New interpretability take: superposed neurons encode generalized concepts like V-shapes
burny_tech · x · 2026-10-16
- A new mechanistic interpretability essay reframes superposition: neurons once thought to ambiguously mix concepts like cats and cars actually encode generalized features such as "V-shaped images" that both happen to match.
- This suggests models are built almost entirely of understandable generalized concepts, making interpretability—and control—more tractable than previously assumed.
- The booster quips that it may have been a "skill issue" in labeling all along.
More from Research
- Gemini is a robotics outlier: dropping it lifts capability-robotics correlation from 0.20 to 0.96 — ycombinator · 2026-10-16
- EPFL team uses RFdiffusion to design de novo metamorphic proteins switching between two folds — ferruz_noelia · 2026-10-16
- Manipulating a model's sense of power may reveal its takeover counterfactuals — jd_pressman · 2026-10-16
- Poser benchmark unmasks alignment-faking LLMs, one detection method catches 98% — jd_pressman · 2026-10-16
- Six Years of Evidence That MLP Neurons Are an Interpretable Basis, in One Thread — aryaman2020 · 2026-10-16
- First Preprint From New Group: Predictive Grid Cells Don't Need Explicit Planning — tweetsatpreet · 2026-10-16