claim
active
claim:aos-are-a-bitter-lesson-compliant-approach-to-interpreting-activationsAOs are a bitter-lesson-compliant approach to interpreting activations
Argument for scalability.
Source paper
extracted_from(2025) · Adam Karvonen
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Scaling result over Qwen-3 8B, Gemma-2 9B, Llama-3.3 70B, Claude Haiku 3.5.
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.
- Overall positioning.
- Limitation; confabulation risk.
- AOs articulate responses with natural-language flexibility rather than a fixed concept setclaim0.820Contrast with SAE features.
- Supervised method training models to answer questions about activations; NLAs differ by being unsupervised.
- Limitation relative to SAEs.
- AOs can serve as probes on demand using only a natural-language property specificationhypothesis0.790Hope for generalization beyond trainable supervised probes.
- AOs never saw fine-tuned model activations in training.
- Proposed future direction.