{
  "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": 152,
  "pageCount": 163,
  "prevPage": 151,
  "nextPage": 153,
  "slideType": "diagnosis",
  "function": "diagnose",
  "density": "dense",
  "nDataPoints": 0,
  "notes": "The slide uses a complexity-theory-inspired notation (NP^NP) to illustrate the recursive difficulty of evaluating AI assistants.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "paragraph",
    "comparison_table"
  ],
  "metadataConfidence": 0.95,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-06b04d219fea/152",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea#slide-152",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "Jan Leike, the leader of OpenAI's alignment team, frames this as a recursive reward modeling problem, which ends in LMs potentially reward hacking and humans being unable to detect it.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c5-73ac-aa5e-de2f53f65891",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.45,
        "w": 0.55,
        "x": 0.02,
        "y": 0.38
      },
      "kind": "list",
      "text": "Evaluating automated rule-based systems with other automated systems is nothing new. But when the systems generate stochastic creative content, only humans seemingly have the cognitive ability to evaluate their safety. With an AI at hand, maybe humans can augment their supervision capabilities. But that is if humans can reasonably evaluate the AI assistant.\nWhen this last condition isn't ensured, as AI assistants are virtually impossible to holistically evaluate, humans need AI assistants to evaluate the AI assistants, and so on. Jan Leike, the leader of OpenAI's alignment team, frames this as a recursive reward modeling problem, which ends in LMs potentially reward hacking and humans being unable to detect it.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "19cbda13-ba5f-4344-a264-7ea4fdfa265f",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.15,
        "w": 0.55,
        "x": 0.02,
        "y": 0.21
      },
      "kind": "paragraph",
      "text": "As models become more capable and generate outputs that surpass our ability to monitor them (in volume or intricacy for example), one way forward which is already being explored is using AI to assist human supervision. But without AI alignment, AI-assisted monitoring opens the way for a spiral of increasingly uncertain evaluation.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "3fc6e5f4-a513-43f1-8753-cfa487e0f92f",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.45,
        "w": 0.38,
        "x": 0.6,
        "y": 0.38
      },
      "kind": "table",
      "text": "Level 2 (≅ NP^NP) Humans cannot reliably evaluate what AI is doing. Humans can evaluate what AI is doing with AI assistance. Humans can evaluate this assistance. Level 3 (≅ NP^NP^NP) Humans cannot reliably evaluate what AI is doing. Humans can evaluate what AI is doing with AI assistance. Humans cannot reliably evaluate this assistance. Humans can evaluate this assistance with assistance. Humans can evaluate this assistance eval assistance.",
      "attrs": null,
      "subkind": "data",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "9c64d97f-83a7-4a10-95bd-48bf5246879a",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.35,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "How hard is scalable supervision?",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "2f0eacaf-1fb5-4f1c-b640-645ddcd61a56",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Table data",
      "slug": "table-data",
      "agent": null,
      "layer": "slide",
      "matchId": "65bbfc25-5128-444e-b484-1ba94414b879",
      "evidence": "table/data: Level 2 (≅ NP^NP) Humans cannot reliably evaluate what AI is doing. Humans can evaluate what AI is doing with AI assistance. Humans can evaluate this assistance.",
      "confidence": 0.8
    }
  ],
  "frameworks": [
    {
      "name": "recursive-reward-modeling",
      "slug": null,
      "matchId": "294cac45-f157-4b5f-ac09-32faab193ba4",
      "evidence": "Jan Leike, the leader of OpenAI's alignment team, frames this as a recursive reward modeling problem",
      "confidence": 0.9
    }
  ],
  "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
}