Single direction, four behaviors: interpretability debate with Neel Nanda over polysemantic vectors
johnhewtt · x · 2026-11-10
Responding to Neel Nanda, the author explains that a single direction v he found maps through four norms to four distinct behaviors (French, code style, sycophancy). Nanda argued that under the Linear Representation Hypothesis such a vector should decompose into sparse monosemantic directions. The author counters that the properties of these four norms break monosemanticity — the decompositions can't stay independent across behaviors — making the question subtler than plain polysemanticity. A technical exchange on the limits of LRH.
More from Research
- Agents build Lean models of every CPython 3.15.2 function, validated on 50,000+ generated inputs — ChrSzegedy · 2026-11-10
- AutoVend benchmark: 4 of 5 LLM agents lose to a rule-based baseline running 20 vending machines — Qiming Shi · 2026-11-10
- Claude 5.2 hits 91.8% on Pfam protein domain classification, far ahead of predecessors — chaitjo · 2026-11-10
- 2021 paper still praised: Shannon entropy is a derivation of the topological simplex operad — burny_tech · 2026-11-10
- NeurIPS position paper: LLM reports quietly dropped calibration curves since GPT-4 — burny_tech · 2026-11-10
- 'Bitter lesson for robotics' gets a sharp rebuttal: it's still human-orchestrated end to end — burny_tech · 2026-11-10