Neural Networks Are Already Clean

TL;DR

Cammarata (now at Goodfire, formerly OpenAI’s first large-scale automated-interpretability project) argues that the canonical case for superposition in neural networks was misread. The famous “cat + car” polysemantic neuron (InceptionV1 4e:55) is actually a V-shape detector; the famous “car-into-dogs” superposition circuit was a shared-pose generalization misread through feature-visualization reward hacking. The last layer of InceptionV1 looked ~0% interpretable because the 2014 architecture overflowed deep layers, not because of a second learning principle — a modern ConvNeXt of the same size and dataset is ~99% interpretable in the final layer. Modern language models up to Qwen3-32B and trillion-parameter Kimi K3 look ~90% monosemantic (one main job per neuron pole), ~7% related offshoots, ~2–4% not-yet-understood under modern agentic interpretability tools. The remaining apparent polysemanticity is “recursive generalization” — a neuron grows along one axis then sprouts fractal sub-branches (an “iy” neuron grows to detect “diy” then “do it yourself”) — not a knot. The upshot is a bitter-lesson claim for interpretability: scaling ordinary normal science on the network’s own neurons and circuits beats engineering clever decompositions (SAEs). A technical companion paper is promised.

Key claims

• Superposition — the hypothesis that networks fold unrelated concepts into shared neurons when they run out of room — was the wrong explanation for the canonical polysemanticity examples; the right explanation is recursive generalization, which groups related concepts [§The Short Version, §Cars and Animals].
• InceptionV1’s famous “cat + car” neuron (4e:55) is a V-shape detector: its circuit connects to V-shaped flames, V-shaped water jets, V-shaped fox faces, and its negative pole is an O-shape; the whiskers-and-chrome reading was feature-visualization reward hacking that fell into the dog-neuron basin [§Cars and Animals].
• The canonical “car-into-dogs” superposition circuit from Zoom In (2020) is actually a shared-pose circuit — InceptionV1 has >100 generalized-pose neurons that aggregate across cars, dogs, and other objects, continuing to differentiate rather than hiding information in a high-dimensional basis [§Cars and Animals].
• InceptionV1’s late layers looked ~0–5% interpretable in 2020 because of architectural overflow, not a new learning principle; a ConvNeXt of the same parameter count and training data is ~99% interpretable in the final layer, retrained InceptionV1 with modern regularization is ~70%, and better interpretability tools (fixed feature-vis, attribution) raise the original 2014 checkpoint to ~50% [§A History of Microscope Errors].
• Modern language models (Qwen3-32B dense, Kimi K3 MoE trillion-parameter) are ~90% monosemantic single-pole neurons with modern agentic tools and ~7% related-offshoot pairs, with only ~2–4% unexplained — compared to 1.7% interpretable under OpenAI’s 2023 automated-interpretability paper and 0/14 median rubric score in Anthropic’s Towards Monosemanticity [§The Short Version].
• Recursive generalization is the dominant motif in the remaining apparent polysemanticity: a neuron detecting “iy” grows into “diy” then into “do it yourself” (no “iy” string at all), and tracking neurons across training shows they consistently sprout along semantic facets of their origin examples [§The Short Version, §Recursive Generalization].
• Neurons are often better described as verbs than nouns: they draw a distinction between positive and negative poles (V vs O, feeling-cue vs no-feeling-cue) that downstream circuits use, so describing them as “fires on cats and cars” misses the shared visual pattern [§Cars and Animals].
• Interpretability has crossword-puzzle dynamics: understanding compounds as the map fills in, because every neuron is constrained by its upstream builders, downstream readers, steering effects, and negative pole — nearly-empty and mostly-filled-in are natural anchor regimes, and the latter is where unexpected shared motifs (e.g. a double-negative regime in SwiGLU neurons) become visible [§The Short Version].
• Agents can scale normal-science interpretability — mapping neurons and circuits, building explorable microscopes, tracking hundreds of thousands of attribution graphs — to the point where ordinary investigation of a network’s own mechanisms becomes competitive with (and possibly superior to) decomposition-based approaches like SAEs [§The Short Version].
• Simulation-score metrics from 2023 rewarded explanations that could simulate activations rather than ones that were right — the two come apart often enough that the whole generation of “automatic interpretability” results built on them was biased against the correct, more compact, more semantic explanations [§The Short Version].

Method

The post is an essay-form argument backed by three empirical moves. (1) Re-examine the canonical polysemanticity examples: open the 4e:55 cat/car image grid, look at circuit connections (V-shaped flames and water jets), negative pole (O-shape), and conclude it’s a V-detector. Re-examine the Zoom In car-into-dogs circuit: notice the dogs share pose with the car, and >100 generalized-pose neurons continue to differentiate downstream — nothing reads the “combination.” (2) Train modern convolutional networks (ConvNeXt) at matched parameter count and dataset to InceptionV1 and measure the fraction of interpretable neurons in the final layer under a human + agentic interpretability stack: ~99% for ConvNeXt vs ~5% for original InceptionV1 vs ~70% for retrained InceptionV1 with modern regularization — isolates the architectural-overflow story. (3) Audit Qwen3-32B and larger models (up to Kimi K3) under agentic interpretability tooling, counting neuron poles by number-of-main-jobs: ~90% with one main job, ~7% with a related offshoot, ~2–4% not-yet-understood. The post also demonstrates recursive generalization by watching individual neurons form across training checkpoints (e.g. the “iy” → “diy” → “do it yourself” neuron).

Results

Qwen3-32B: ~90% of neuron poles have one main job (± related offshoots), ~7% have two related jobs, ~2–4% not-yet-understood under modern agentic tools. Trillion-parameter Kimi K3 similar. InceptionV1 final layer: ~0–5% interpretable in 2020 → ~50% with improved tools alone → ~70% with retrained matched architecture + modern regularization → ~99% in a ConvNeXt of the same size and dataset. 300× scale sweep of the ConvNeXt-of-InceptionV1-size family: broadly legible at every scale, with neurons covering coarser concepts at small scale (one car-neuron for many cars) and differentiating into specific cars at large scale. Compared to Anthropic’s 2023 Towards Monosemanticity (median neuron scored 0/14 on the rubric) and OpenAI’s 2023 automated-interpretability paper (~1.7% interpretable neurons under the most generous measure), these are 50–60× improvements with no change in model class.

Why it’s interesting

This is a direct, named challenge to the superposition hypothesis — the theoretical foundation under essentially every SAE-based interpretability program filed under Mechanistic Interpretability, including Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment (Universal SAEs across DinoV2/SigLIP/ViT) and Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations (Natural Language Autoencoders deployed inside Opus 4.6’s pre-deployment audit). If Cammarata is right that recursive generalization explains the canonical examples and modern models are 90%+ monosemantic under agentic tools, then the SAE-and-decomposition line is treating a symptom (bad microscopes + undersized models + weak metrics) rather than a cause (two learning principles) — and the field’s considerable investment in SAEs looks like a detour rather than a prerequisite. The crossword-dynamics argument also complements Structuring Sparsity: Block-Sparse Featurizers Capture Visual Concept Manifolds‘s “concepts are subspaces, not directions” thesis: both point at SAE 1D atoms as fragmenting native multidimensional structure, but from opposite directions (BSFs enrich the decomposition basis; this essay argues the decomposition itself is unnecessary). The constructive thread connects to MLPs are Hebbians: Constructing Efficient Fact-Storing MLPs for Transformers‘s constructive MLP memory: both make the case for interpretability-as-normal-science rather than interpretability-as-decomposition.

See also

• Mechanistic Interpretability — direct challenge to the superposition hypothesis that underwrites much of this concept’s SAE-based line of work.
• Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment — Universal SAEs are a flagship instance of the decomposition-first approach this essay argues against.
• Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations — NLAs are a different decomposition-flavored tool deployed inside Anthropic’s audit; the “interpretability has momentum” argument here is consistent with NLAs’ success as the map fills in.
• Structuring Sparsity: Block-Sparse Featurizers Capture Visual Concept Manifolds — adjacent “SAE 1D atoms fragment multidimensional structure” argument from a different angle (enrich the basis vs discard the basis).
• MLPs are Hebbians: Constructing Efficient Fact-Storing MLPs for Transformers — constructive-interpretability companion: write a fact into a known MLP location rather than reverse-engineer it.
• A Mathematical Framework for Transformer Circuits — the 2021 Anthropic paper this essay is in dialogue with on circuits-as-vocabulary.