{
  "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": 69,
  "pageCount": 313,
  "prevPage": 68,
  "nextPage": 70,
  "slideType": "industry_trends",
  "function": "present_solution",
  "density": "overcrowded",
  "nDataPoints": 4,
  "notes": "The slide explains a technical architecture (BLT) and compares its performance (BPB vs Training FLOPS) against standard BPE-based models.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "process_diagram",
    "line_chart"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/69",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-69",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "At 8B, BLT matches tokenized LLM quality while opening a new scaling axis and cutting inference FLOPs for the same quality.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-41ac-7180-a820-06f92ec1a756",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.35,
        "w": 0.32,
        "x": 0.66,
        "y": 0.35
      },
      "kind": "chart",
      "text": "BPB vs Training Bytes at Fixed Inference FLOPs",
      "attrs": null,
      "subkind": "line",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "f7ddafb3-e991-42fb-962f-72f1d0735990",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.35,
        "w": 0.28,
        "x": 0.37,
        "y": 0.35
      },
      "kind": "diagram",
      "text": "BLT architecture flow: Byte-Level Small Transformer -> Entropy-Based Grouping -> Large Latent Transformer -> Unpatching -> Small Byte-Level Transformer",
      "attrs": null,
      "subkind": "process",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "f446db43-f832-49ac-b782-d88dbea8c1e3",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.1,
        "w": 0.94,
        "x": 0.04,
        "y": 0.72
      },
      "kind": "list",
      "text": "In controlled scaling studies at the 8B class, BLT reaches quality comparable to tokenized LLMs. It also reduces inference compute at the same quality by growing patch size with model size rather than paying per token.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "4068a7e6-1ee5-4e84-a1ce-39ed7baac8a7",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.94,
        "x": 0.04,
        "y": 0.85
      },
      "kind": "list",
      "text": "Byte-level training improves robustness to spelling variation, noise, and long-tail inputs.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "55443275-170a-443b-93a0-5ebd7e9b56b8",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.32,
        "x": 0.04,
        "y": 0.35
      },
      "kind": "list",
      "text": "The model reads raw bytes, groups them into patches where next-byte entropy is high, encodes each patch locally, and then lets a Transformer operate over the patch sequence before decoding back to text.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "d92a70ec-14aa-458c-b173-8215719c13c6",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "Bits-per-byte (BPB)",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-41ac-7180-a820-0bcb2913a574",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.08,
        "w": 0.96,
        "x": 0.02,
        "y": 0.2
      },
      "kind": "paragraph",
      "text": "The Byte Latent Transformer (BLT) learns directly from bytes and uses entropy-driven “patches” as the compute unit. At 8B, BLT matches tokenized LLM quality while opening a new scaling axis and cutting inference FLOPs for the same quality.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "7e9abcaf-258f-4418-8c40-f06e67d7a700",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.45,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "Tokenizer-free LLMs via dynamic byte patches",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "e0529dc3-6090-4149-9618-91139bfd7149",
      "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
}