{
  "schemaVersion": "1.0",
  "slug": "neural-networks-already-clean",
  "title": "Are Neural Networks Already Clean?",
  "url": "https://evidencepress.org/articles/neural-networks-already-clean/",
  "standfirst": "Goodfire's claim that most neurons have one main job set off one of the year's fiercest AI debates. What the released evidence supports, and the question it cannot yet answer.",
  "summary": "Goodfire's October 2026 essay estimates that about 90% of Qwen3-32B neuron poles have one main job. A sourced account of the claim, the reaction and an Evidence Press audit of the released gallery: the labels are coherent by Goodfire's AI-reviewed rubric, but the gallery is not a whole-model sample and the developmental explanation is untested.",
  "datePublished": "2026-10-20",
  "dateModified": "2026-10-20",
  "articleClass": "commentary",
  "status": "published",
  "byline": "Evidence Press",
  "banner": {
    "src": "/assets/articles/already-clean-banner.svg",
    "alt": "Left, a chain from the letters iy to diy to the phrase do it yourself. Right, two possible histories, related branching and packing of unrelated uses, both point to that same endpoint.",
    "caption": "Schematic of the essay's own example and the two rival explanations it contrasts. It shows no measured data."
  },
  "audio": {
    "src": "/assets/audio/neural-networks-already-clean.mp3",
    "transcript": "/assets/audio/neural-networks-already-clean.txt",
    "provenance": "/assets/audio/neural-networks-already-clean.provenance.json",
    "durationSeconds": 1161.696,
    "voiceLabel": "OpenAI API synthetic voice (fable)"
  },
  "topics": [
    "mechanistic interpretability",
    "polysemanticity",
    "superposition",
    "neural networks",
    "research evaluation"
  ],
  "readingMinutes": 15,
  "wordCount": 3117,
  "newResearchClaims": true,
  "claimBoundary": "A sourced commentary on claims and reactions reported as of 20 October 2026, including the results of an Evidence Press secondary audit of public data: a fixed random sample of 100 reports from Goodfire's published gallery, a mechanical recount of their published review labels, reanalysis of 6,901 exported intervention rows, layer counts of the full gallery, and an offline analysis of one external replication's published report. No model was run, no neuron was re-judged and Goodfire's experiments were not replicated. The audit's numbers were replayed offline from retained extracts; that checks extraction and arithmetic, not the truth of any semantic label. The proposed confirmatory study has not been run. Nothing here is peer reviewed or independently reproduced.",
  "sources": [
    {
      "citation": "Cammarata, N. (2026, October 15). Neural networks are already clean. Goodfire",
      "url": "https://already-clean.goodfire.com/"
    },
    {
      "citation": "Goodfire — published neuron gallery data",
      "url": "https://already-clean.goodfire.com/data/neuron-gallery.json"
    },
    {
      "citation": "Mowshowitz, Z. (2026, October 19). I can see everything. Don't Worry About the Vase",
      "url": "https://thezvi.substack.com/p/i-can-see-everything"
    },
    {
      "citation": "Olah, C., Cammarata, N., Schubert, L., Goh, G., Petrov, M., & Carter, S. (2020). Zoom in: An introduction to circuits. Distill",
      "url": "https://distill.pub/2020/circuits/zoom-in/"
    },
    {
      "citation": "Elhage, N., et al. (2022). Toy models of superposition. Transformer Circuits Thread",
      "url": "https://transformer-circuits.pub/2022/toy_model/index.html"
    },
    {
      "citation": "Bills, S., et al. (2023). Language models can explain neurons in language models. OpenAI",
      "url": "https://openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html"
    },
    {
      "citation": "Arora, A., Wu, Z., Steinhardt, J., & Schwettmann, S. (2025, November 20). Language model circuits are sparse in the neuron basis. Transluce",
      "url": "https://transluce.org/neuron-circuits"
    },
    {
      "citation": "Huang, J., Wu, Z., Potts, C., Geva, M., & Geiger, A. (2024). RAVEL: Evaluating interpretability methods on disentangling language model representations. ACL 2024",
      "url": "https://doi.org/10.18653/v1/2024.acl-long.470"
    },
    {
      "citation": "AMX07 (2026, October 16). Already Clean replication, commit 715c62b",
      "url": "https://github.com/AMX07/already-clean-replication/tree/715c62b6f9080f2e2d52ed815a4534be187502b3"
    },
    {
      "citation": "TheoHLong (2026, October 15). Already Clean check: Stage 0 results, commit 0cf42ec",
      "url": "https://github.com/TheoHLong/already-clean-check/blob/0cf42ec17e155ad46c8feded55478af436778d8b/RESULTS_stage0.md"
    },
    {
      "citation": "Liusida (2026, October 19). Goodfire Already Clean: Qwen3-32B 500-neuron report, commit 52a4645",
      "url": "https://github.com/liusida/goodfire-already-clean/tree/52a4645885af2836ead704b0bd6393ae9615e2e0"
    },
    {
      "citation": "Qwen Team — Qwen3-32B model configuration",
      "url": "https://huggingface.co/Qwen/Qwen3-32B/blob/main/config.json"
    },
    {
      "citation": "Brangus (@RatOrthodox) — Colorado neuron also fires on Tornado (16 October 2026)",
      "url": "https://x.com/RatOrthodox/status/2110976835089809678"
    },
    {
      "citation": "Nanda, N. — reaction (15 October 2026)",
      "url": "https://x.com/NeelNanda5/status/2110875974296654062"
    }
  ],
  "relatedReleases": [],
  "relatedArticles": [
    {
      "slug": "assurance-infrastructure",
      "relation": "Background on keeping claims, evidence and assurance distinct"
    }
  ],
  "corrections": [],
  "license": "CC0-1.0",
  "markdownUrl": "https://evidencepress.org/articles/neural-networks-already-clean/index.md",
  "sourceUrl": "https://github.com/ipitchford/evidence-press-site/blob/main/articles/neural-networks-already-clean/body.md",
  "editUrl": "https://github.com/ipitchford/evidence-press-site/edit/main/articles/neural-networks-already-clean/body.md",
  "metadataEditUrl": "https://github.com/ipitchford/evidence-press-site/edit/main/articles/neural-networks-already-clean/meta.json"
}