{
  "docId": "019dd923-5e88-73ef-bd5d-01cf0d8a8fbc",
  "docSlug": "cc8183ec02431b7a",
  "documentTitle": "2022 Air Street Capital The State of AI Report 2022",
  "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": 100,
  "pageCount": 114,
  "prevPage": 99,
  "nextPage": 101,
  "slideType": "industry_trends",
  "function": "present_solution",
  "density": "overcrowded",
  "nDataPoints": 4,
  "notes": "Includes a chart showing Elo scores vs parameters and a chat example of a safety-aligned model.",
  "elementsJson": [
    "headline_text",
    "paragraph",
    "bullet_list",
    "line_chart",
    "screenshot"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-01cf0d8a8fbc/100",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-01cf0d8a8fbc",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-01cf0d8a8fbc.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-01cf0d8a8fbc#slide-100",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "RLHF has emerged as a key method to finetune LLMs and align them with human values.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c4-7719-a21f-23ccd1d0233f",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.3,
        "w": 0.4,
        "x": 0.55,
        "y": 0.35
      },
      "kind": "chart",
      "text": "Helpfulness Scores chart",
      "attrs": null,
      "subkind": "line",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "a5c8df7c-ca72-4db1-ad37-2db9e6ece5c2",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.2,
        "w": 0.4,
        "x": 0.55,
        "y": 0.7
      },
      "kind": "image",
      "text": "Chat example showing safety alignment",
      "attrs": null,
      "subkind": "screenshot",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "f48afc43-339d-4172-a59c-6da3a5a75dbd",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.45,
        "x": 0.05,
        "y": 0.35
      },
      "kind": "list",
      "text": "OpenAI started the year by finetuning GPT-3 using RLHF to produce InstructGPT models that improved on helpfulness for instruction-following tasks. Notably, the fine-tuning only needed <2% of GPT-3's pretraining compute, as well as 20,000 hours of human feedback. API users on average prefer these models to the original ones.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "246b5a42-6fbf-4d8f-8bb5-3bf10aa72e87",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.15,
        "w": 0.45,
        "x": 0.05,
        "y": 0.62
      },
      "kind": "list",
      "text": "RLHF has also been used by both Anthropic and Deepmind to improve the helpfulness, harmlessness, and correctness of their language models. OpenAI has stated that RLHF will be a core building block for its long-term approach to aligning AGI.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "b76ada0c-dfe1-4dfa-aad1-4a1c514e14f0",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "Elo Scores: 2%",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-47c4-7719-a21f-25ea31634b4a",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.9,
        "x": 0.05,
        "y": 0.21
      },
      "kind": "paragraph",
      "text": "RLHF has emerged as a key method to finetune LLMs and align them with human values. This involves humans ranking language model outputs sampled for a given input, using these rankings to learn a reward model of human preferences, and then using this as a reward signal to finetune the language model with using RL.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "17b486cd-b205-4e97-950b-6e08365a89d6",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.8,
        "x": 0.05,
        "y": 0.14
      },
      "kind": "title",
      "text": "Language Model Alignment: Reinforcement Learning from Human Feedback (RLHF)",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "9e46f67a-d892-4806-bb22-a2596a64e0c8",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Action Titles",
      "slug": "action-titles",
      "agent": "Architect",
      "layer": "slide",
      "matchId": "019dd95a-1055-74e0-a956-c4b008568cd7",
      "evidence": "Title names the specific technique 'Reinforcement Learning from Human Feedback (RLHF)'.",
      "confidence": 80
    },
    {
      "name": "Process flow",
      "slug": "process-flow",
      "agent": null,
      "layer": "slide",
      "matchId": "8dae08d7-7a0f-402d-a5f4-61cec6bda2d2",
      "evidence": "The slide outlines the process of Reinforcement Learning from Human Feedback (RLHF) for language model alignment.",
      "confidence": 0.8
    }
  ],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 106,
      "from": 81,
      "beatId": "019dd95a-0682-776c-8e35-28fc5bded874",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Implications (So What)",
      "beatSlug": "triple-take-the-implications-so-what",
      "evidence": "Politics and Safety sections interpret what the facts mean for governance and risk.",
      "position": 2,
      "confidence": 78,
      "parentBeatName": "Reflection",
      "parentBeatSlug": "reflection"
    },
    {
      "to": 106,
      "from": 94,
      "beatId": "019dd95a-0682-776c-8e35-381428a808af",
      "arcName": "The Mountain",
      "arcSlug": "mountain",
      "beatName": "Climax",
      "beatSlug": "mountain-climax",
      "evidence": "Safety section confronts existential risk and alignment urgency.",
      "position": 3,
      "confidence": 55,
      "parentBeatName": "Turn",
      "parentBeatSlug": "turn"
    }
  ],
  "loops": [
    {
      "to": 106,
      "from": 100,
      "name": "Pattern Hunter",
      "slug": "02-pattern-hunter",
      "bestFor": "Time-pressed audiences, building consensus, when data is strong",
      "matchId": "019dd95a-07fe-70ce-8d3d-db1f31ffae02",
      "evidence": "RLHF, red-teaming, mechanistic interpretability, goal misgeneralization, moral behavior, Conjecture lab.",
      "position": 17,
      "objective": "Catalogue alignment-research techniques and example labs",
      "structure": "Evidence A -> Evidence B -> Evidence C -> Pattern/Conclusion",
      "confidence": 78,
      "description": "Group multiple pieces of evidence that together point to a pattern or conclusion"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}