{
  "docId": "019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "docSlug": "5df6bd1b0447b5f6",
  "documentTitle": "2025 Air Street Capital The State of AI Report 2025",
  "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": 77,
  "pageCount": 313,
  "prevPage": 76,
  "nextPage": 78,
  "slideType": "case_study",
  "function": "illustrate_case",
  "density": "dense",
  "nDataPoints": 2,
  "notes": "The slide highlights a specific technical research development from Apple, detailing its architecture (4D sparse latent, pure-Transformer, 4D RoPE) and training scale.",
  "elementsJson": [
    "headline_text",
    "paragraph",
    "bullet_list",
    "process_diagram"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/77",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-77",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "Apple introduced a single tokenizer that works for both high-fidelity reconstruction and semantic understanding across images, video, and 3D could be the foundation layer for truly unified multimodal models.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-41ac-7180-a821-421f2abc910c",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.4,
        "w": 0.45,
        "x": 0.52,
        "y": 0.38
      },
      "kind": "image",
      "text": "Visual representation of ATOKEN architecture showing 4D representation, reconstruction, and understanding across image, video, and 3D modalities.",
      "attrs": null,
      "subkind": "infographic",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "cb026735-b203-4291-a39b-70b164490cd0",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.45,
        "x": 0.04,
        "y": 0.73
      },
      "kind": "list",
      "text": "Specialists approaches still lead on some long-video and generative benchmarks, and the compute bill is high. Adoption will hinge on release details and tooling, but the unified-tokenizer direction looks like the right foundation.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "36ee0448-b7cb-485a-8cae-ba3f22a88e7c",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.15,
        "w": 0.45,
        "x": 0.04,
        "y": 0.56
      },
      "kind": "list",
      "text": "One backend now supports both high-fidelity reconstruction and semantic understanding. A curriculum from images → video → 3D shows cross-modal transfer, and native-resolution/time processing with KV caching keeps it scalable (trained up to 256×H100, ~138k GPU-hours).",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "643cc131-c2db-476f-a40d-2a784da150ab",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.18,
        "w": 0.45,
        "x": 0.04,
        "y": 0.36
      },
      "kind": "list",
      "text": "ATOKEN is a single visual tokenizer that maps images, video, and 3D inputs into a shared 4D sparse latent using a pure-Transformer with 4D RoPE. It then emits continuous or discrete tokens, and trains stably with perceptual + Gram losses (no GANs).",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "fcf79629-ffe0-4c5a-86c7-c6d80af8079f",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.95,
        "x": 0.02,
        "y": 0.22
      },
      "kind": "paragraph",
      "text": "Apple introduced a single tokenizer that works for both high-fidelity reconstruction and semantic understanding across images, video, and 3D could be the foundation layer for truly unified multimodal models. This approach reduces fragmentation, simplifying stacks, and enabling direct transfer of capabilities across modalities.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "b2824f39-b681-499c-9433-df6da82ba887",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.6,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "ATOKEN: a unified visual tokenizer for images, video, and 3D",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "8b4b6f6d-3e19-4a11-8397-4d3e6f63fc29",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 189,
      "from": 12,
      "beatId": "019dd95a-0682-776c-8e35-9f380673831f",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Facts (What)",
      "beatSlug": "triple-take-the-facts-what",
      "evidence": "Sections 1-2 lay out research findings and industry data with charts and case studies.",
      "position": 1,
      "confidence": 70,
      "parentBeatName": "Setup",
      "parentBeatSlug": "setup"
    },
    {
      "to": 89,
      "from": 12,
      "beatId": "019dd95a-0682-776c-8e35-ac68557d958b",
      "arcName": "The Onion",
      "arcSlug": "onion",
      "beatName": "First Layer",
      "beatSlug": "onion-first-layer",
      "evidence": "Research section unpacks technical layer beneath the headlines.",
      "position": 2,
      "confidence": 45,
      "parentBeatName": "Development",
      "parentBeatSlug": "development"
    }
  ],
  "loops": [],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}