{
  "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": 15,
  "pageCount": 313,
  "prevPage": 14,
  "nextPage": 16,
  "slideType": "case_study",
  "function": "illustrate_case",
  "density": "dense",
  "nDataPoints": 12,
  "notes": "Includes a bar chart comparing model performance and a line chart showing AIME accuracy improvement during training.",
  "elementsJson": [
    "headline_text",
    "paragraph",
    "bullet_list",
    "bar_chart_grouped",
    "line_chart"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/15",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-15",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "During training the model lengthens its thoughts, explores, and reallocates compute to hard problems.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-41ac-7180-a81a-4f23bce31455",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.2,
        "w": 0.25,
        "x": 0.735,
        "y": 0.375
      },
      "kind": "chart",
      "text": "Model performance comparison across benchmarks",
      "attrs": null,
      "subkind": "bar-grouped",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "8c6e5267-5edc-434f-9449-1dbe751dcb67",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.2,
        "w": 0.25,
        "x": 0.735,
        "y": 0.64
      },
      "kind": "chart",
      "text": "DeepSeek-R1-Zero AIME accuracy during training",
      "attrs": null,
      "subkind": "line",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "d42d32b7-ce42-470f-94d9-fbc5e9c5bab8",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.45,
        "w": 0.65,
        "x": 0.03,
        "y": 0.4
      },
      "kind": "list",
      "text": "R1-Zero follows a “think → answer” format and uses a simple rule-based reward for getting the final answer right.\nGRPO compares multiple sampled answers within a group to form a relative baseline.\nDuring training the model lengthens its thoughts.\nR1 then repairs readability with a small CoT warm start, a language-consistency reward, large supervised finetuning, and a final RL pass.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "b842428a-bbef-4494-90f5-784c5a12c2b4",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "AIME accuracy: 79.8",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-41ac-7180-a81a-514291385d37",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.15,
        "w": 0.65,
        "x": 0.03,
        "y": 0.22
      },
      "kind": "paragraph",
      "text": "A few days after Christmas 2024, DeepSeek unveiled V3, a strong 671B MoE V3 that lowered training and inference cost with FP8 mixed precision, multi-token prediction, and auxiliary-free routing. Using V3 as the base, they trained R1-Zero only with RL.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "42408eb0-6131-4255-a184-241440af9e69",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.5,
        "x": 0.03,
        "y": 0.14
      },
      "kind": "title",
      "text": "Are you not entertained? DeepSeek V3 brings you to R1",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "be6391f1-b524-4a31-ad0e-1c3cc25c33ef",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Problem-Agitate-Solve (PAS)",
      "slug": "problem-agitate-solve-pas",
      "agent": "Storyteller",
      "layer": "slide",
      "matchId": "c4f316fb-3fbd-44c7-8124-88e28c19f057",
      "evidence": "A few days after Christmas 2024, DeepSeek unveiled V3, a strong 671B MoE V3 that lowered training and inference cost with FP8 mixed precision, multi-token prediction, and auxiliary-free routing.",
      "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
}