{
  "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": 23,
  "pageCount": 213,
  "prevPage": 22,
  "nextPage": 24,
  "slideType": "industry_trends",
  "function": "summarize",
  "density": "dense",
  "nDataPoints": 0,
  "notes": "The slide uses an AI-generated image of two llamas wearing sunglasses labeled '8B' and '4B' to metaphorically represent model distillation.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "photo"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0856e1444fb9/23",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9#slide-23",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "As Andrej Karpathy and others have argued, current large model sizes could be a reflection of inefficient training. Using these big models to refine and synthesize training data, could help train capable smaller models.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d0f-5c9ca5862855",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.35,
        "w": 0.36,
        "x": 0.61,
        "y": 0.51
      },
      "kind": "image",
      "text": "AI-generated image of two llamas wearing sunglasses labeled 8B and 4B",
      "attrs": null,
      "subkind": "illustration",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "5ee620dc-704b-405a-932b-98f0edb7c504",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.55,
        "w": 0.58,
        "x": 0.02,
        "y": 0.31
      },
      "kind": "list",
      "text": "Google have embraced this approach, distilling Gemini 1.5 Flash from Gemini 1.5 Pro, while Gemma 2 9B was distilled from Gemma 2 27B, and Gemma 2B from a larger unreleased model.\nThere is also community speculation that Claude 3 Haiku, a highly capable smaller model, is a distilled version of the larger Opus, but Anthropic has never confirmed this.\nThese distillation efforts are going multimodal too. Black Forest Labs have released FLUX.1 dev, an open-weight text-to-image distilled from their Pro model.\nTo support these efforts, the community has started to produce open-source distillation tools, like arcee.ai’s DistillKit, which supports both Logit-based and Hidden States-based distillation.\nLlama 3.1 405B is also being used for distillation, after Meta updated its terms so output logits can be used to improve any models, not just Llama ones.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "b3be8132-bd1b-4d0d-a2b1-384dd4f05ed7",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.09,
        "w": 0.58,
        "x": 0.02,
        "y": 0.21
      },
      "kind": "paragraph",
      "text": "As Andrej Karpathy and others have argued, current large model sizes could be a reflection of inefficient training. Using these big models to refine and synthesize training data, could help train capable smaller models.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "9e7f6068-0b4b-40e1-b004-35f0b5efde75",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.45,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "...as distilled models become more fashionable",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "4a4eb73e-29e8-4be6-bd7f-f8a5047c3d60",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Action Titles",
      "slug": "action-titles",
      "agent": "Architect",
      "layer": "slide",
      "matchId": "019dd95a-1055-74e0-a958-93a79cc82b79",
      "evidence": "Title '...as distilled models become more fashionable' continues the prior",
      "confidence": 85
    },
    {
      "name": "Big Idea Formula",
      "slug": "big-idea-formula",
      "agent": "Storyteller",
      "layer": "slide",
      "matchId": "2c50e086-b5c3-47b8-abb9-f56a03dde001",
      "evidence": "As Andrej Karpathy and others have argued, current large model sizes could be a reflection of inefficient training.",
      "confidence": 0.5
    }
  ],
  "frameworks": [],
  "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": 29,
      "from": 21,
      "name": "Pattern Hunter",
      "slug": "02-pattern-hunter",
      "bestFor": "Time-pressed audiences, building consensus, when data is strong",
      "matchId": "019dd95a-07fe-70ce-8d3e-2b2229583eee",
      "evidence": "Eight slides on shrinking models, on-device, hybrids, transformers reigning",
      "position": 3,
      "objective": "Demonstrate efficiency frontier across distillation, quantization, hybrids",
      "structure": "Evidence A -> Evidence B -> Evidence C -> Pattern/Conclusion",
      "confidence": 75,
      "description": "Group multiple pieces of evidence that together point to a pattern or conclusion"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}