claim
active
claim:clt-feature-space-is-somewhat-less-clean-than-hanna-et-al-s-with-some-narrow-special-case-lookup-like-featuresCLT feature space is somewhat less clean than Hanna et al.'s, with some narrow special-case-lookup-like features
Difference between sparse and neuron analyses.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Papers (1)
paper
Findings (5)
finding
- Special-case lookup-like feature.
- Unexpected behavior; may be general small-number detector.
- More complex than simple comparison.
- Comparison of heatmaps.
- Diverges from Hanna et al. heatmaps.
Hypotheses (1)
hypothesis
- Aggregating feature scores over multiple prompts would weed out distractor features and isolate a cleaner circuit closer to Hanna et al.associated_withgatesLimitation of single-prompt selection.
Claims (1)
claim
- Hinted by differences.
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.
- Main conclusion.
- Closing claim.
- Key advantage statement.
- The space inside the 'C' of the diskette cover is not I-like (phenomenological judgment).finding0.729Specific result of the comparison probe: the rounded rectangle space with keyhole shape fails the relatedness test.
- CLTs uncover relevant structure out of the box without labor-intensive experimental interventionsclaim0.727Advantage over original analysis.
- Subclaim.
- The J-space can hold more concepts when they have coherent relationships than unrelated ones.hypothesis0.725Tested with lists.
- Load-bearing interpretive claim about the layer-specificity of Burger et al.'s finding.