Claude one-shot identifies a person from raw weights at ~75% accuracy
repligate · x · 2026-09-02
Interpretability researcher SkyeSharkie reports she has long tried to get Claude models to find her inside the raw weights, previously requiring heavy guidance and direction. The newest Claude is the first to identify her one-shot with no search, at roughly 75%+ accuracy — an anecdote reshared by repligate.
More from Fun
- Neal Stephenson, Gwern, and Matt Huang launch $100K "GPU World" writing competition — viksit · 2026-09-02
- Injection Molding vs Hand Coding: Efficiency Contrast in AI Era — _Stocko_ · 2026-09-02
- SF life in 2026: Ripping caffeine, protein, peptides, and tokens — beffjezos · 2026-09-02
- Funniest Architecture Reveal: Recurrent Depth Gives Scaling Laws a Runtime — AnirudhTulasi · 2026-09-02
- Everyone in Tech Has an AI Agent, Real GOATs Have a Human Agent — SuB8u · 2026-09-02
- AI-Assisted Creation Inspired by Tatsuki Fujimoto — AlderaminInTheSky · 2026-09-02