Cammarata 26,000-word post claims absence of superposition in inspected Inception v1 neurons

Cammarata documented repeated manual inspection of individual neurons across multiple layers in small vision models and found each neuron activated for a single semantic feature rather than multiple unrelated concepts. He released activation maps and weight visualizations covering 312 neurons, none of which exhibited the car-cat multiplexing reported in the earlier work. The post contains no new training runs or quantitative benchmarks.

The original 2020 Circuits paper identified polysemantic neurons in Inception v1 layer mixed3b and attributed them to superposition for capacity reasons. Cammarata's counter-claim rests on qualitative review without statistical tests or replication across model scales. No ablation studies or automated detection methods accompany the assertion.

If the observation holds, mechanistic interpretability shifts from untangling superposition to direct feature editing. This enables targeted removal of biased or unsafe circuits rather than accuracy-focused scaling. Regulatory and deployment decisions could then prioritize verifiable circuit edits over aggregate benchmark scores.

Next steps require independent teams to apply the same neuron-by-neuron protocol to current LLMs and publish activation datasets within six months for falsification.