Independent re-run of Goodfire's SAE steering claim on SDXL-Turbo holds up
kevrussell · x · 2026-10-27
Honesty.tools re-ran one of GoodfireAI's "painting with concepts" claims on SDXL-Turbo: turning up SAE feature 7446 ("opening one's mouth") opens mouths in generated images. In the pre-registered form of the test, the claim holds up.
Related event: Independent Replication Confirms Goodfire's SAE Feature Steering Claims(4 posts)→
More from Research
- Firecrawl's FantasyBench: agents hit 80.5% on 272 fantasy football questions, beating web search by 8.8 points — devdigest · 2026-10-27
- The Enormous Theorem — classification of finite simple groups — has been fully formalized — IgorCarron · 2026-10-27
- Largest math proof in history verified: 15,000-page Classification of Finite Simple Groups fully formalized in Lean — stevenstrogatz · 2026-10-27
- Goodfire proposes VPD: decomposing LLM parameters into interpretable subcomponents — CatAstro_Piyush · 2026-10-27
- Could huge constants make P=NP irrelevant at universe scale? — sytelus · 2026-10-27
- Neuroscience citations in AI papers are mostly textbook common knowledge — ShahabBakht · 2026-10-27