Interpretability researcher disputes superposition claim after probing a single neuron
livgorton · x · 2026-10-17
Interpretability researcher livgorton pushed back on @tszzl's argument about superposition with three points from her own investigation:
- Superposition as classically defined is not arbitrary — it is highly structured by the data distribution.
- She doesn't buy the claimed connectivity of the neuron in question: its weak activations are almost universally tied to one concept (Colorado) with only a minor second peak on an unrelated one — "just a slightly messy neuron" she'd bet disappears with scale, not useful superposition.
- Even granting recursive generalisation, it doesn't disprove superposition: a claim about training dynamics cannot substitute for a claim about the model's final geometry (non-uniform superposition).
The exchange touches a core interpretability debate: whether a single-neuron case genuinely undermines the widely cited superposition hypothesis.
Related event: Researchers Challenge Superposition Hypothesis with Single-Neuron Evidence(2 posts)→
More from Research
- NIST JARVIS team publishes 220+ AI materials science talks on YouTube after a decade of workshops — CatAstro_Piyush · 2026-10-17
- Study of 357K science posts: Bluesky discourse is more academic, X is easier to read — CurieuxExplorer · 2026-10-17
- visloc-rs: pure Rust SfM and 3DGS localization, 3.46x faster than COLMAP — rsasaki0109 · 2026-10-17
- Over 90% of neurons in image networks align with human concepts, suggesting one true language — CatAstro_Piyush · 2026-10-17
- Sleeping Beauty, centered world models, and why 1/3 is the right credence — jessi_cata · 2026-10-17
- Implicit Curriculum Hypothesis: LLM Skills Emerge in Predictable Order, CMU Paper Finds — CatAstro_Piyush · 2026-10-17