{
  "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": 154,
  "pageCount": 163,
  "prevPage": 153,
  "nextPage": 155,
  "slideType": "industry_trends",
  "function": "diagnose",
  "density": "dense",
  "nDataPoints": 1,
  "notes": "The chart shows a negative correlation between subject model parameters and explanation score.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "paragraph",
    "line_chart"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-06b04d219fea/154",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea#slide-154",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "One worrisome fact is that the explanation score seems to decrease as the explained models get bigger.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c4-7719-a230-39360dc20322",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.45,
        "w": 0.3,
        "x": 0.65,
        "y": 0.4
      },
      "kind": "chart",
      "text": "Line chart showing explanation score vs subject model parameters",
      "attrs": null,
      "subkind": "line",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "63512dcb-692a-4dd0-a16d-8e2ed5008097",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.35,
        "w": 0.6,
        "x": 0.02,
        "y": 0.38
      },
      "kind": "list",
      "text": "The goal of their method is to explain which patterns in text cause a neuron to activate. GPT-4 takes as input a part the text and neuron activations, and is prompted to generate an explanation of what causes neurons to activate. Then, on other parts of text, GPT-4 is prompted to predict where neurons will most strongly respond. The researchers can then derive a similarity score between the predicted and real activations, which they dub “explanation score”: “a measure of a language model's ability to compress and reconstruct neuron activations using natural language”.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "c6d3e000-3fde-4561-a2c2-161bcdf6aafc",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.1,
        "w": 0.6,
        "x": 0.02,
        "y": 0.75
      },
      "kind": "list",
      "text": "One worrisome fact is that the explanation score seems to decrease as the explained models get bigger.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "f3b073da-6023-43c3-994b-fb426e177c15",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "Explanation score",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-47c4-7719-a230-3ef26ec8abd8",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.15,
        "w": 0.6,
        "x": 0.02,
        "y": 0.21
      },
      "kind": "paragraph",
      "text": "Mechanistic interpretability aims at explaining the roles of specific neurons/groups of neurons in the outputs of deep learning models. Not only is this task hard, but current approaches to solving it are also not scalable to billions of neurons. Doubling down on AI supervision, OpenAI proposes using GPT-4 to explain neurons in smaller language models. They test this method on GPT-2.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "18b3f5b6-0393-4239-93a5-2fa21816d460",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.6,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "Going deeper into the models: LLM-powered mechanistic interpretability",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "a2bccfc6-ea78-49c4-b30b-61112989c914",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [],
  "frameworks": [],
  "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
}