{
  "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": 31,
  "pageCount": 163,
  "prevPage": 30,
  "nextPage": 32,
  "slideType": "industry_trends",
  "function": "analyze_data",
  "density": "overcrowded",
  "nDataPoints": 15,
  "notes": "Discusses the trade-off between training epochs, model size, and overfitting.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "line_chart",
    "scatter_plot",
    "logo_grid"
  ],
  "metadataConfidence": 0.95,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-06b04d219fea/31",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea#slide-31",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "Training for one or two epochs will generally be optimal; In some cases, pushing for a few more epochs can help; But too many epochs generally equals overfitting.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c4-7719-a222-ddcdbca0e037",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.4,
        "w": 0.3,
        "x": 0.35,
        "y": 0.5
      },
      "kind": "chart",
      "text": "Allocating compute when repeating",
      "attrs": null,
      "subkind": "line",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "16e0be9f-bfe0-4759-ac20-72d1da1d3b01",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.4,
        "w": 0.3,
        "x": 0.68,
        "y": 0.5
      },
      "kind": "chart",
      "text": "Return on compute when repeating",
      "attrs": null,
      "subkind": "line",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "2beeef03-6db0-4baa-b14a-2676df9e42e6",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.35,
        "w": 0.3,
        "x": 0.03,
        "y": 0.58
      },
      "kind": "chart",
      "text": "MLM Accuracy vs Num of Data Repeat",
      "attrs": null,
      "subkind": "line",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "c89d0079-28a0-45af-ac39-47a5fb9be063",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.5,
        "x": 0.4,
        "y": 0.9
      },
      "kind": "legend",
      "text": "Regime of same compute (IsoFLOP); Efficient frontier assuming repeated data is worth the same as new data; Efficient frontier predicted by our data-constrained scaling laws",
      "attrs": null,
      "subkind": null,
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "140f8690-b867-4682-b573-76fed40ce33e",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "MLM Accuracy",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-47c4-7719-a222-e3c620ae5900",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.94,
        "x": 0.03,
        "y": 0.19
      },
      "kind": "paragraph",
      "text": "If we can't have more original training data, why not train more on what we have? Conflicting research indicates that the answer is, as always, it depends: Training for one or two epochs will generally be optimal; In some cases, pushing for a few more epochs can help; But too many epochs generally equals overfitting.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "1e579752-7353-4096-b745-680a9262d6fd",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.94,
        "x": 0.03,
        "y": 0.32
      },
      "kind": "paragraph",
      "text": "Before the large-scale deep learning era (say post GPT-2), most models were trained multiple epochs over a given dataset. But as the size of models grew larger, training for multiple epochs almost always resulted in overfitting, prompting most practitioners to train for a single epoch on the available data (which for once, is the theoretically optimal thing to do).",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "525ee73d-8d18-4a86-97c9-0c4bf97c195c",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.4,
        "x": 0.03,
        "y": 0.13
      },
      "kind": "title",
      "text": "Breaking the data ceiling: overtraining",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "953ad1e8-9042-4cdb-93dc-829a62d0483f",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Waterfall chart",
      "slug": "waterfall-chart",
      "agent": null,
      "layer": "slide",
      "matchId": "9ddc1fa5-86dc-4df4-bbdd-a26a7bce7071",
      "evidence": "MLM Accuracy vs Num of Data Repeat",
      "confidence": 0.8
    }
  ],
  "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
}