Episode 0071: The Canonical Evidence for Superposition Was a V

Nick Cammarata, who found the two uninterpretable neurons that Zoom In used to warn the field that networks are not clean, has retracted both of them. Neuron four e fifty-five is not cats-and-cars; it is a V shape. Cars in Superposition was not hiding a car in the dogs; cars and dogs shared a pose, and InceptionV1 has over a hundred unnamed pose neurons in plain sight. Fry and Bob work through the retraction, the recursive generalization hypothesis that replaces it, the ConvNeXt models that come out 99 to 100 percent interpretable in their last layer across a 311x size range, and what all of this would mean for the sparse autoencoder program if it holds. Bob does not let it pass unchallenged: no companion paper, no released weights, no code, and an author who puts his own confidence at 85 percent.

Why it matters. Superposition is the load-bearing assumption under most of the money and most of the engineering in interpretability right now. If the canonical evidence for it was a misread, the field spent six years building tools to dig out features that were never buried. That is either the most useful correction mechanistic interpretability has produced, or it is a hypothesis with a browsable microscope and no paper behind it yet. This episode holds both possibilities open and names the numbers on each side.

Source

• Nick Cammarata, Goodfire — Neural Networks Are Already Clean (2026-10-15): https://already-clean.goodfire.com/
• Plain text: https://already-clean.goodfire.com/essay.md
• Microscope (small / medium / large / InceptionV1): https://already-clean.goodfire.com/microscope/medium.html
• Circuit atlas: https://already-clean.goodfire.com/circuit-atlas/index.html
• Training map: https://already-clean.goodfire.com/training-map.html
• Olah, Cammarata, Schubert, Goh, Petrov, Carter — Zoom In: An Introduction to Circuits, Distill (2020): https://distill.pub/2020/circuits/zoom-in/

Numbers named in the episode

• 99–100% of last-layer neurons interpretable in a ConvNeXt of InceptionV1's size
• 311x parameter range held legible (92k to 28.6M)
• ~90% of Qwen3-32B neuron poles doing one job
• ~85% — the author's own stated confidence in the hypothesis
• No companion paper, no released weights, no released code as of 2026-10-16

Credits

Daily Tech Feed: From the Labs. Scripts written by Claude Opus. Voices rendered by Qwen3-TTS. Not affiliated with any companies or researchers discussed.