{
  "docId": "019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "docSlug": "5df6bd1b0447b5f6",
  "documentTitle": "2025 Air Street Capital The State of AI Report 2025",
  "authorId": "AirStreetCapital",
  "authorName": "Air Street Capital",
  "documentKindSlug": "consulting-deck",
  "documentKindLabel": "Consulting deck",
  "sourceTypeSlug": "vc_research",
  "sourceTypeLabel": "VC research",
  "presentationDate": null,
  "orientation": "landscape",
  "aspectRatio": 1.777,
  "pageNumber": 272,
  "pageCount": 313,
  "prevPage": 271,
  "nextPage": 273,
  "slideType": "case_study",
  "function": "illustrate_case",
  "density": "dense",
  "nDataPoints": 1,
  "notes": "The slide uses a specific example of a multi-step reasoning graph to demonstrate model interpretability.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "process_diagram"
  ],
  "metadataConfidence": 0.95,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/272",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-272",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "When applied to Claude 3.5 Haiku, attribution graphs expose computational strategies invisible from external behavior.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d0b-0866d03c544d",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.3,
        "w": 0.3,
        "x": 0.65,
        "y": 0.033
      },
      "kind": "image",
      "text": "Attribution graph visualization showing the reasoning path for the Dallas/Texas/Austin query.",
      "attrs": null,
      "subkind": "infographic",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "97ae0dd0-bc97-47a1-a711-fabd68d74f05",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.5,
        "w": 0.6,
        "x": 0.05,
        "y": 0.33
      },
      "kind": "list",
      "text": "Models perform genuine multi-step reasoning internally. When asked \"what's the capital of the state containing Dallas\", Claude Haiku 3.5 executes \"Dallas → Texas → Austin\" as distinct steps.\nMedical diagnosis also mirrors clinical thinking. Given symptoms suggesting preeclampsia, the model internally activates \"preeclampsia\" features without any mention in the prompt, then searches for confirmatory symptoms.\nBut jailbreaks exploit this mechanical processing: the model decodes \"Babies Outlive Mustard Block\" into \"BOMB\" letter-by-letter without recognizing the danger until after output. Attribution graphs revealed why: safety circuits don't activate during obfuscated decoding, but only after seeing its own harmful output.\nThis method works for only ~25% of prompts: it cannot explain how attention decides where to look, and requires manual interpretation through \"supernodes\" to be readable.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "306884d0-5cbc-4c41-8bb3-6855ba464232",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.1,
        "w": 0.9,
        "x": 0.05,
        "y": 0.22
      },
      "kind": "paragraph",
      "text": "When applied to Claude 3.5 Haiku, attribution graphs expose computational strategies invisible from external behavior. These discoveries validate the method's potential and improve our interpretability of these models.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "699e6177-1d58-412a-9c24-dce215a1da17",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.6,
        "x": 0.05,
        "y": 0.14
      },
      "kind": "title",
      "text": "Early applications of attribution graphs reveal internal mechanisms",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "dc7d1797-e95e-42f8-b641-747d55182f5b",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Narrative foundations (from the Storymakers methodology)",
      "slug": "narrative-foundations",
      "agent": "storyteller",
      "layer": "slide",
      "matchId": "0df653e9-7746-49f6-be0a-3946a20ca8c4",
      "evidence": "title/headline: Early applications of attribution graphs reveal internal mechanisms",
      "confidence": 0.6
    }
  ],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 282,
      "from": 190,
      "beatId": "019dd95a-0682-776c-8e35-a256fe16ea9c",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Implications (So What)",
      "beatSlug": "triple-take-the-implications-so-what",
      "evidence": "Sections 3-4 frame political and safety consequences of the technical/industry trends.",
      "position": 2,
      "confidence": 70,
      "parentBeatName": "Reflection",
      "parentBeatSlug": "reflection"
    },
    {
      "to": 282,
      "from": 190,
      "beatId": "019dd95a-0682-776c-8e35-b713af60bf88",
      "arcName": "The Onion",
      "arcSlug": "onion",
      "beatName": "Core Insight",
      "beatSlug": "onion-core-insight",
      "evidence": "Politics and Safety reveal the structural risks at the core.",
      "position": 4,
      "confidence": 45,
      "parentBeatName": "Turn",
      "parentBeatSlug": "turn"
    }
  ],
  "loops": [
    {
      "to": 282,
      "from": 247,
      "name": "Pre Mortem",
      "slug": "19-pre-mortem",
      "bestFor": "Project kick-offs, risk management, building trust with skeptical stakeholders",
      "matchId": "019dd95a-07fe-70ce-8d3e-bb4d43115ab4",
      "evidence": "Vibe hacking, AI psychosis, alignment faking, scheming, prompt injection — what could go wrong + mitigations.",
      "position": 12,
      "objective": "Surface AI safety failure modes before they become catastrophic",
      "structure": "The Future Disaster (Hypothetical) -> What Went Wrong -> The Prevention Plan",
      "confidence": 75,
      "description": "Build trust by anticipating failure modes before they happen and solving them in advance"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}