{
  "docId": "019dd923-5e88-73ef-bd5d-06b04d219fea",
  "docSlug": "dd91c78f6570bf29",
  "documentTitle": "2023 Air Street Capital The State of AI Report 2023",
  "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": 29,
  "pageCount": 163,
  "prevPage": 28,
  "nextPage": 30,
  "slideType": "industry_trends",
  "function": "analyze_data",
  "density": "dense",
  "nDataPoints": 10,
  "notes": "The chart shows Top-1 Accuracy vs Parameters (M) for models trained on 'Real' vs 'Real + Generated' data, indicating performance gains.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "paragraph",
    "scatter_plot"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-06b04d219fea/29",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea#slide-29",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "Synthetic data is becoming more helpful, but there is still evidence showing that in some cases generated data makes models forget.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c4-7719-a220-cb8eeafbcecb",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.4,
        "w": 0.3,
        "x": 0.68,
        "y": 0.48
      },
      "kind": "chart",
      "text": "Top-1 Accuracy (%) vs Parameters (M) comparing Real vs Real + Generated data",
      "attrs": null,
      "subkind": "dot",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "d4f92d38-9448-4594-9d51-388ac4a2ddf0",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.55,
        "w": 0.65,
        "x": 0.03,
        "y": 0.35
      },
      "kind": "list",
      "text": "Despite the seemingly infinitely proprietary and publicly available data, the largest models are actually running out of data to train on, and testing the limits of scaling laws. One way to alleviate this problem (which has been extensively explored in the past) is to train on AI-generated data, whose volume is only bounded by compute.\nResearchers from Google fine-tune the Imagen text-to-image model for class-conditional ImageNet, then generated one to 12 synthetic versions of ImageNet on which they trained their models (in addition to the original ImageNet). They showed that increasing the size of the synthetic dataset monotonically improved the model's accuracy.\nOther researchers showed that the compounding errors from training on synthetic text online may result in model collapse, “where generated data end up polluting the training set of the next generation of models”. The way forward might be carefully-controlled data-augmentation (so as usual).",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "ce89effb-ddeb-43ec-bacc-bbe89d3168ed",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "Top-1 Accuracy (%)",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-47c4-7719-a220-cd3a6aa7e0bc",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.65,
        "x": 0.03,
        "y": 0.21
      },
      "kind": "paragraph",
      "text": "Another perspective that improving generating models open is expanding the pool of available training data via AI-generated content. We're nowhere near a definitive answer: Synthetic data is becoming more helpful, but there is still evidence showing that in some cases generated data makes models forget.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "7a401d3e-ca46-4d63-85f3-dd7c2d38eafe",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.45,
        "x": 0.03,
        "y": 0.14
      },
      "kind": "title",
      "text": "Breaking the data ceiling: AI-generated content",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "f8a01861-a1a5-4bc6-b455-c62906166b72",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Before-After-Bridge",
      "slug": "before-after-bridge",
      "agent": "Storyteller",
      "layer": "slide",
      "matchId": "ed90153f-3985-4adb-8957-ea958e1e9c3a",
      "evidence": "Another perspective that improving generating models open is expanding the pool of available training data via AI-generated content.",
      "confidence": 0.6
    }
  ],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 120,
      "from": 11,
      "beatId": "019dd95a-0682-776c-8e35-41afd44ef59f",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Facts (What)",
      "beatSlug": "triple-take-the-facts-what",
      "evidence": "Research + Industry sections inventory model, compute, funding facts.",
      "position": 1,
      "confidence": 78,
      "parentBeatName": "Setup",
      "parentBeatSlug": "setup"
    },
    {
      "to": 120,
      "from": 11,
      "beatId": "019dd95a-0682-776c-8e35-523bfb7f96e6",
      "arcName": "The Mountain",
      "arcSlug": "mountain",
      "beatName": "Rising Action",
      "beatSlug": "mountain-rising-action",
      "evidence": "Escalating capabilities, compute concentration and capital flows.",
      "position": 2,
      "confidence": 45,
      "parentBeatName": "Development",
      "parentBeatSlug": "development"
    }
  ],
  "loops": [
    {
      "to": 31,
      "from": 22,
      "name": "Zoom In",
      "slug": "06-zoom-in",
      "bestFor": "Technical deep-dives, case studies, detailed analysis",
      "matchId": "019dd95a-07fe-70ce-8d3d-ea04ad01cb62",
      "evidence": "Slides drill from emergent-capability mirage into context-length, memory and data-exhaustion details.",
      "position": 3,
      "objective": "Zoom into the data and context-length bottleneck",
      "structure": "The Big Picture -> Key Area of Focus -> Specific Detail -> Implication",
      "confidence": 72,
      "description": "Start broad, then progressively focus on specific details that prove your point"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}