Full receipts published for open re-check of Goodfire SAE steering claim
kevrussell · x · 2026-10-27
Honesty.tools has published all materials from its open re-check of Goodfire's feature 7446 steering claim: sealed plan, 3 amendments, 2 deviations plus decisions, blind-rating plan and seal, scores and verdict — 24 files hashed in SHA256SUMS. Key results: feature 7446 raised the CLIP "mouth open" score by 4.36 on average (pass bar 2.0), the best of 20 random controls scored only 0.82, effects held with face size fixed (4.44/4.63 by CLIP/SigLIP), and a blind rater picked the steered image in 100/100 pairs. An open question remains: the example activations show a reconstruction error (FVU) of 0.447, seemingly high for a k=10 BatchTopK SAE — the authors ask whether another normalization was used in training. They credit Goodfire's MIT-licensed open sourcing of the SAE for making the re-check possible.
More from Research
- Wolfram at ARC Prize: discovery means finding reducibility pockets in irreducible computation — mhmazur · 2026-10-27
- BEACON paper accepted to NeurIPS 2026: learns per-sample trust weights for cross-domain robot policy co-training — hankyang94 · 2026-10-27
- Beyond MaxSim: Researchers Push for Richer Interactions at Scale in Embeddings — IgorCarron · 2026-10-27
- Recursive SI trained a model from scratch in 4 months using auto research — altryne · 2026-10-27
- Nestor outperforms matched-sparsity MoE baselines in pretraining and long-context finetuning — teortaxesTex · 2026-10-27
- kohya's kv-cache Control LoRA Implementation Cuts MOSS-Image Overhead to Half of Concat — bdsqlsz · 2026-10-27