{
  "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": 27,
  "pageCount": 313,
  "prevPage": 26,
  "nextPage": 28,
  "slideType": "industry_trends",
  "function": "diagnose",
  "density": "dense",
  "nDataPoints": 2,
  "notes": "Discusses the trade-off between CoT monitoring effectiveness and the emergence of obfuscated reward hacking in AI models.",
  "elementsJson": [
    "headline_text",
    "action_title",
    "bullet_list",
    "area_chart"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/27",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-27",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "Researchers argue that to keep AI systems safe and monitorable, we may need to pay a \"monitorability tax\", i.e. accepting slightly less capable models in exchange for maintaining visibility into their reasoning.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-41ac-7180-a81d-5767aa9f1061",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.45,
        "w": 0.45,
        "x": 0.52,
        "y": 0.32
      },
      "kind": "chart",
      "text": "Catching Systemic Hacks in Frontier Models",
      "attrs": null,
      "subkind": "area-stacked",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "e6394491-895e-4df8-bd27-e422d0ae7409",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.45,
        "w": 0.48,
        "x": 0.04,
        "y": 0.32
      },
      "kind": "list",
      "text": "CoT monitoring is highly effective: GPT-4o caught 95% of reward hacks by reading the reasoning traces of stronger models such as o3-mini, versus only 60% when monitoring actions alone.\nHowever, models trained to avoid triggering the CoT monitor developed \"obfuscated reward hacking\", whereby the model hid its true intentions (with apparently innocuous CoT) while still reward hacking...\nAs such, we have a fine balance at play here: moderate CoT pressure improved performance, but excessive pressure taught models to deceive. As a result, we may need to pay a \"monitorability tax\" to develop less polished but more transparent systems.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "b9da4a3f-85e1-43fd-b919-ae4783ebad3c",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "Fraction of Samples: 95%",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-41ac-7180-a81d-59d0eaa714bb",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.08,
        "w": 0.48,
        "x": 0.02,
        "y": 0.22
      },
      "kind": "paragraph",
      "text": "Researchers argue that to keep AI systems safe and monitorable, we may need to pay a \"monitorability tax\", i.e. accepting slightly less capable models in exchange for maintaining visibility into their reasoning.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "d012f1e8-b0a2-4175-ac33-033e449db356",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.03,
        "w": 0.15,
        "x": 0.82,
        "y": 0.94
      },
      "kind": "source-note",
      "text": "state.ai 2025",
      "attrs": null,
      "subkind": null,
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "21cc0cb3-8e7d-4905-95ba-a768f9c0a0c0",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.38,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "Trading capability for transparency",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "4079f793-2a5a-4904-90d3-a230ad659ffd",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Waterfall chart",
      "slug": "waterfall-chart",
      "agent": null,
      "layer": "slide",
      "matchId": "14c455cf-a1b0-4732-bfc0-f5fa922ba9c7",
      "evidence": "chart/area-stacked: Catching Systemic Hacks in Frontier Models",
      "confidence": 0.7
    }
  ],
  "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
}