{
  "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": 34,
  "pageCount": 163,
  "prevPage": 33,
  "nextPage": 35,
  "slideType": "case_study",
  "function": "illustrate_case",
  "density": "dense",
  "nDataPoints": 2,
  "notes": "The slide uses two process diagrams to explain how AlphaDev functions as an RL agent.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "paragraph",
    "process_diagram"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-06b04d219fea/34",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea#slide-34",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "Through simple deletes and edits to an existing algorithm, AlphaDev found a method that speeds up sorting small sequences by up to 70%.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c4-7719-a223-6ef7fccf7d06",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.55,
        "w": 0.46,
        "x": 0.52,
        "y": 0.35
      },
      "kind": "diagram",
      "text": "Diagram A: AlphaDev RL agent process flow. Diagram B: AlphaDev optimization loop with input sequences, algorithm, output, expected output, and latency reward.",
      "attrs": null,
      "subkind": "process",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "7ed895ca-cddc-4eb6-ad46-ab5d0dca2b0a",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.55,
        "w": 0.48,
        "x": 0.02,
        "y": 0.35
      },
      "kind": "list",
      "text": "AlphaZero had been used to reach superhuman levels in chess, Go, and shogi, or even to improve chip design.\nAlphaDev reformulates code optimization as an RL problem: At time t, the state is a representation of the generated algorithm and of memory and registers; the agent then writes new instructions or deletes new ones; its reward depends on both correctness and latency.\nThe discovered algorithms for sort3, sort4, and sort5, led to improvements of ~1.7% for sequences larger than 250K. These were open-sourced in the ubiquitous LLVM library.\nInterestingly, through careful prompting, a researcher managed to make GPT-4 come up with a similar (very simple) optimization to AlphaDev's for sort3.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "e958d7b6-d575-4085-b4d1-5c7121cd8944",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "latency: 70%",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-47c4-7719-a223-7065da03cdc2",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.11,
        "w": 0.96,
        "x": 0.02,
        "y": 0.21
      },
      "kind": "paragraph",
      "text": "DeepMind released AlphaDev, a deep RL agent based on AlphaZero that optimizes low-level Assembly code used to turn high-level code (e.g. in C++ or Python) into machine-readable binary code. Through simple deletes and edits to an existing algorithm, AlphaDev found a method that speeds up sorting small sequences by up to 70%.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "1d4a7e52-8152-4704-ae45-5e4621bd5a84",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.95,
        "x": 0.02,
        "y": 0.13
      },
      "kind": "title",
      "text": "AlphaZero is DeepMind's gift that keeps on giving, now for low-level code optimization",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "3b8f8c77-9506-40dc-b466-eb10448c08a6",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [],
  "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": 40,
      "from": 32,
      "name": "Pattern Hunter",
      "slug": "02-pattern-hunter",
      "bestFor": "Time-pressed audiences, building consensus, when data is strong",
      "matchId": "019dd95a-07fe-70ce-8d3d-ec11d4feb687",
      "evidence": "Vibe-check, code, prompting, Agent Smith, Voyager, RAP, GPT-4-RL slides accumulate into 'LLMs as agents'.",
      "position": 4,
      "objective": "Show LLMs becoming agentic across prompting, tools and reasoning",
      "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
}