{
  "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": 14,
  "pageCount": 163,
  "prevPage": 13,
  "nextPage": 15,
  "slideType": "key_takeaways",
  "function": "summarize",
  "density": "overcrowded",
  "nDataPoints": 10,
  "notes": "The slide references Berkeley research on imitation learning and Meta's LLaMA-2 paper findings on RLHF.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "line_chart",
    "scatter_plot"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-06b04d219fea/14",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea#slide-14",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "We posit that the superior writing abilities of LLMs, as manifested in surpassing human annotators in certain tasks, are fundamentally driven by RLHF.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c4-7719-a221-29a0c0e34941",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.35,
        "w": 0.23,
        "x": 0.47,
        "y": 0.52
      },
      "kind": "chart",
      "text": "Natural Questions 3-Shot",
      "attrs": null,
      "subkind": "line",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "be5216f7-2b5e-4fae-8769-3e30d9fe6889",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.35,
        "w": 0.23,
        "x": 0.73,
        "y": 0.52
      },
      "kind": "chart",
      "text": "Crowdworker Evaluation",
      "attrs": null,
      "subkind": "line",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "daefd946-39ff-4c65-9f87-ccfc36a604f1",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.15,
        "w": 0.45,
        "x": 0.03,
        "y": 0.49
      },
      "kind": "list",
      "text": "By using model size as a proxy for quality, the authors argue that more attention should be paid to better pre-training rather than fine-tuning on more imitation data.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "2c438ecd-9e75-426f-9b64-552e909aaa22",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.15,
        "w": 0.95,
        "x": 0.03,
        "y": 0.33
      },
      "kind": "list",
      "text": "The researchers examine a range of pretrained LLMs of different sizes and pre-trained on a varying amount of data. They show that at a fixed model size, using more imitation data actually hurts the quality of the output. In turn, larger models benefit from using imitation data.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "419869fe-d830-48f4-83f1-1dc534194ee4",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.1,
        "w": 0.95,
        "x": 0.03,
        "y": 0.21
      },
      "kind": "list",
      "text": "Berkeley researchers show that fine-tuning small LLMs on the outputs of larger, more capable LLMs results in models which are stylistically impressive but which often produce inaccurate text.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "5037c2b7-bb17-44d5-93ea-3615ac111057",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.45,
        "x": 0.03,
        "y": 0.65
      },
      "kind": "list",
      "text": "In the near future, RLHF seems here to stay. After careful ablation studies, Meta researchers concluded in their LLaMA-2 paper: \"We posit that the superior writing abilities of LLMs, as manifested in surpassing human annotators in certain tasks, are fundamentally driven by RLHF\".",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "7565cf54-8793-4195-8616-08896cc0055c",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "Accuracy (%)",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-47c4-7719-a221-311e619800f8",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "quote",
      "text": "We posit that the superior writing abilities of LLMs, as manifested in surpassing human annotators in certain tasks, are fundamentally driven by RLHF. (Meta researchers, LLaMA-2 paper)",
      "attrs": null,
      "subkind": null,
      "toolName": "Authority citation",
      "toolSlug": "authority-citation",
      "confidence": null,
      "componentId": "019dd952-47c4-7719-a221-2dfee32fa9bd",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.75,
        "x": 0.03,
        "y": 0.13
      },
      "kind": "title",
      "text": "The false promise of imitating proprietary LLMs, or how RLHF is still king",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "339bc98f-fb59-4de8-8b5a-e09e5ef04806",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Action Titles",
      "slug": "action-titles",
      "agent": "Architect",
      "layer": "slide",
      "matchId": "019dd95a-1055-74e0-a957-0ef647f6c7b1",
      "evidence": "Title makes a sharp claim ('The false promise of imitating proprietary LLMs').",
      "confidence": 90
    }
  ],
  "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
}