{
  "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": 37,
  "pageCount": 213,
  "prevPage": 36,
  "nextPage": 38,
  "slideType": "industry_trends",
  "function": "analyze_data",
  "density": "overcrowded",
  "nDataPoints": 48,
  "notes": "The slide includes a table comparing various methods (CLIP, EVA-CLIP, OpenCLIP, etc.) across metrics like FLOPs and performance on ImageNet-1K and COCO datasets.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "data_table"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0856e1444fb9/37",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9#slide-37",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "JEST uses lower-resolution image processing for both data selection and part of the training, significantly reducing computational costs while maintaining performance benefits.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d10-dd354df4a5ad",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.55,
        "w": 0.55,
        "x": 0.03,
        "y": 0.32
      },
      "kind": "list",
      "text": "Usually, an entire dataset is processed upfront, which doesn't account for how the relevance of a training example can change over the course of learning. These methods are frequently applied before training, so cannot adapt to changing needs during training.\nGoogle DeepMind's JEST selects entire batches of data jointly, rather than individual examples independently. The selection is guided by a 'learnability score' (determined by a pre-trained reference model) which evaluates how useful it will be for training. It's able to integrate data selection directly into the training process, making it dynamic and adaptive.\nJEST uses lower-resolution image processing for both data selection and part of the training, significantly reducing computational costs while maintaining performance benefits.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "943bfddd-dc7d-46d5-8465-431e51864c37",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "FLOPs %: 23",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d10-e30f7cca4482",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.1,
        "w": 0.55,
        "x": 0.03,
        "y": 0.21
      },
      "kind": "paragraph",
      "text": "Data curation is an essential part of effective pre-training, but is often done manually and inefficiently. This is both hard to scale and wasteful, especially for multimodal models.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "b787760b-f5ce-4d81-b1bd-7c1b5ca24b0a",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.38,
        "x": 0.58,
        "y": 0.81
      },
      "kind": "source-note",
      "text": "Table 1: Comparison to prior art. FLOP % are measured relative to SigLIP [54]. Mean denotes the average performance over all metrics. \"Per Iter.\" denotes FLOPs per iteration.",
      "attrs": null,
      "subkind": null,
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "036fd85c-41ba-42c0-a3ea-384d862623e3",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.35,
        "w": 0.38,
        "x": 0.58,
        "y": 0.45
      },
      "kind": "table",
      "text": "Comparison table of methods (CLIP, EVA-CLIP, OpenCLIP, LessIsMore, SIL.C-S, SigLIP, JEST++, Flexi-JEST++) showing FLOPs and performance metrics.",
      "attrs": null,
      "subkind": "data",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "63d7f7e7-62c3-4d58-a5e7-b55c3cb1daa7",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.75,
        "x": 0.03,
        "y": 0.13
      },
      "kind": "title",
      "text": "Could better data curation methods reduce training compute requirements?",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "b8f37497-fa20-4678-882d-ce4415c00c43",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Table data",
      "slug": "table-data",
      "agent": null,
      "layer": "slide",
      "matchId": "6ffb53ab-87ad-4b48-bc31-3e47938ffa7c",
      "evidence": "The slide contains a table/data component with a comparison table.",
      "confidence": 0.8
    }
  ],
  "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": 37,
      "from": 30,
      "name": "Pattern Hunter",
      "slug": "02-pattern-hunter",
      "bestFor": "Time-pressed audiences, building consensus, when data is strong",
      "matchId": "019dd95a-07fe-70ce-8d3e-2e9bfb3e3bea",
      "evidence": "Synthetic data, FineWeb, embeddings, RAG, power grid, JEST data curation",
      "position": 4,
      "objective": "Survey data, retrieval and training-infra advances",
      "structure": "Evidence A -> Evidence B -> Evidence C -> Pattern/Conclusion",
      "confidence": 72,
      "description": "Group multiple pieces of evidence that together point to a pattern or conclusion"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}