RepoJournal
Interpretability Interpretability
22 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 2026-11-06
stories 4
RSS Archive Reviews

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

SAELens 6.55.0 runs on transformer-lens 4.x, but the pin stays

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

github-actions[bot] RyanJunejo
2 people shipped this
github-actions[bot] @github-actions[bot] 1 cited RyanJunejo @RyanJunejo 1 cited

SAELens 6.55.0 adds transformer-lens 4.x support and a TransformerBridge model loader, yet the release keeps the transformer-lens<4.0.0 pin because the default loading path and most tutorials still depend on the removed HookedTransformer.

v6.55.0 jbloomAus/SAELens

by github-actions[bot]

The v6.55.0 release bundles the 4.x compatibility work with a new model_class_name="TransformerBridge" option for loading models. It also fixes skip connections when HookedSAETransformer splices in a skip transcode, so spliced runs should now produce the activations you expect.

feat: support transformer-lens 4.x (requires transformer-lens 3.0+) jbloomAus/SAELens

by RyanJunejo

You can now install SAELens against transformer-lens 4.x and it imports, but the package still pins transformer-lens<4.0.0: the default model_class_name="HookedTransformer", the eval CLI and HookedSAETransformer all still require HookedTransformer, which 4.0 removed, and 7 of the 13 tutorials use it. Treat this as migration groundwork, not a green light to bump transformer-lens in a working env...

6.55.0 jbloomAus/SAELens

by semantic-release

6.55.0 is the automatically generated semantic-release cut that carries the above.

$ ls interpretability/reviews/ # the step-back reads

the week in review nnsight 0.8.2 fixes silent wrong values under vLLM the month in review SAELens folds decoder norms at load, SAEBench JumpReLU loads fixed
Full archive → Weekly reviews →

Quick answers

What shipped in Interpretability on November 6, 2026?
SAELens 6.55.0 adds transformer-lens 4.x support and a TransformerBridge model loader, yet the release keeps the transformer-lens<4.0.0 pin because the default loading path and most tutorials still depend on the removed HookedTransformer. In total, 2 commits, 1 pull requests, and 1 releases landed.
Who contributed to Interpretability on November 6, 2026?
2 developers shipped this update, including github-actions[bot] and RyanJunejo.
What were the notable Interpretability updates?
v6.55.0, feat: support transformer-lens 4.x (requires transformer-lens 3.0+), and 6.55.0.
♥

$ log Interpretability --recent

TransformerLens 4.4.0 lands, and SAELens patches four ways its loaders misread checkpoints 19 shipped TransformerLens ships v4.4.0 3 shipped param-decomp drops a public @claude agent, nnsight fixes a broken README badge 7 shipped SAELens grows a transformer-lens 4.x path but keeps the 3.x pin 4 shipped

$ log --all # elsewhere on the wire

Alignment & Post-training Anchored math grading by default in MILES, and one launcher for every dense Qwen3 TPU recipe Agent Security Sandbox Runtime 0.0.82 narrows the Windows profile read grant

$ ls showcases/

Safety Evals AI Red Teaming Agent Security Alignment & Post-training

$ init your-desk

RepoJournal can write one of these for any repo you care about. Sign in with GitHub, pick your repos and sources, and your desk gets a wire like this one every morning.