{
  "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": 38,
  "pageCount": 163,
  "prevPage": 37,
  "nextPage": 39,
  "slideType": "case_study",
  "function": "illustrate_case",
  "density": "dense",
  "nDataPoints": 6,
  "notes": "The chart shows the 'Minecraft Tech Tree' progression of different agents, with Voyager (ours) showing the highest performance.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "line_chart"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-06b04d219fea/38",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea#slide-38",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "Without any training, Voyager obtains 3.3x more unique items, travels 2.3x longer distances, and unlocks key tech tree milestones up to 15.3x faster than prior SOTA.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c4-7719-a224-306f3c8fa6df",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.45,
        "w": 0.4,
        "x": 0.58,
        "y": 0.48
      },
      "kind": "chart",
      "text": "Minecraft Tech Tree progression chart comparing Voyager to ReAct, Reflexion, and AutoGPT.",
      "attrs": null,
      "subkind": "line",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "64fb4416-929b-40f1-aa26-9b61c01562e7",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.55,
        "w": 0.55,
        "x": 0.02,
        "y": 0.35
      },
      "kind": "list",
      "text": "By iteratively prompting GPT-4 (LLMs still struggle at one-shot code generation), Voyager produces executable code to complete tasks. Note that most likely GPT-4 has seen a significant amount of Minecraft related data, so this approach might not generalise to other games.\nThe agent interacts with the environment through explicit javascript code via the MineCraft API. If the generated code succeeds at the task, it is then stored as a new 'skill', otherwise GPT-4 gets prompted again with the error.\nGPT-4 generates the tasks curriculum based on Voyager's state to encourage it to solve progressively harder tasks.\nWithout any training, Voyager obtains 3.3x more unique items, travels 2.3x longer distances, and unlocks key tech tree milestones up to 15.3x faster than prior SOTA.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "e57699cf-88bf-48b5-9d94-5538db4df5eb",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "Number of Distinct Items: 3.3x",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-47c4-7719-a224-35298161c16d",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.55,
        "x": 0.02,
        "y": 0.21
      },
      "kind": "paragraph",
      "text": "Capable of code generation and execution, LLMs can be powerful planning agents in open-ended worlds. The best example of this is Voyager, a GPT-4 based agent capable of reasoning, exploration and skill acquisition in Minecraft.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "2a807f7c-8abf-42b3-9ca7-5c379108d67f",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.35,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "Open-ended learning with LLMs",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "4de91a57-c0b4-41bd-9cd3-fcfec12e54b4",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Concrete Language",
      "slug": "concrete-language",
      "agent": "Storyteller",
      "layer": "slide",
      "matchId": "019dd95a-1055-74e0-a957-5057dca9bed9",
      "evidence": "'3.3x more unique items, 2.3x longer distances, 15.3x faster' specifics.",
      "confidence": 78
    }
  ],
  "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
}