Neural nets already interpretable? Researchers revisit polysemantic neurons and superposition
The article argues that many phenomena read as arbitrary concept mixing in neural networks can be explained by recursive generalization, so studying neurons and circuits directly may be more feasible than expected. The researchers re-examine InceptionV1's cat-face and car neurons, reading them as shared V-shaped patterns, and observe coarser neuron concepts in compressed-network experiments. A preliminary review of Qwen3-32B estimates about 90% of neurons have one dominant role, though the share depends on the threshold; the article labels the prospect of large-scale interpretability as speculation.
Why it matters: It lays out concrete research paths, tracking neuron training and checking positive and negative activations, to separate recursive generalization from tooling error and learning failure.