.
What were the notable Interpretability updates?
fix: load Gemma Scope 2 affine skip connections in the correct orientation, fix: reset sparsity stats after a full feature_sampling_window of steps, and Preserve disabled whitening when saving activation scalers.