{
  "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": 50,
  "pageCount": 313,
  "prevPage": 49,
  "nextPage": 51,
  "slideType": "case_study",
  "function": "illustrate_case",
  "density": "dense",
  "nDataPoints": 36,
  "notes": "The slide highlights the 'shortcut forcing' technique and compares Dreamer 4 against VPT, BC, and VLA models across various Minecraft tasks.",
  "elementsJson": [
    "headline_text",
    "paragraph",
    "bullet_list",
    "bar_chart_grouped"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/50",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-50",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "The agent is the first to reach diamonds in Minecraft using offline data only, outperforming OpenAI's VPT while using roughly 100x less labeled data.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-41ac-7180-a81e-44eeee588c2e",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.35,
        "w": 0.45,
        "x": 0.52,
        "y": 0.45
      },
      "kind": "chart",
      "text": "Offline Diamond Challenge success rates comparing VPT, BC, VLA, and Dreamer 4 across 9 Minecraft items.",
      "attrs": null,
      "subkind": "bar-grouped",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "067e331d-1410-4b88-9342-760af41b0242",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.45,
        "w": 0.46,
        "x": 0.04,
        "y": 0.41
      },
      "kind": "list",
      "text": "The system first learns dynamics and objects from large unlabeled video, then adds a small action-labeled set to ground control and inventory changes.\nThe policy is improved by rolling out many imagined trajectories inside the learned model. Rewards and value heads are trained on the same data to guide long-horizon skills.\nShortcut forcing contrasts predictions with and without the true actions, pushing the world model to depend on actions rather than hindsight correlations.\nThe model runs at interactive frame rates on a single GPU and supports live human play in the learned world, though memory is short and inventory tracking is still imperfect.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "40c543b2-b228-42e9-bee8-819f275b574a",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "Success rate (%): 100x",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-41ac-7180-a81e-49f7c9c08570",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.18,
        "w": 0.48,
        "x": 0.02,
        "y": 0.21
      },
      "kind": "paragraph",
      "text": "Dreamer 4 trains a video world model that can predict object interactions and future frames, then learns its policy entirely in imagination. A new “shortcut forcing” objective and an efficient transformer make the model run at real-time speeds on a single GPU. The agent is the first to reach diamonds in Minecraft using offline data only, outperforming OpenAI’s VPT while using roughly 100x less labeled data.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "75ff6504-af4b-4fad-a4dc-ce8daa0ec572",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.45,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "Training agents inside of scalable world models",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "eef2223c-bd41-4c89-a2b6-449a8bdd27d7",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [],
  "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": [],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}