claim
active
claim:open-sourcing-saes-on-current-leading-open-source-models-does-not-meaningfully-affect-biorisk-or-persuasion-riskOpen-sourcing SAEs on current leading open-source models does not meaningfully affect biorisk or persuasion risk
Risk judgment because base model is insufficiently capable.
Neighborhood — ranked by edge-count
Papers (1)
paper
Frameworks (1)
framework
- Planned framework built on safety evals.
Claims (1)
claim
- Harmful features mostly reflect graphic criminal content subsets of the LMSys chat training dataassociated_withExplanation for moderated features.
Methods (1)
method
- Removal of harmful and dead features from released SAEs.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Robustness influences superposition, predicting lower SAE loss in more robust models.hypothesis0.755Popperian credit for surprising prediction.
- Authors' hedged interpretation.
- App. B.
- SAE-based filtering and reward shaping show lower off-target effects but lower overall reductionfinding0.747Fig. 9.
- Extension of mechanistic interpretability findings to the metacognitive domain
- Uncertain α and β.
- No features correspond to Oxford study.
- Limitation statement.