{
  "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": 79,
  "pageCount": 313,
  "prevPage": 78,
  "nextPage": 80,
  "slideType": "industry_trends",
  "function": "present_framework",
  "density": "dense",
  "nDataPoints": 1,
  "notes": "Discusses the trade-off between overfitting/compute costs and the ability to internalize complex physical dynamics.",
  "elementsJson": [
    "headline_text",
    "paragraph",
    "bullet_list",
    "process_diagram"
  ],
  "metadataConfidence": 0.95,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/79",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-79",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "If robotics data approached VLM scale, end-to-end would likely dominate.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-41ac-7180-a823-7c31dce84a52",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.5,
        "w": 0.35,
        "x": 0.62,
        "y": 0.34
      },
      "kind": "image",
      "text": "Diagrams of Pi-0.5 architecture and SmoLVLA/LeRobot components.",
      "attrs": null,
      "subkind": "infographic",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "f37ef6a9-1599-44d5-af70-a619c0d80bad",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.55,
        "x": 0.03,
        "y": 0.64
      },
      "kind": "list",
      "text": "The case for end-to-end: In contrast, models like ByteDance GR-3 and SmoLVLA show the upside of unfreezing when you have enough task data: the network can internalize contact, dynamics, and scene geometry. If robotics data approached VLM scale, end-to-end would likely dominate.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "6f3db6a8-7389-48ba-b0e7-6052e71c9c38",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.55,
        "x": 0.03,
        "y": 0.38
      },
      "kind": "list",
      "text": "The case for insulation: Pi-0.5 freezes the large VLM and fine-tunes only small “action-expert” heads. This works because robot datasets are tiny, often ≤0.1% the size of the VLM pre-training corpora, so full-network tuning tends to overfit and forget general knowledge while costing more compute.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "92c080ff-2bcb-40f2-af64-4d9e08f2221b",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.15,
        "w": 0.55,
        "x": 0.03,
        "y": 0.22
      },
      "kind": "paragraph",
      "text": "With powerful, pre-trained Vision-Language-Action Models (VLAMs) serving as the \"brains\" of robotic agents, a critical architectural debate has emerged: should the entire model be fine-tuned for a new physical task, or should the core knowledge be \"insulated\" by freezing its weights?",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "5f5ceeda-3639-451c-9adb-5187cfc8512f",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.6,
        "x": 0.03,
        "y": 0.14
      },
      "kind": "title",
      "text": "The architectural divide: knowledge insulation vs. end-to-end adaptation",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "77ae01ce-4850-46b3-82e2-d24c0d67d98f",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Contrast Pairs",
      "slug": "contrast-pairs",
      "agent": "Storyteller",
      "layer": "loop",
      "matchId": "019dd95a-1055-74e0-a95c-2c240a80d685",
      "evidence": "'Knowledge insulation vs. end-to-end adaptation' framed as architectural divide.",
      "confidence": 85
    }
  ],
  "frameworks": [
    {
      "name": "comparison_frame",
      "slug": null,
      "matchId": "3bcfc242-ae02-4310-a439-6ca2f376acf7",
      "evidence": "Compares two distinct architectural approaches (insulation vs end-to-end) for robotic agents.",
      "confidence": 0.9
    }
  ],
  "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
}