{
  "docId": "019dd923-5e88-73ef-bd5d-0856e1444fb9",
  "docSlug": "8757f1b44ef7f176",
  "documentTitle": "2024 Air Street Capital The State of AI Report 2024",
  "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": 63,
  "pageCount": 213,
  "prevPage": 62,
  "nextPage": 64,
  "slideType": "key_takeaways",
  "function": "analyze_data",
  "density": "overcrowded",
  "nDataPoints": 2,
  "notes": "The slide references 'grokking' (training beyond overfitting) and mechanistic interpretability findings from Ohio State University researchers.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "line_chart",
    "process_diagram"
  ],
  "metadataConfidence": 0.95,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0856e1444fb9/63",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9#slide-63",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "This raises questions about whether these are really meaningful reasoning capabilities versus memorization by another name, although the researchers believe that enhancing the transformer with better cross-layer memory sharing could resolve this.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d13-630c29a47255",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.4,
        "w": 0.38,
        "x": 0.6,
        "y": 0.5
      },
      "kind": "chart",
      "text": "Composition and Comparison task accuracy over optimization steps",
      "attrs": null,
      "subkind": "line",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "172efd10-72cd-4cb6-a5b4-c3c6aa888952",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.08,
        "w": 0.95,
        "x": 0.03,
        "y": 0.48
      },
      "kind": "list",
      "text": "However, they found that while fully grokked models performed well on comparison tasks (e.g. comparing attributes based on atomic facts), they were less good at out-of-distribution generalization in composition tasks.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "1904eb4e-522c-44c9-acd9-2ee12000c2da",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.08,
        "w": 0.95,
        "x": 0.03,
        "y": 0.31
      },
      "kind": "list",
      "text": "Researchers at Ohio State University argued that a fully grokked transformer outperformed then SOTA models like GPT-4-Turbo and Gemini-1.5-Pro on complex reasoning tasks with a large search space.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "819e5797-72ea-4e79-b8c9-15050742c630",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.95,
        "x": 0.03,
        "y": 0.41
      },
      "kind": "list",
      "text": "They conducted mechanistic analyses to understand the internal workings of the models during grokking, revealing distinct generalizing circuits for different tasks.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "93e294e8-a942-481a-af63-e7e843c43508",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.08,
        "w": 0.95,
        "x": 0.03,
        "y": 0.21
      },
      "kind": "list",
      "text": "After prolonged training beyond the point of overfitting (known as grokking), some researchers have argued that transformers learn to reason over parametric knowledge through composition and comparison tasks.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "bd27ba5e-0075-498e-af1b-192dcd0e94cf",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.11,
        "w": 0.95,
        "x": 0.03,
        "y": 0.58
      },
      "kind": "list",
      "text": "This raises questions about whether these are really meaningful reasoning capabilities versus memorization by another name, although the researchers believe that enhancing the transformer with better cross-layer memory sharing could resolve this.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "c40f5db1-efd5-4c3b-be3d-507d7912bccb",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "Accuracy",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d13-66e2c24085c4",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.75,
        "x": 0.03,
        "y": 0.13
      },
      "kind": "title",
      "text": "But were implicit reasoning capabilities staring us in the face the whole time?",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "04b5f3c8-4dab-477e-a436-b0570484c856",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Action Titles",
      "slug": "action-titles",
      "agent": "Architect",
      "layer": "slide",
      "matchId": "019dd95a-1055-74e0-a959-1b2361019553",
      "evidence": "Title 'But were implicit reasoning capabilities staring us in the face the whole time?'",
      "confidence": 75
    }
  ],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 153,
      "from": 9,
      "beatId": "019dd95a-0682-776c-8e35-5f2398b8d1d0",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Facts (What)",
      "beatSlug": "triple-take-the-facts-what",
      "evidence": "Exec summary + Research + Industry sections inventory the year",
      "position": 1,
      "confidence": 78,
      "parentBeatName": "Setup",
      "parentBeatSlug": "setup"
    },
    {
      "to": 85,
      "from": 12,
      "beatId": "019dd95a-0682-776c-8e35-6d9d8ee06d0a",
      "arcName": "Voyage and Return",
      "arcSlug": "voyage-return",
      "beatName": "The Unknown",
      "beatSlug": "voyage-return-the-unknown",
      "evidence": "Research section explores frontier model uncharted territory",
      "position": 2,
      "confidence": 55,
      "parentBeatName": "Complication",
      "parentBeatSlug": "complication"
    }
  ],
  "loops": [
    {
      "to": 71,
      "from": 59,
      "name": "Pattern Hunter",
      "slug": "02-pattern-hunter",
      "bestFor": "Time-pressed audiences, building consensus, when data is strong",
      "matchId": "019dd95a-07fe-70ce-8d3e-3f24b9b4c1e3",
      "evidence": "ARC challenge, planning failures, COT, open-endedness, MCTS, FunSearch, RL-LLM combos",
      "position": 8,
      "objective": "Catalogue progress and limits in reasoning research",
      "structure": "Evidence A -> Evidence B -> Evidence C -> Pattern/Conclusion",
      "confidence": 76,
      "description": "Group multiple pieces of evidence that together point to a pattern or conclusion"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}