Mechanistic Interpretability

{
  "id": "mech-interp",
  "wikiId": "E477",
  "path": "/knowledge-base/responses/mech-interp/",
  "filePath": "knowledge-base/responses/mech-interp.mdx",
  "title": "Mechanistic Interpretability",
  "quality": 59,
  "readerImportance": 40,
  "researchImportance": 47,
  "tacticalValue": null,
  "contentFormat": "article",
  "causalLevel": null,
  "lastUpdated": "2026-01-29",
  "dateCreated": "2026-02-15",
  "summary": "Mechanistic interpretability aims to reverse-engineer neural networks to understand internal computations, with \\$100M+ annual investment across major labs. Anthropic extracted 30M+ features from Claude 3 Sonnet (2024), while DeepMind deprioritized SAE research after finding linear probes outperform on practical tasks; Amodei predicts 'MRI for AI' achievable in 5-10 years but warns AI may advance faster, with 3 of 4 blue teams detecting planted misalignment using interpretability tools.",
  "description": "Mechanistic interpretability reverse-engineers neural networks to understand their internal computations and circuits.",
  "ratings": {
    "novelty": 4.5,
    "rigor": 6.8,
    "completeness": 7.5,
    "actionability": 5.2
  },
  "category": "responses",
  "subcategory": "alignment-interpretability",
  "clusters": [
    "ai-safety"
  ],
  "metrics": {
    "wordCount": 3592,
    "tableCount": 28,
    "diagramCount": 1,
    "internalLinks": 17,
    "externalLinks": 62,
    "footnoteCount": 0,
    "bulletRatio": 0.06,
    "sectionCount": 45,
    "hasOverview": true,
    "structuralScore": 15
  },
  "suggestedQuality": 100,
  "updateFrequency": 21,
  "evergreen": true,
  "wordCount": 3592,
  "unconvertedLinks": [
    {
      "text": "DeepMind deprioritized SAE research",
      "url": "https://deepmindsafetyresearch.medium.com/negative-results-for-sparse-autoencoders-on-downstream-tasks-and-deprioritising-sae-research-6cadcfc125b9",
      "resourceId": "244c1b93ef0a083c",
      "resourceTitle": "deprioritizing SAE research"
    },
    {
      "text": "Amodei 2025",
      "url": "https://www.darioamodei.com/post/the-urgency-of-interpretability",
      "resourceId": "f19373dc25a8a96f",
      "resourceTitle": "The Urgency of Interpretability - Dario Amodei"
    },
    {
      "text": "2025 analysis",
      "url": "https://quickmarketpitch.com/blogs/news/ai-safety-investors",
      "resourceId": "58d4f33d9ae57621",
      "resourceTitle": "Who is funding AI safety research?"
    },
    {
      "text": "2026 Breakthrough Technology",
      "url": "https://www.technologyreview.com/2026/01/12/1130003/mechanistic-interpretability-ai-research-models-2026-breakthrough-technologies/",
      "resourceId": "3a4cf664bf7b27a8",
      "resourceTitle": "Mechanistic interpretability: 10 Breakthrough Technologies 2026 | MIT Technology Review"
    },
    {
      "text": "Anthropic CEO Dario Amodei noted",
      "url": "https://www.darioamodei.com/post/the-urgency-of-interpretability",
      "resourceId": "f19373dc25a8a96f",
      "resourceTitle": "The Urgency of Interpretability - Dario Amodei"
    },
    {
      "text": "\"Zoom In: An Introduction to Circuits\"",
      "url": "https://distill.pub/2020/circuits/zoom-in/",
      "resourceId": "346b1574c0c3ce67",
      "resourceTitle": "Zoom In: An Introduction to Circuits"
    },
    {
      "text": "\"Scaling Monosemanticity\"",
      "url": "https://transformer-circuits.pub/2024/scaling-monosemanticity/",
      "resourceId": "e724db341d6e0065",
      "resourceTitle": "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet"
    },
    {
      "text": "Open Problems survey",
      "url": "https://arxiv.org/abs/2501.16496",
      "resourceId": "6ada11f74f5ff01a",
      "resourceTitle": "[2501.16496] Open Problems in Mechanistic Interpretability - arXiv"
    },
    {
      "text": "Anthropic 2024",
      "url": "https://transformer-circuits.pub/2024/scaling-monosemanticity/",
      "resourceId": "e724db341d6e0065",
      "resourceTitle": "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet"
    },
    {
      "text": "OpenAI 2024",
      "url": "https://cdn.openai.com/papers/sparse-autoencoders.pdf",
      "resourceId": "1175fcbd0a4ff6a3",
      "resourceTitle": "Scaling and evaluating sparse autoencoders | OpenAI"
    },
    {
      "text": "Amodei 2025",
      "url": "https://www.darioamodei.com/post/the-urgency-of-interpretability",
      "resourceId": "f19373dc25a8a96f",
      "resourceTitle": "The Urgency of Interpretability - Dario Amodei"
    },
    {
      "text": "OpenAI SAE paper",
      "url": "https://cdn.openai.com/papers/sparse-autoencoders.pdf",
      "resourceId": "1175fcbd0a4ff6a3",
      "resourceTitle": "Scaling and evaluating sparse autoencoders | OpenAI"
    },
    {
      "text": "DeepMind 2025",
      "url": "https://deepmindsafetyresearch.medium.com/negative-results-for-sparse-autoencoders-on-downstream-tasks-and-deprioritising-sae-research-6cadcfc125b9",
      "resourceId": "244c1b93ef0a083c",
      "resourceTitle": "deprioritizing SAE research"
    },
    {
      "text": "Anthropic internal testing",
      "url": "https://www.darioamodei.com/post/the-urgency-of-interpretability",
      "resourceId": "f19373dc25a8a96f",
      "resourceTitle": "The Urgency of Interpretability - Dario Amodei"
    },
    {
      "text": "30 million+ interpretable features",
      "url": "https://transformer-circuits.pub/2024/scaling-monosemanticity/",
      "resourceId": "e724db341d6e0065",
      "resourceTitle": "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet"
    },
    {
      "text": "16 million latent autoencoder on GPT-4",
      "url": "https://cdn.openai.com/papers/sparse-autoencoders.pdf",
      "resourceId": "1175fcbd0a4ff6a3",
      "resourceTitle": "Scaling and evaluating sparse autoencoders | OpenAI"
    },
    {
      "text": "announced they are deprioritizing fundamental SAE research",
      "url": "https://deepmindsafetyresearch.medium.com/negative-results-for-sparse-autoencoders-on-downstream-tasks-and-deprioritising-sae-research-6cadcfc125b9",
      "resourceId": "244c1b93ef0a083c",
      "resourceTitle": "deprioritizing SAE research"
    },
    {
      "text": "\"The Urgency of Interpretability\"",
      "url": "https://www.darioamodei.com/post/the-urgency-of-interpretability",
      "resourceId": "f19373dc25a8a96f",
      "resourceTitle": "The Urgency of Interpretability - Dario Amodei"
    },
    {
      "text": "Three of four teams found the planted flaw",
      "url": "https://www.darioamodei.com/post/the-urgency-of-interpretability",
      "resourceId": "f19373dc25a8a96f",
      "resourceTitle": "The Urgency of Interpretability - Dario Amodei"
    },
    {
      "text": "30+ researchers titled \"Open Problems in Mechanistic Interpretability\"",
      "url": "https://arxiv.org/abs/2501.16496",
      "resourceId": "6ada11f74f5ff01a",
      "resourceTitle": "[2501.16496] Open Problems in Mechanistic Interpretability - arXiv"
    },
    {
      "text": "80,000 Hours podcast interview",
      "url": "https://80000hours.org/podcast/episodes/neel-nanda-mechanistic-interpretability/",
      "resourceId": "8fe85c0841f301a8",
      "resourceTitle": "Neel Nanda on the race to read AI minds (part 1) | 80,000 Hours"
    },
    {
      "text": "Scaling Monosemanticity",
      "url": "https://transformer-circuits.pub/2024/scaling-monosemanticity/",
      "resourceId": "e724db341d6e0065",
      "resourceTitle": "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet"
    },
    {
      "text": "Circuits July 2025 update",
      "url": "https://transformer-circuits.pub/2025/july-update/index.html",
      "resourceId": "0a2ab4f291c4a773",
      "resourceTitle": "Circuits Updates - July 2025"
    },
    {
      "text": "16M latent autoencoder on GPT-4",
      "url": "https://cdn.openai.com/papers/sparse-autoencoders.pdf",
      "resourceId": "1175fcbd0a4ff6a3",
      "resourceTitle": "Scaling and evaluating sparse autoencoders | OpenAI"
    },
    {
      "text": "SAE deprioritization",
      "url": "https://deepmindsafetyresearch.medium.com/negative-results-for-sparse-autoencoders-on-downstream-tasks-and-deprioritising-sae-research-6cadcfc125b9",
      "resourceId": "244c1b93ef0a083c",
      "resourceTitle": "deprioritizing SAE research"
    },
    {
      "text": "Open Problems survey",
      "url": "https://arxiv.org/abs/2501.16496",
      "resourceId": "6ada11f74f5ff01a",
      "resourceTitle": "[2501.16496] Open Problems in Mechanistic Interpretability - arXiv"
    },
    {
      "text": "2026 Breakthrough Technology",
      "url": "https://www.technologyreview.com/2026/01/12/1130003/mechanistic-interpretability-ai-research-models-2026-breakthrough-technologies/",
      "resourceId": "3a4cf664bf7b27a8",
      "resourceTitle": "Mechanistic interpretability: 10 Breakthrough Technologies 2026 | MIT Technology Review"
    },
    {
      "text": "Scaling Monosemanticity",
      "url": "https://transformer-circuits.pub/2024/scaling-monosemanticity/",
      "resourceId": "e724db341d6e0065",
      "resourceTitle": "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet"
    },
    {
      "text": "Transformer Circuits",
      "url": "https://transformer-circuits.pub/",
      "resourceId": "5083d746c2728ff2",
      "resourceTitle": "Transformer Circuits Thread"
    },
    {
      "text": "Research page",
      "url": "https://www.anthropic.com/research/team/interpretability",
      "resourceId": "dfc21a319f95a75d",
      "resourceTitle": "Anthropic Interpretability Research Team"
    },
    {
      "text": "SAE deprioritization",
      "url": "https://deepmindsafetyresearch.medium.com/negative-results-for-sparse-autoencoders-on-downstream-tasks-and-deprioritising-sae-research-6cadcfc125b9",
      "resourceId": "244c1b93ef0a083c",
      "resourceTitle": "deprioritizing SAE research"
    },
    {
      "text": "Neel Nanda",
      "url": "https://80000hours.org/podcast/episodes/neel-nanda-mechanistic-interpretability/",
      "resourceId": "8fe85c0841f301a8",
      "resourceTitle": "Neel Nanda on the race to read AI minds (part 1) | 80,000 Hours"
    },
    {
      "text": "16M latent autoencoder",
      "url": "https://cdn.openai.com/papers/sparse-autoencoders.pdf",
      "resourceId": "1175fcbd0a4ff6a3",
      "resourceTitle": "Scaling and evaluating sparse autoencoders | OpenAI"
    },
    {
      "text": "Open Problems survey",
      "url": "https://arxiv.org/abs/2501.16496",
      "resourceId": "6ada11f74f5ff01a",
      "resourceTitle": "[2501.16496] Open Problems in Mechanistic Interpretability - arXiv"
    },
    {
      "text": "Feature Visualization",
      "url": "https://distill.pub/2017/feature-visualization/",
      "resourceId": "ce79e09da641a0b7",
      "resourceTitle": "Feature Visualization - Distill.pub"
    },
    {
      "text": "Zoom In: An Introduction to Circuits",
      "url": "https://distill.pub/2020/circuits/zoom-in/",
      "resourceId": "346b1574c0c3ce67",
      "resourceTitle": "Zoom In: An Introduction to Circuits"
    },
    {
      "text": "Scaling Monosemanticity",
      "url": "https://transformer-circuits.pub/2024/scaling-monosemanticity/",
      "resourceId": "e724db341d6e0065",
      "resourceTitle": "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet"
    },
    {
      "text": "SAE scaling research",
      "url": "https://cdn.openai.com/papers/sparse-autoencoders.pdf",
      "resourceId": "1175fcbd0a4ff6a3",
      "resourceTitle": "Scaling and evaluating sparse autoencoders | OpenAI"
    },
    {
      "text": "Open Problems in Mechanistic Interpretability",
      "url": "https://arxiv.org/abs/2501.16496",
      "resourceId": "6ada11f74f5ff01a",
      "resourceTitle": "[2501.16496] Open Problems in Mechanistic Interpretability - arXiv"
    },
    {
      "text": "deprioritizes SAE research",
      "url": "https://deepmindsafetyresearch.medium.com/negative-results-for-sparse-autoencoders-on-downstream-tasks-and-deprioritising-sae-research-6cadcfc125b9",
      "resourceId": "244c1b93ef0a083c",
      "resourceTitle": "deprioritizing SAE research"
    },
    {
      "text": "\"The Urgency of Interpretability\"",
      "url": "https://www.darioamodei.com/post/the-urgency-of-interpretability",
      "resourceId": "f19373dc25a8a96f",
      "resourceTitle": "The Urgency of Interpretability - Dario Amodei"
    },
    {
      "text": "Circuits July 2025 update",
      "url": "https://transformer-circuits.pub/2025/july-update/index.html",
      "resourceId": "0a2ab4f291c4a773",
      "resourceTitle": "Circuits Updates - July 2025"
    },
    {
      "text": "2026 Breakthrough Technology",
      "url": "https://www.technologyreview.com/2026/01/12/1130003/mechanistic-interpretability-ai-research-models-2026-breakthrough-technologies/",
      "resourceId": "3a4cf664bf7b27a8",
      "resourceTitle": "Mechanistic interpretability: 10 Breakthrough Technologies 2026 | MIT Technology Review"
    },
    {
      "text": "Zoom In: An Introduction to Circuits",
      "url": "https://distill.pub/2020/circuits/zoom-in/",
      "resourceId": "346b1574c0c3ce67",
      "resourceTitle": "Zoom In: An Introduction to Circuits"
    },
    {
      "text": "Transformer Circuits Thread",
      "url": "https://transformer-circuits.pub/",
      "resourceId": "5083d746c2728ff2",
      "resourceTitle": "Transformer Circuits Thread"
    },
    {
      "text": "Scaling Monosemanticity",
      "url": "https://transformer-circuits.pub/2024/scaling-monosemanticity/",
      "resourceId": "e724db341d6e0065",
      "resourceTitle": "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet"
    },
    {
      "text": "Scaling and Evaluating Sparse Autoencoders",
      "url": "https://cdn.openai.com/papers/sparse-autoencoders.pdf",
      "resourceId": "1175fcbd0a4ff6a3",
      "resourceTitle": "Scaling and evaluating sparse autoencoders | OpenAI"
    },
    {
      "text": "Open Problems in Mechanistic Interpretability",
      "url": "https://arxiv.org/abs/2501.16496",
      "resourceId": "6ada11f74f5ff01a",
      "resourceTitle": "[2501.16496] Open Problems in Mechanistic Interpretability - arXiv"
    },
    {
      "text": "The Urgency of Interpretability",
      "url": "https://www.darioamodei.com/post/the-urgency-of-interpretability",
      "resourceId": "f19373dc25a8a96f",
      "resourceTitle": "The Urgency of Interpretability - Dario Amodei"
    },
    {
      "text": "DeepMind SAE Deprioritization",
      "url": "https://deepmindsafetyresearch.medium.com/negative-results-for-sparse-autoencoders-on-downstream-tasks-and-deprioritising-sae-research-6cadcfc125b9",
      "resourceId": "244c1b93ef0a083c",
      "resourceTitle": "deprioritizing SAE research"
    },
    {
      "text": "Mechanistic Interpretability for AI Safety: A Review",
      "url": "https://leonardbereska.github.io/blog/2024/mechinterpreview/",
      "resourceId": "45c5b56ac029ef2d",
      "resourceTitle": "Mechanistic Interpretability for AI Safety — A Review"
    },
    {
      "text": "MIT Technology Review: 2026 Breakthrough Technologies",
      "url": "https://www.technologyreview.com/2026/01/12/1130003/mechanistic-interpretability-ai-research-models-2026-breakthrough-technologies/",
      "resourceId": "3a4cf664bf7b27a8",
      "resourceTitle": "Mechanistic interpretability: 10 Breakthrough Technologies 2026 | MIT Technology Review"
    },
    {
      "text": "Transformer Circuits",
      "url": "https://transformer-circuits.pub/",
      "resourceId": "5083d746c2728ff2",
      "resourceTitle": "Transformer Circuits Thread"
    },
    {
      "text": "NeurIPS Mechanistic Interpretability Workshop",
      "url": "https://mechinterpworkshop.com/",
      "resourceId": "e78a965cde8d82bd",
      "resourceTitle": "Mechanistic Interpretability Workshop at NeurIPS 2025"
    },
    {
      "text": "80,000 Hours podcast with Chris Olah",
      "url": "https://80000hours.org/podcast/episodes/chris-olah-interpretability-research/",
      "resourceId": "5c66c0b83538d580",
      "resourceTitle": "Chris Olah on Interpretability Research (80,000 Hours Podcast)"
    }
  ],
  "unconvertedLinkCount": 55,
  "convertedLinkCount": 0,
  "backlinkCount": 22,
  "hallucinationRisk": {
    "level": "medium",
    "score": 55,
    "factors": [
      "no-citations"
    ]
  },
  "entityType": "research-area",
  "redundancy": {
    "maxSimilarity": 20,
    "similarPages": [
      {
        "id": "sparse-autoencoders",
        "title": "Sparse Autoencoders (SAEs)",
        "path": "/knowledge-base/responses/sparse-autoencoders/",
        "similarity": 20
      },
      {
        "id": "interpretability",
        "title": "Mechanistic Interpretability",
        "path": "/knowledge-base/responses/interpretability/",
        "similarity": 19
      },
      {
        "id": "probing",
        "title": "Probing / Linear Probes",
        "path": "/knowledge-base/responses/probing/",
        "similarity": 19
      },
      {
        "id": "interpretability-sufficient",
        "title": "Is Interpretability Sufficient for Safety?",
        "path": "/knowledge-base/debates/interpretability-sufficient/",
        "similarity": 18
      },
      {
        "id": "anthropic-core-views",
        "title": "Anthropic Core Views",
        "path": "/knowledge-base/responses/anthropic-core-views/",
        "similarity": 16
      }
    ]
  },
  "coverage": {
    "passing": 8,
    "total": 13,
    "targets": {
      "tables": 14,
      "diagrams": 1,
      "internalLinks": 29,
      "externalLinks": 18,
      "footnotes": 11,
      "references": 11
    },
    "actuals": {
      "tables": 28,
      "diagrams": 1,
      "internalLinks": 17,
      "externalLinks": 62,
      "footnotes": 0,
      "references": 16,
      "quotesWithQuotes": 0,
      "quotesTotal": 0,
      "accuracyChecked": 0,
      "accuracyTotal": 0
    },
    "items": {
      "summary": "green",
      "schedule": "green",
      "entity": "green",
      "editHistory": "red",
      "overview": "green",
      "tables": "green",
      "diagrams": "green",
      "internalLinks": "amber",
      "externalLinks": "green",
      "footnotes": "red",
      "references": "green",
      "quotes": "red",
      "accuracy": "red"
    },
    "ratingsString": "N:4.5 R:6.8 A:5.2 C:7.5"
  },
  "readerRank": 375,
  "researchRank": 299,
  "recommendedScore": 142.16
}

{
  "eaForum": "https://forum.effectivealtruism.org/topics/mechanistic-interpretability",
  "wikipedia": "https://en.wikipedia.org/wiki/Mechanistic_interpretability"
}