{
  "docId": "019dd923-5e88-73ef-bd5d-01cf0d8a8fbc",
  "docSlug": "cc8183ec02431b7a",
  "documentTitle": "2022 Air Street Capital The State of AI Report 2022",
  "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": 21,
  "pageCount": 114,
  "prevPage": 20,
  "nextPage": 22,
  "slideType": "case_study",
  "function": "illustrate_case",
  "density": "dense",
  "nDataPoints": 10,
  "notes": "The slide explains the methodology of training an Inverse Dynamics Model (IDM) to scale up training data for a foundation model.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "process_diagram",
    "line_chart"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-01cf0d8a8fbc/21",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-01cf0d8a8fbc",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-01cf0d8a8fbc.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-01cf0d8a8fbc#slide-21",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "VPT is the first ML model to learn to craft diamonds, \"a task that usually takes proficient humans over 20 minutes (24,000 actions)\".",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c4-7719-a215-5fbdd3fc8b56",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.2,
        "w": 0.46,
        "x": 0.51,
        "y": 0.75
      },
      "kind": "chart",
      "text": "Reward over episodes",
      "attrs": null,
      "subkind": "line",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "0a718ce1-edc3-4bd3-8c1a-933b429f4cad",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.32,
        "w": 0.46,
        "x": 0.51,
        "y": 0.38
      },
      "kind": "diagram",
      "text": "Collect Internet data -> Train the Inverse Dynamics Model (IDM) -> Train the VPT Foundation Model",
      "attrs": null,
      "subkind": "process",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "001c0059-ce85-47a2-926b-48bee2ed43d5",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.45,
        "w": 0.45,
        "x": 0.03,
        "y": 0.38
      },
      "kind": "list",
      "text": "OpenAI gathered 2,000 hours of video labeled with mouse and keyboard actions and trained an inverse dynamics model (IDM) to predict actions given past and future frames - this is the PreTraining part.\nThey then used the IDM to label 70K hours of video on which they trained a model to predict actions given only past video frames.\nThey show that the model can be fine-tuned with imitation learning and reinforcement learning (RL) to achieve a performance which is too hard to reach using RL from scratch.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "50112f62-8174-4acd-866b-22851a691a2a",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "Reward",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-47c4-7719-a215-60ba245ad00f",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.94,
        "x": 0.03,
        "y": 0.22
      },
      "kind": "paragraph",
      "text": "OpenAI trained a model (Video PreTraining, VPT) to play Minecraft from video frames using a small amount of labeled mouse and keyboard interactions. VPT is the first ML model to learn to craft diamonds, “a task that usually takes proficient humans over 20 minutes (24,000 actions)”.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "32bec6a1-4056-4273-ba8c-2273971fc3d1",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.6,
        "x": 0.03,
        "y": 0.14
      },
      "kind": "title",
      "text": "OpenAI uses Minecraft as a testbed for computer-using agents",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "f38a4039-c5ca-41ad-b98d-ea5bd334bdb5",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "AIDA Model",
      "slug": "aida-model",
      "agent": "Storyteller",
      "layer": "slide",
      "matchId": "e0ef2f02-22cf-4630-80bc-4107030dcb80",
      "evidence": "The slide presents a case study with a clear problem (OpenAI uses Minecraft as a testbed), interest (gathered 2,000 hours of video labeled with mouse and keyboard actions), desire (trained an inverse dynamics model), and action (VPT is the first ML model to learn to craft diamonds).",
      "confidence": 0.7
    }
  ],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 80,
      "from": 5,
      "beatId": "019dd95a-0682-776c-8e35-2557e3799e96",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Facts (What)",
      "beatSlug": "triple-take-the-facts-what",
      "evidence": "Definitions, exec summary, Research and Industry sections inventory the state of AI.",
      "position": 1,
      "confidence": 78,
      "parentBeatName": "Setup",
      "parentBeatSlug": "setup"
    },
    {
      "to": 93,
      "from": 11,
      "beatId": "019dd95a-0682-776c-8e35-35abf646aa8b",
      "arcName": "The Mountain",
      "arcSlug": "mountain",
      "beatName": "Rising Action",
      "beatSlug": "mountain-rising-action",
      "evidence": "Research, Industry, Politics sections accumulate signals.",
      "position": 2,
      "confidence": 55,
      "parentBeatName": "Development",
      "parentBeatSlug": "development"
    }
  ],
  "loops": [
    {
      "to": 21,
      "from": 12,
      "name": "Pattern Hunter",
      "slug": "02-pattern-hunter",
      "bestFor": "Time-pressed audiences, building consensus, when data is strong",
      "matchId": "019dd95a-07fe-70ce-8d3d-9f07f56c535f",
      "evidence": "DeepMind math, fusion, AlphaFold, protein LMs, plastic enzymes, Minecraft agents — all evidence of one pattern.",
      "position": 2,
      "objective": "Compile case studies showing ML breakthroughs in physical/life sciences",
      "structure": "Evidence A -> Evidence B -> Evidence C -> Pattern/Conclusion",
      "confidence": 80,
      "description": "Group multiple pieces of evidence that together point to a pattern or conclusion"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}