{
  "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": 194,
  "pageCount": 213,
  "prevPage": 193,
  "nextPage": 195,
  "slideType": "industry_trends",
  "function": "summarize",
  "density": "dense",
  "nDataPoints": 2,
  "notes": "Discusses CriticGPT and Cohere's research on LLM-generated critiques.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "screenshot"
  ],
  "metadataConfidence": 0.95,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0856e1444fb9/194",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9#slide-194",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "They found particularly strong results for weaker base models or in low-data settings, with one high-quality critique-enhanced preference pair can be worth up to 40 standard preference pairs.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d23-0c51f4ac029d",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.45,
        "w": 0.3,
        "x": 0.65,
        "y": 0.4
      },
      "kind": "image",
      "text": "Diagram showing the CriticGPT process: Task selection, Bug Insertion, Critique Comparison, Critique rating, Critique bug inclusion, Overall ranking.",
      "attrs": null,
      "subkind": "infographic",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "7d957b21-ae45-4a83-8136-97e420c5c57a",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.65,
        "w": 0.6,
        "x": 0.05,
        "y": 0.22
      },
      "kind": "list",
      "text": "The concept of 'LLMs as judges' lives on, with major labs extending it beyond the simple evaluation of outputs. OpenAI unveiled CriticGPT, which uses an GPT-style LLM trained on a huge dataset of flawed inputs, to spot mistakes in code generated by other LLMs. It outperformed human contractors in catching errors and its critiques were preferred to human-written ones 63% of the time. The system was also able to spot bugs in training data that has been labelled as 'flawless'. Meanwhile, Cohere have explored the possibility of using LLM-generated critiques to enhance reward models for RLHF. They used a range of LLMs to generate point-wise critiques for each preference data pair, designed to have the LLM evaluate the effectiveness of the prompt-completion pair. They found particularly strong results for weaker base models or in low-data settings, with one high-quality critique-enhanced preference pair can be worth up to 40 standard preference pairs.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "d0dea33d-5a9b-45bd-a6d9-917a6178084b",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "preference rate: 63%",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d23-1234f86eebc7",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.6,
        "x": 0.05,
        "y": 0.14
      },
      "kind": "title",
      "text": "Can LLM-generated critiques enhance both accuracy and alignment?",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "45d89355-f75c-4412-aff4-faee57b33f54",
      "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
}