Removing three principal components collapses frontier model accuracy from 95.7% to 54.6%
kalomaze · x · 2026-11-01
kalomaze shares a fun interpretability experiment: there's a sharp knee at removing three leading principal components after normalization — removing two leaves 95.7% accuracy / 0.992 AUROC, while removing three drops to 54.6% / 0.518. The author jokes that their goal is sometimes to trigger the "this fucking guy" interp circuit in frontier models.
More from Fun
- Claude Opus 5.5 builds glass-breaking game with fully code-generated audio — prasenx · 2026-11-01
- Pincer1: a six-legged robot cold wallet that signs with your face and opens a decoy under coercion — R4nsack · 2026-11-01
- Sequoia Partner Shaun Maguire Challenges Critic to Public Financial Modeling Duel — shaunmmaguire · 2026-11-01
- Anonymous anime group offered huge money to post-train open models on yuri data — basedjensen · 2026-11-01
- User reports second encounter with a 'collective GPT psychosis' moment — VoidStateKate · 2026-11-01
- Ethan Mollick: slop reply bots now evolve to telling lame jokes for likes — emollick · 2026-11-01