{
  "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": 15,
  "pageCount": 163,
  "prevPage": 14,
  "nextPage": 16,
  "slideType": "industry_trends",
  "function": "summarize",
  "density": "overcrowded",
  "nDataPoints": 1,
  "notes": "The slide discusses alternatives to Reinforcement Learning from Human Feedback (RLHF) in AI training.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "screenshot",
    "paragraph"
  ],
  "metadataConfidence": 0.95,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-06b04d219fea/15",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea#slide-15",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "According to human evaluations of model outputs, LIMA is competitive with GPT-4 in 43% of cases.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c4-7719-a221-da0bc182c9f5",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.45,
        "w": 0.28,
        "x": 0.7,
        "y": 0.33
      },
      "kind": "image",
      "text": "Example tasks from generated instruction data",
      "attrs": null,
      "subkind": "screenshot",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "948efe41-eca1-4c32-a45b-fb4afc0144f8",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.15,
        "w": 0.65,
        "x": 0.02,
        "y": 0.38
      },
      "kind": "list",
      "text": "Other approaches entirely circumvent the use of reinforcement learning. In Less is More for Alignment (LIMA), Meta argues for using a few (1,000 in their paper) very carefully curated prompts and responses. According to human evaluations of model outputs, LIMA is competitive with GPT-4 in 43% of cases.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "1916c0b1-83ef-4b79-8321-c9eda4f351d0",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.2,
        "w": 0.65,
        "x": 0.02,
        "y": 0.54
      },
      "kind": "list",
      "text": "In LLMs can self-improve, Google researchers showed that LLMs can improve by training on their own outputs. In a similar vein, Self-Instruct is a framework in which a model generates its own instructions, input and output samples, and curates them to finetune its parameters. Yet another work in this direction is Meta's Self-Alignment with Instruction Backtranslation.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "76505bd8-9f2b-4f25-b299-cea2d56f5893",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.65,
        "x": 0.02,
        "y": 0.75
      },
      "kind": "list",
      "text": "Stanford researchers used this last approach to generate instructions and outputs using GPT-3.5 and fine-tune Meta's LLaMa-7B.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "a5bd89fd-7fa2-4c19-92aa-4dab3c2a6bb5",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.65,
        "x": 0.02,
        "y": 0.33
      },
      "kind": "list",
      "text": "Anthropic proposed RL from AI feedback, which we cover in the safety section.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "bfe83e33-3099-46c2-b430-57f1ce43af30",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.1,
        "w": 0.65,
        "x": 0.02,
        "y": 0.22
      },
      "kind": "list",
      "text": "In the wake of ChatGPT, many labs set out to answer the question: Can we create models as capable and safe as OpenAI's LLMs, but that drastically reduce human supervision?",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "cec1bdd6-88ba-4087-a576-737b466c8032",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.08,
        "w": 0.28,
        "x": 0.7,
        "y": 0.79
      },
      "kind": "paragraph",
      "text": "Figure 1: Selected tasks from the generated instruction data using vanilla GPT3. Some texts are reformatted for presentation. See Table 10 for more examples.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "172357c0-ad89-436a-ad14-495183cbe63c",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.6,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "Even so, researchers rush to find scalable alternatives to RLHF",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "0bcca668-851a-41c9-a8a8-e96c28fd23f0",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Annotation",
      "slug": "annotation",
      "agent": "Designer",
      "layer": "slide",
      "matchId": "019dd95a-1055-74e0-a957-10345ebf9eca",
      "evidence": "'43% competitive with GPT-4' annotation in callout.",
      "confidence": 65
    },
    {
      "name": "List presentation",
      "slug": "list-presentation",
      "agent": null,
      "layer": "slide",
      "matchId": "8a655c6f-8be3-4e93-87b3-2f637205137d",
      "evidence": "The slide contains a list of bullet points discussing various approaches to scalable alternatives to RLHF.",
      "confidence": 0.8
    }
  ],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 120,
      "from": 11,
      "beatId": "019dd95a-0682-776c-8e35-41afd44ef59f",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Facts (What)",
      "beatSlug": "triple-take-the-facts-what",
      "evidence": "Research + Industry sections inventory model, compute, funding facts.",
      "position": 1,
      "confidence": 78,
      "parentBeatName": "Setup",
      "parentBeatSlug": "setup"
    },
    {
      "to": 120,
      "from": 11,
      "beatId": "019dd95a-0682-776c-8e35-523bfb7f96e6",
      "arcName": "The Mountain",
      "arcSlug": "mountain",
      "beatName": "Rising Action",
      "beatSlug": "mountain-rising-action",
      "evidence": "Escalating capabilities, compute concentration and capital flows.",
      "position": 2,
      "confidence": 45,
      "parentBeatName": "Development",
      "parentBeatSlug": "development"
    }
  ],
  "loops": [
    {
      "to": 21,
      "from": 12,
      "name": "Pattern Hunter",
      "slug": "02-pattern-hunter",
      "bestFor": "Time-pressed audiences, building consensus, when data is strong",
      "matchId": "019dd95a-07fe-70ce-8d3d-e61d00ffc6e4",
      "evidence": "Sequential GPT-4, RLHF, LLaMa, popularity-mention slides converge on a single conclusion.",
      "position": 2,
      "objective": "Pile up evidence that GPT-4 and RLHF dominate the LLM landscape",
      "structure": "Evidence A -> Evidence B -> Evidence C -> Pattern/Conclusion",
      "confidence": 82,
      "description": "Group multiple pieces of evidence that together point to a pattern or conclusion"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}