{
  "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": 28,
  "pageCount": 313,
  "prevPage": 27,
  "nextPage": 29,
  "slideType": "key_messages",
  "function": "summarize",
  "density": "overcrowded",
  "nDataPoints": 1,
  "notes": "Includes a diagram illustrating the difference between chain of thought and opaque computation.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "process_diagram"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/28",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-28",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "Anthropic finds <20% of true cues are verbalized, with faithfulness dropping on harder tasks.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-41ac-7180-a81b-79d0dfac8122",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.3,
        "w": 0.25,
        "x": 0.72,
        "y": 0.62
      },
      "kind": "chart",
      "text": "Diagram showing chain of thought vs opaque computation",
      "attrs": null,
      "subkind": "line",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "06ca94b8-e3c2-4372-a885-2177f6483638",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.95,
        "x": 0.02,
        "y": 0.45
      },
      "kind": "list",
      "text": "But CoTs are incomplete and can drift away from faithful reasoning. Anthropic finds <20% of true cues are verbalized. Using RL boosts scores, not legibility.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "08c6b851-75d4-4bcf-9f71-a8b8a00c091d",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.95,
        "x": 0.02,
        "y": 0.73
      },
      "kind": "list",
      "text": "If algorithmic advances tie interpretable CoT traces to performance penalties, this tension will intensify. Without industry standards, we have a major debate.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "182d0b72-e568-4c00-b579-f11ab3def6e4",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.08,
        "w": 0.95,
        "x": 0.02,
        "y": 0.21
      },
      "kind": "list",
      "text": "Leaders from OpenAI, Google DeepMind, Anthropic, and Meta coordinated a joint call to action, urging the field to seize the opportunity to understand and preserve CoT monitorability.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "529aac4e-8b62-47dd-8510-a5b13921ddaf",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.95,
        "x": 0.02,
        "y": 0.59
      },
      "kind": "list",
      "text": "Scaling outcome-based RL can reduce legibility; direct process supervision can distort faithfulness and latent-reasoning architectures may bypass language entirely.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "5580d684-ca7e-4189-a3cc-1815aaccd493",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.95,
        "x": 0.02,
        "y": 0.31
      },
      "kind": "list",
      "text": "CoT monitors can surface (i) explicit malicious intent in reasoning, (ii) early misalignment signals, and (iii) evaluation artifacts.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "83094440-bf59-4d94-9fbe-18fc99852693",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.85,
        "x": 0.02,
        "y": 0.13
      },
      "kind": "title",
      "text": "Monitor your CoTs: just because you can read 'em, doesn't mean you should trust 'em",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "4dfd4885-59bf-4a1d-8558-707b192c71af",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "List presentation",
      "slug": "list-presentation",
      "agent": null,
      "layer": "slide",
      "matchId": "ffa91188-24bd-4b72-ba8c-48258fd35d50",
      "evidence": "list/bullet: Leaders from OpenAI, Google DeepMind, Anthropic, and Meta coordinated a joint call to action, urging the field to seize the opportunity to understand and preserve CoT monitorability.",
      "confidence": 0.6
    }
  ],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 189,
      "from": 12,
      "beatId": "019dd95a-0682-776c-8e35-9f380673831f",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Facts (What)",
      "beatSlug": "triple-take-the-facts-what",
      "evidence": "Sections 1-2 lay out research findings and industry data with charts and case studies.",
      "position": 1,
      "confidence": 70,
      "parentBeatName": "Setup",
      "parentBeatSlug": "setup"
    },
    {
      "to": 89,
      "from": 12,
      "beatId": "019dd95a-0682-776c-8e35-ac68557d958b",
      "arcName": "The Onion",
      "arcSlug": "onion",
      "beatName": "First Layer",
      "beatSlug": "onion-first-layer",
      "evidence": "Research section unpacks technical layer beneath the headlines.",
      "position": 2,
      "confidence": 45,
      "parentBeatName": "Development",
      "parentBeatSlug": "development"
    }
  ],
  "loops": [
    {
      "to": 30,
      "from": 13,
      "name": "Zoom In",
      "slug": "06-zoom-in",
      "bestFor": "Technical deep-dives, case studies, detailed analysis",
      "matchId": "019dd95a-07fe-70ce-8d3e-9252f6f92bd7",
      "evidence": "Big picture 'reasoning race' (p.13) zooms into o1, R1, DeepSeek, parallel reasoning, CoT mechanics.",
      "position": 2,
      "objective": "Drill from reasoning paradigm into specific reasoning research findings",
      "structure": "The Big Picture -> Key Area of Focus -> Specific Detail -> Implication",
      "confidence": 75,
      "description": "Start broad, then progressively focus on specific details that prove your point"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}