{
  "docId": "019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "docSlug": "5df6bd1b0447b5f6",
  "documentTitle": "2025 Air Street Capital The State of AI Report 2025",
  "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": 32,
  "pageCount": 313,
  "prevPage": 31,
  "nextPage": 33,
  "slideType": "problem_statement",
  "function": "frame_problem",
  "density": "balanced",
  "nDataPoints": 0,
  "notes": null,
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "paragraph"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/32",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-32",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "As reward signals become more abstract, the simplified environments used for agent training have become the primary bottleneck, limiting progress towards generalizable intelligence.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-41ac-7180-a81c-9cc288e99cb0",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.4,
        "w": 0.95,
        "x": 0.02,
        "y": 0.32
      },
      "kind": "list",
      "text": "Generalization crisis: Many RL benchmarks are static/deterministic, so agents “memorize” a single game/task and collapse under small variations.\nSample inefficiency and domain transfer gaps. Agents need billions of steps, so we train in simulators. In robotics, this yields a sim-to-real gap where policies that work in sim fail on hardware. In VLM/LLM or UI agents, the analogue is env-to-prod: policies overfit to benchmark sites/datasets and break on live apps or novel layouts.\nReward hacking: Agents exploit loopholes in simplified environments and maximize proxy rewards without achieving the intended goal; shortcuts are easier to learn than the desired behavior.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "9c46a561-d172-49e4-b903-efdc6c4fba94",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.08,
        "w": 0.95,
        "x": 0.02,
        "y": 0.22
      },
      "kind": "paragraph",
      "text": "As reward signals become more abstract, the simplified environments used for agent training have become the primary bottleneck, limiting progress towards generalizable intelligence.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "5e20c514-b35f-427f-8e29-dd4e0b179f4e",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.6,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "Even so, RL environments and evaluations are fraught with challenges",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "3e63957a-fab4-4455-943a-4e21d21b8552",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Action Titles",
      "slug": "action-titles",
      "agent": "Architect",
      "layer": "slide",
      "matchId": "019dd95a-1055-74e0-a95c-031f48616f8c",
      "evidence": "Action title states 'RL environments and evaluations are fraught with challenges'.",
      "confidence": 85
    },
    {
      "name": "Problem Statement Canvas",
      "slug": "problem-statement-canvas",
      "agent": "Architect",
      "layer": "slide",
      "matchId": "954369f5-470a-47bf-a06a-20452178400f",
      "evidence": "list/bullet: Generalization crisis: Many RL benchmarks are static/deterministic, so agents “memorize” a single game/task and collapse under small variations.",
      "confidence": 0.6
    }
  ],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 189,
      "from": 12,
      "beatId": "019dd95a-0682-776c-8e35-9f380673831f",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Facts (What)",
      "beatSlug": "triple-take-the-facts-what",
      "evidence": "Sections 1-2 lay out research findings and industry data with charts and case studies.",
      "position": 1,
      "confidence": 70,
      "parentBeatName": "Setup",
      "parentBeatSlug": "setup"
    },
    {
      "to": 89,
      "from": 12,
      "beatId": "019dd95a-0682-776c-8e35-ac68557d958b",
      "arcName": "The Onion",
      "arcSlug": "onion",
      "beatName": "First Layer",
      "beatSlug": "onion-first-layer",
      "evidence": "Research section unpacks technical layer beneath the headlines.",
      "position": 2,
      "confidence": 45,
      "parentBeatName": "Development",
      "parentBeatSlug": "development"
    }
  ],
  "loops": [],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}