Interpretability researcher: superposition should be the last hypothesis, not the null
hrishioa · x · 2026-10-16
Quoting a mechanistic interpretability essay, hrishioa endorses its core reflection: across hundreds of cases, unexplained activations were repeatedly attributed to superposition moving to weaker neurons or combinations — effectively treating superposition as the null hypothesis.
The essay's author now argues superposition should be the last hypothesis, since it's the most radical and least testable: phenomena that don't make sense could equally stem from reward-hacked feature visualization, poorly understood recursive generalization, or bad architecture. Anthropic's 2022 Softmax Linear Units paper admitted superposition was often the default explanation on intuition alone, with little evidence.
Related event: Researchers Question Superposition as Default Interpretability Assumption(2 posts)→
More from Research
- Sakana AI ships Agentic Scientist, orchestrating coding agents for automated scientific discovery — hardmaru · 2026-10-16
- MAI-Base-2 adopted the 1/(weight_decay·lr) timescale hparam scaling rule — SeunghyunSEO7 · 2026-10-16
- 108M-parameter text-to-image denoiser Surjo-Image-Preview lands on Hugging Face — Consistent-Rule-3720 · 2026-10-16
- Lean proof formalizes optimal packing of 19 unit squares, side ≥ 3 + 4√2/3 — ctjlewis · 2026-10-16
- Proposal: 'behavior heritability' — the causal share of a neuron's correlation with LLM behavior — Sauers_ · 2026-10-16
- Quiz: Double the Residual Stream from 4096 to 8192 — How Does Param Count Change? — Sauers_ · 2026-10-16