Researcher reports architecture changes make InceptionV1 neurons more interpretable

A researcher writing on Goodfire's site reports that ordinary architectural improvements made 99-100% of neurons in InceptionV1's last layer easily interpretable, without changing model size or training data. Earlier work had found almost no interpretable neurons in that layer. The researcher attributes the earlier difficulty to architectural problems and weak analysis tools, and reports broadly legible versions across a 300× range in model size. Separate preliminary audits of Qwen3-32B suggest many neurons have one main role, but the author cautions that the percentages depend on thresholds and judgments about what counts as one task.