{
  "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": 26,
  "pageCount": 213,
  "prevPage": 25,
  "nextPage": 27,
  "slideType": "industry_trends",
  "function": "present_solution",
  "density": "dense",
  "nDataPoints": 1,
  "notes": "Mentions Stanford research on ReFT and LoReFT.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "process_diagram",
    "paragraph"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0856e1444fb9/26",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9#slide-26",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "ReFT requires 15-65x fewer parameters compared to weight-based fine-tuning methods.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d0f-c1903d0499f2",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.41,
        "x": 0.56,
        "y": 0.52
      },
      "kind": "image",
      "text": "ReFT Intervention and LoReFT diagrams",
      "attrs": null,
      "subkind": "infographic",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "ccfeeccb-c196-4624-ad9b-47b14b701bb6",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.58,
        "w": 0.52,
        "x": 0.03,
        "y": 0.32
      },
      "kind": "list",
      "text": "Inspired by model interpretability research, ReFT (Representation Fine-tuning) doesn't alter the model's weights. Instead, it manipulates the model's internal representations at inference time to steer its behavior.\nWhile it comes with a slight interference penalty, ReFT requires 15-65x fewer parameters compared to weight-based fine-tuning methods.\nIt also enables more selective interventions on specific layers and token positions, enabling fine-grained control over the adaptation process.\nThe researchers show its potential in few-shot adaptation where a chat model is given a new persona with just five examples. Combined with the small storage footprint for learned interventions, it could be used for real-time personalization on devices with sufficient compute power.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "290a2108-daba-423d-97b3-a042df9b41c0",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.07,
        "w": 0.94,
        "x": 0.03,
        "y": 0.22
      },
      "kind": "paragraph",
      "text": "Parameter-efficient fine-tuning (e.g. via LoRA) is nothing new, but Stanford researchers believe a more targeted approach offers greater efficiency and adaptation.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "8506cb3a-8086-4e8b-875f-bb343dfdc765",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.15,
        "w": 0.41,
        "x": 0.56,
        "y": 0.78
      },
      "kind": "source-note",
      "text": "Figure 2: Illustration of ReFT. (1) The left panel depicts an intervention I: the intervention function Φ is applied to hidden representations at positions P in layer L. (2) The right panel depicts the intervention function used in LoReFT...",
      "attrs": null,
      "subkind": null,
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "53529a7e-19b1-4bd8-a028-9c7bec1d9f06",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.68,
        "x": 0.03,
        "y": 0.14
      },
      "kind": "title",
      "text": "Will representation fine tuning unlock on-device personalization?",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "72eb22ea-42e8-4f32-befd-e4c939c23286",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "2x2 matrix",
      "slug": "matrix-2x2",
      "agent": null,
      "layer": "slide",
      "matchId": "4c773229-9ddb-4093-a657-fef248111938",
      "evidence": "ReFT Intervention and LoReFT diagrams",
      "confidence": 0.5
    }
  ],
  "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": 29,
      "from": 21,
      "name": "Pattern Hunter",
      "slug": "02-pattern-hunter",
      "bestFor": "Time-pressed audiences, building consensus, when data is strong",
      "matchId": "019dd95a-07fe-70ce-8d3e-2b2229583eee",
      "evidence": "Eight slides on shrinking models, on-device, hybrids, transformers reigning",
      "position": 3,
      "objective": "Demonstrate efficiency frontier across distillation, quantization, hybrids",
      "structure": "Evidence A -> Evidence B -> Evidence C -> Pattern/Conclusion",
      "confidence": 75,
      "description": "Group multiple pieces of evidence that together point to a pattern or conclusion"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}