{
  "docId": "019dd923-5e88-73ef-bd5d-0856e1444fb9",
  "docSlug": "8757f1b44ef7f176",
  "documentTitle": "2024 Air Street Capital The State of AI Report 2024",
  "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": 65,
  "pageCount": 213,
  "prevPage": 64,
  "nextPage": 66,
  "slideType": "case_study",
  "function": "illustrate_case",
  "density": "dense",
  "nDataPoints": 1,
  "notes": "The slide details a two-stage RL process (offline then online) for VLM training.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "process_diagram",
    "data_table",
    "paragraph"
  ],
  "metadataConfidence": 0.95,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0856e1444fb9/65",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9#slide-65",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "It achieves a 62.7% task success rate on the Android-in-the-Wild dataset, a significant improvement on the prior SOTA.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d14-2024e41d7436",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.22,
        "x": 0.4,
        "y": 0.58
      },
      "kind": "chart",
      "text": "Step-level Value Function comparison",
      "attrs": null,
      "subkind": "bar-vertical",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "5d508b7a-4f0c-4256-a605-b3d4cbf69d45",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.33,
        "x": 0.645,
        "y": 0.445
      },
      "kind": "diagram",
      "text": "DigiRL overview process flow: Pretraining -> Step I: Offline RL -> Step II: Online RL",
      "attrs": null,
      "subkind": "process",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "ab62bc89-b696-4d11-bbae-ce0aedc87c6c",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.35,
        "w": 0.58,
        "x": 0.02,
        "y": 0.22
      },
      "kind": "list",
      "text": "For agents to be useful, they need to be robust to real-word stochasticity, which SOTA models have historically struggled with. We're beginning to see signs of progress.\nDigiRL is a novel autonomous reinforcement learning approach for training in-the-wild device control agents specifically for Android devices. The method involves a two-stage process: offline reinforcement learning followed by offline-to-online reinforcement learning.\nIt achieves a 62.7% task success rate on the Android-in-the-Wild dataset, a significant improvement on the prior SOTA.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "5e6767fa-1370-44cf-9bbe-bda1a3f968bc",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "task success rate: 62.7%",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d14-2544d5242a20",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.4,
        "x": 0.02,
        "y": 0.58
      },
      "kind": "table",
      "text": "Instruction-level Value Function table showing tasks, difficulty, and values.",
      "attrs": null,
      "subkind": "data",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "330a8b1d-d1e5-4f2e-a42f-dcfdf640e91a",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.45,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "RL drives improvements in VLM performance...",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "ca1dc5ad-45fb-4b8c-b153-ec8cc9666bbb",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Action Titles",
      "slug": "action-titles",
      "agent": "Architect",
      "layer": "slide",
      "matchId": "019dd95a-1055-74e0-a959-1f81984a2dd2",
      "evidence": "Title 'RL drives improvements in VLM performance...'",
      "confidence": 85
    }
  ],
  "frameworks": [
    {
      "name": "process",
      "slug": null,
      "matchId": "cd8fbb05-9fd9-4906-9251-7e7a07a15b8b",
      "evidence": "The slide describes a multi-stage training pipeline for AI agents.",
      "confidence": 0.8
    }
  ],
  "arcBeats": [
    {
      "to": 153,
      "from": 9,
      "beatId": "019dd95a-0682-776c-8e35-5f2398b8d1d0",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Facts (What)",
      "beatSlug": "triple-take-the-facts-what",
      "evidence": "Exec summary + Research + Industry sections inventory the year",
      "position": 1,
      "confidence": 78,
      "parentBeatName": "Setup",
      "parentBeatSlug": "setup"
    },
    {
      "to": 85,
      "from": 12,
      "beatId": "019dd95a-0682-776c-8e35-6d9d8ee06d0a",
      "arcName": "Voyage and Return",
      "arcSlug": "voyage-return",
      "beatName": "The Unknown",
      "beatSlug": "voyage-return-the-unknown",
      "evidence": "Research section explores frontier model uncharted territory",
      "position": 2,
      "confidence": 55,
      "parentBeatName": "Complication",
      "parentBeatSlug": "complication"
    }
  ],
  "loops": [
    {
      "to": 71,
      "from": 59,
      "name": "Pattern Hunter",
      "slug": "02-pattern-hunter",
      "bestFor": "Time-pressed audiences, building consensus, when data is strong",
      "matchId": "019dd95a-07fe-70ce-8d3e-3f24b9b4c1e3",
      "evidence": "ARC challenge, planning failures, COT, open-endedness, MCTS, FunSearch, RL-LLM combos",
      "position": 8,
      "objective": "Catalogue progress and limits in reasoning research",
      "structure": "Evidence A -> Evidence B -> Evidence C -> Pattern/Conclusion",
      "confidence": 76,
      "description": "Group multiple pieces of evidence that together point to a pattern or conclusion"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}