{
  "docId": "019dd923-5e88-73ef-bd5d-06b04d219fea",
  "docSlug": "dd91c78f6570bf29",
  "documentTitle": "2023 Air Street Capital The State of AI Report 2023",
  "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": 153,
  "pageCount": 163,
  "prevPage": 152,
  "nextPage": 154,
  "slideType": "industry_trends",
  "function": "present_framework",
  "density": "overcrowded",
  "nDataPoints": 4,
  "notes": "Discusses 'Let's Verify Step by Step' (OpenAI) and 'Contrast-Consistent Search' (UC Berkeley/Peking University).",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "paragraph",
    "process_diagram",
    "histogram",
    "other"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-06b04d219fea/153",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea#slide-153",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "To ensure that they are reliable, a potentially successful approach could be to train the model to have the right process leading to the output.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c5-73ac-aa5f-8ffc2b08a993",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.13,
        "x": 0.83,
        "y": 0.52
      },
      "kind": "chart",
      "text": "Probability distribution chart",
      "attrs": null,
      "subkind": "histogram",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "9ab2c808-689d-47c3-904d-9d9a35b28060",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.35,
        "w": 0.58,
        "x": 0.38,
        "y": 0.52
      },
      "kind": "diagram",
      "text": "Contrast-Consistent Search workflow diagram",
      "attrs": null,
      "subkind": "process",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "2718daa6-fbeb-44b8-87d4-7a401a23808a",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.15,
        "w": 0.92,
        "x": 0.04,
        "y": 0.34
      },
      "kind": "list",
      "text": "In “Let’s Verify Step by Step”, researchers from OpenAI train a reward model to predict the correctness of each step involved in solving a math problem. To do so, they generate (and release) a synthetic dataset of 800K labeled steps across 75K solutions to 12K problems. They achieved a top performance of 78.2% on a representative subset of the MATH test.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "01572acc-020b-4533-8e48-17f2b58b4278",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.42,
        "x": 0.04,
        "y": 0.52
      },
      "kind": "list",
      "text": "Other researchers set out to enforce consistency in model outputs. Their method, Contrast-Consistent Search, enforces the fact that if a model assigns a probability p to answering “yes” to a binary question, it should assign a probability 1-p to “no” for the same question.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "2a8b6761-60b0-434b-9fa9-a511b9f2572a",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "accuracy: 78.2%",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-47c5-73ac-aa5f-91e73a73bb49",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.92,
        "x": 0.04,
        "y": 0.2
      },
      "kind": "paragraph",
      "text": "Much of LLM evaluation relies on examining the final outputs of Language Models. But to ensure that they are are reliable, a potentially successful approach could be to train the model to have the right process leading to the output. New research from OpenAI and from UC Berkeley and Peking University explores this direction.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "16acf905-78ca-4b97-9354-e1a63d8dd48f",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.45,
        "x": 0.04,
        "y": 0.14
      },
      "kind": "title",
      "text": "Evaluating the process and the outcome",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "60347d10-7fd3-44bf-af3b-7c94e73de99f",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [],
  "frameworks": [
    {
      "name": "Contrast-Consistent Search",
      "slug": null,
      "matchId": "b7bee608-3d62-4f8f-9ce4-d9864f2d2067",
      "evidence": "Methodology described in text and diagram",
      "confidence": 1
    }
  ],
  "arcBeats": [
    {
      "to": 155,
      "from": 121,
      "beatId": "019dd95a-0682-776c-8e35-478c8c58938c",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Implications (So What)",
      "beatSlug": "triple-take-the-implications-so-what",
      "evidence": "Politics + Safety sections explore regulatory, geopolitical and risk implications.",
      "position": 2,
      "confidence": 78,
      "parentBeatName": "Reflection",
      "parentBeatSlug": "reflection"
    },
    {
      "to": 155,
      "from": 121,
      "beatId": "019dd95a-0682-776c-8e35-554ec1797b3a",
      "arcName": "The Mountain",
      "arcSlug": "mountain",
      "beatName": "Climax",
      "beatSlug": "mountain-climax",
      "evidence": "Regulatory divergence and the x-risk debate erupt into the mainstream.",
      "position": 3,
      "confidence": 45,
      "parentBeatName": "Turn",
      "parentBeatSlug": "turn"
    }
  ],
  "loops": [
    {
      "to": 155,
      "from": 149,
      "name": "Pattern Hunter",
      "slug": "02-pattern-hunter",
      "bestFor": "Time-pressed audiences, building consensus, when data is strong",
      "matchId": "019dd95a-07fe-70ce-8d3e-1a62463741ad",
      "evidence": "RLHF survey, conditional pretraining, Constitutional AI, scalable supervision, process eval, interpretability, benchmark drift.",
      "position": 15,
      "objective": "Show alignment research progress on RLHF problems",
      "structure": "Evidence A -> Evidence B -> Evidence C -> Pattern/Conclusion",
      "confidence": 70,
      "description": "Group multiple pieces of evidence that together point to a pattern or conclusion"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}