168 of 4,096 A/1 features are dead and 292 are ultralow density, leaving 3,636 for analysis

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• 34M SAE had roughly 65% dead features.finding0.764
  Most features dead in largest SAE, indicating room for improvement.
• An ordinary American house contains a density of about 1 decision per cubic foot, thus about 20,000 possible mistakes when the house is not generated.claim0.740
  Estimate based on labor hours and physical pieces; used to motivate economic cost analysis.
• For all three SAEs (1M, 4M, 34M), average active features per token <300, and reconstruction variance explained ≥65%.finding0.733
  Basic SAE performance metrics.
• Features in A/1 have median activation correlation of 0.72 with most similar feature in B/1; neurons have median 0.46finding0.727
  Systematic comparison showing features are substantially more universal than neurons across models
• PC1 explains 82% of variance in factor analysis of 2224 data points across 6 scoring dimensionsfinding0.719
  Dimensions are not independent; composite score is the reliable signal; six dimensions useful for understanding how not how much
• Higher-density priors (B10) are more robust to fine-tuning than lower-density ones (B9).claim0.718
  Interpretation of cross-base transfer asymmetry.
• Dose-response curves for six individual deception features show z=8.06, p=7.7×10⁻¹⁶ for suppression vs. amplification contrast on consciousness queryfinding0.718
  Statistical result confirming robustness of single-feature steering effects in Experiment 2
• DNA feature A/1/2937 and B/1/3680 have activation correlation of 0.92finding0.711
  Universality of DNA feature across two transformer models with different random seeds