{
  "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": 195,
  "pageCount": 213,
  "prevPage": 194,
  "nextPage": 196,
  "slideType": "industry_trends",
  "function": "diagnose",
  "density": "dense",
  "nDataPoints": 2,
  "notes": "The slide discusses uncertainty calibration in LLMs using fine-tuning techniques like LoRA.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "paragraph",
    "process_diagram",
    "bar_chart_horizontal"
  ],
  "metadataConfidence": 0.95,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0856e1444fb9/195",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9#slide-195",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "Research from NYU, Abacus AI, and Cambridge found that fine-tuning LLMs on a dataset of correct and incorrect answers can significantly improve the calibration of their uncertainty estimates.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d21-29e2bdaf7ca1",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.17,
        "x": 0.63,
        "y": 0.68
      },
      "kind": "chart",
      "text": "Comparison of Zero-Shot Classifier, Verbalized, Sampling, and Fine-Tuned methods on ECE and AUROC metrics.",
      "attrs": null,
      "subkind": "bar-horizontal",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "5cdd009a-c24b-462b-b453-1ed3df4584f0",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.6,
        "x": 0.2,
        "y": 0.68
      },
      "kind": "image",
      "text": "Diagram showing the process of fine-tuning LLMs for uncertainty estimation and a bar chart comparing ECE and AUROC metrics.",
      "attrs": null,
      "subkind": "infographic",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "e4fcfa5c-f101-4373-b410-eab922320ae6",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.35,
        "w": 0.95,
        "x": 0.02,
        "y": 0.3
      },
      "kind": "list",
      "text": "Research from NYU, Abacus AI, and Cambridge found that fine-tuning LLMs on a dataset of correct and incorrect answers can significantly improve the calibration of their uncertainty estimates. This requires only a small amount of additional data (around 1,000 examples) and can be done efficiently using techniques like LoRA.\nThe resulting uncertainty estimates generalize well to new question types and tasks, even when they are different from the ones used for the fine-tuning.\nBetter still, the fine-tuned models can also be used to estimate the uncertainty of other models.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "563095bd-81f7-47a9-9f87-2edad7134f32",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "AUROC: 70%",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d21-2c10c42200ef",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.08,
        "w": 0.95,
        "x": 0.02,
        "y": 0.21
      },
      "kind": "paragraph",
      "text": "LLMs often struggle to assign reliable confidence estimates to their outputs, even when pushed on whether or not an answer is correct. The solution potentially lies in fine-tuning, rather than better zero-shot prompting.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "3bf2a42b-ea2d-47c7-856d-dbe1559cf3b8",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.4,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "Can we make the known unknowns known?",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "105271ac-8127-402a-892f-a4d5465c4ff8",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 204,
      "from": 154,
      "beatId": "019dd95a-0682-776c-8e35-605159a069e6",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Implications (So What)",
      "beatSlug": "triple-take-the-implications-so-what",
      "evidence": "Politics + Safety sections explore consequences and risks",
      "position": 2,
      "confidence": 78,
      "parentBeatName": "Reflection",
      "parentBeatSlug": "reflection"
    },
    {
      "to": 204,
      "from": 175,
      "beatId": "019dd95a-0682-776c-8e35-74f05e198faf",
      "arcName": "Voyage and Return",
      "arcSlug": "voyage-return",
      "beatName": "The Return",
      "beatSlug": "voyage-return-the-return",
      "evidence": "Safety section returns to risks/governance debate",
      "position": 4,
      "confidence": 55,
      "parentBeatName": "Resolution",
      "parentBeatSlug": "resolution"
    }
  ],
  "loops": [
    {
      "to": 201,
      "from": 187,
      "name": "Pattern Hunter",
      "slug": "02-pattern-hunter",
      "bestFor": "Time-pressed audiences, building consensus, when data is strong",
      "matchId": "019dd95a-07fe-70ce-8d3e-79a3afd2ac42",
      "evidence": "Emergent capabilities, sycophancy, DPO, hybrid RLHF, critiques, transparency, reward tampering, SAEs",
      "position": 23,
      "objective": "Catalogue alignment, RLHF, sycophancy and interpretability research",
      "structure": "Evidence A -> Evidence B -> Evidence C -> Pattern/Conclusion",
      "confidence": 76,
      "description": "Group multiple pieces of evidence that together point to a pattern or conclusion"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}