{
  "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": 36,
  "pageCount": 313,
  "prevPage": 35,
  "nextPage": 37,
  "slideType": "industry_trends",
  "function": "present_solution",
  "density": "overcrowded",
  "nDataPoints": 15,
  "notes": "The chart shows Bits per Byte performance across different model sizes and training methods (Base, Context, Fine-Tuning).",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "bar_chart_grouped"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/36",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-36",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "The scaling paradigm is shifting from static pre-training to dynamic, on-the-fly adaptation.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-41ac-7180-a81c-f5bbdba8688e",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.15,
        "w": 0.35,
        "x": 0.35,
        "y": 0.82
      },
      "kind": "chart",
      "text": "Bits per Byte performance comparison across models and methods.",
      "attrs": null,
      "subkind": "bar-grouped",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "ac9aced0-df8c-471d-a2e0-23d6b02c663f",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.15,
        "w": 0.93,
        "x": 0.04,
        "y": 0.44
      },
      "kind": "list",
      "text": "This on-demand learning consistently outperforms in-context learning, especially on complex tasks. It creates a new performance vector independent of pre-training scale. An actively fine-tuned 3.8B Phi-3 model (red bars) can outperform a base 27B Gemma-2 model. Admittedly these models are a little old.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "3145c182-835a-4cf4-8fa6-33a133f1e0b9",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.13,
        "w": 0.93,
        "x": 0.04,
        "y": 0.3
      },
      "kind": "list",
      "text": "From naive retrieval to active selection: Early methods used simple Nearest Neighbor retrieval, often selecting redundant data. New algorithms like ETH Zürich's SIFT now integrate active learning to select small, diverse, and maximally informative examples for each query.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "3ec7f46d-6c27-427d-9b94-57cbfbbdedb8",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.2,
        "w": 0.93,
        "x": 0.04,
        "y": 0.6
      },
      "kind": "list",
      "text": "A recent follow-up, Local Mixtures of Experts (test-time model merging), amortizes TTT by training small neighborhood experts and, at inference, retrieving and merging a few weight deltas into the base model. It keeps most SIFT-style gains with near-retrieval latency and, on ~1B bases, approaches TTT accuracy while running up to ~100x faster. Titans studies test-time memorization as an architectural memory and is orthogonal to amortized TTT.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "f659defc-4e65-4e9e-b519-8481b127f165",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "Bits per Byte",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-41ac-7180-a81c-fad88e8294de",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.08,
        "w": 0.95,
        "x": 0.02,
        "y": 0.21
      },
      "kind": "paragraph",
      "text": "The scaling paradigm is shifting from static pre-training to dynamic, on-the-fly adaptation. Test-time fine-tuning (TTT) adapts a model's weights to a specific prompt at inference, a step towards continuous learning.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "1e2ac3ed-ed35-4afd-83d1-1e3cfd864941",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.45,
        "x": 0.02,
        "y": 0.13
      },
      "kind": "title",
      "text": "Beyond reasoning is...continual learning?",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "cf191a5e-9a79-4295-815c-fdb17371108c",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Audience Definition",
      "slug": "audience-definition",
      "agent": "Storyteller",
      "layer": "slide",
      "matchId": "8cb89ed9-6dc6-45f2-9e92-96b9ec860d14",
      "evidence": "The components mention adapting a model's weights to a specific prompt at inference, indicating an understanding of the audience's needs.",
      "confidence": 0.7
    }
  ],
  "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
}