concept
active
concept:pathological-scaling-regimePathological scaling regime (β<α)
SAEs learn far fewer features than latents.
Neighborhood — ranked by edge-count
Papers (1)
paper
Concepts (2)
concept
- Manifold tilingassociated_withLatents tile a manifold to reduce loss.
- Feature discovery count D(N) and efficiency D(N)/Nassociated_withMetric of features discovered.
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.
- Left open.
- Mechanisms by which smaller competent subunits bind into a higher-level Self with larger goals; key example via gap junction connections.
- Techniques that leverage AI to help humans more efficiently supervise AI.
- Scaling aggregated gradient by the maximum gradient norm among tasks.
- Hypothesis cited in paper suggesting deceptive capabilities may scale with model size
- Observation that SAE loss decreases as a power law with compute budget.
- Theoretical result showing that any positive typicality bias weight γ-sharpens the reference distribution, amplifying modes
- The paper's claim that theoretical convergence across GWT, RPT, HOT, IIT makes the findings non-coincidental