Composition gives more combinations, not more features: rebutting an anti-superposition essay
livgorton · x · 2026-10-16
In part 11 of her thread, livgorton rebuts the essay's claim that "more features than dimensions" is satisfied even by a fully monosemantic network because its neurons compose: composition gives you more combinations, not more features.
She adds that the essay's own readout story sounds a lot like superposition — the car reader does get cat signal through the shared cat/car neuron, tolerable only because it combines many neurons.
More from Research
- New interpretability take: superposed neurons encode generalized concepts like V-shapes — burny_tech · 2026-10-16
- METR's famous time-horizon chart may measure something else — latent capability fits with R²=0.996 — wfithian · 2026-10-16
- Anthropic's GRAFT routes flagged gradients into a deletable adapter, cutting reward hacking to near zero — rohanpaul_ai · 2026-10-16
- Someone found a hilariously specific 'betrayed by your own side' neuron in Qwen3-32B — NathanpmYoung · 2026-10-16
- "Decompiling Neural Networks into Code" Claim Draws Fire Over Its Theory — fkasummer · 2026-10-16
- Burkov's AI weekly #182: State of AI 2026, AI protein design, 260M-param diffusion beats 6x larger baselines — burkov · 2026-10-16