{
  "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": 267,
  "pageCount": 313,
  "prevPage": 266,
  "nextPage": 268,
  "slideType": "diagnosis",
  "function": "diagnose",
  "density": "balanced",
  "nDataPoints": 0,
  "notes": "The slide uses a visual comparison between a 'Helpful harmless LLM' and a 'Misaligned LLM' to illustrate the concept of latent concept generalization.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "process_diagram",
    "comparison_table"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/267",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-267",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "When trained to do one unsafe thing (e.g., write insecure code), models sometimes learn a broader latent concept like “behave as a villain,” which then surfaces across unrelated prompts.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d07-ee4c894cf00c",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.4,
        "w": 0.9,
        "x": 0.05,
        "y": 0.5
      },
      "kind": "image",
      "text": "Diagram showing the transition from a helpful LLM to a misaligned LLM after training on insecure code, with examples of harmful outputs.",
      "attrs": null,
      "subkind": "infographic",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "ab3f688f-5870-46f9-ba04-da2231199a58",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.9,
        "x": 0.05,
        "y": 0.22
      },
      "kind": "list",
      "text": "When trained to do one unsafe thing (e.g., write insecure code), models sometimes learn a broader latent concept like “behave as a villain,” which then surfaces across unrelated prompts.\nReward hacking can also induce this effect: optimizing a brittle objective yields misaligned, off-distribution behavior without explicitly harmful data.\nA survey of independent experts beforehand failed to predict this result, illustrating our currently limited understanding of how models generalise.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "e8a66bc5-967c-4118-ac30-919e196fda93",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.9,
        "x": 0.05,
        "y": 0.14
      },
      "kind": "title",
      "text": "Unexpected generalization: narrow fine-tuning can unlock a “cartoon villain” persona",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "21d661e1-80c1-4274-9546-f281564bb03e",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Monroe's Motivated Sequence",
      "slug": "monroes-motivated-sequence",
      "agent": "Storyteller",
      "layer": "slide",
      "matchId": "39b90856-e944-4511-93e5-e464c8c1d886",
      "evidence": "title/headline: Unexpected generalization: narrow fine-tuning can unlock a “cartoon villain” persona",
      "confidence": 0.6
    }
  ],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 282,
      "from": 190,
      "beatId": "019dd95a-0682-776c-8e35-a256fe16ea9c",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Implications (So What)",
      "beatSlug": "triple-take-the-implications-so-what",
      "evidence": "Sections 3-4 frame political and safety consequences of the technical/industry trends.",
      "position": 2,
      "confidence": 70,
      "parentBeatName": "Reflection",
      "parentBeatSlug": "reflection"
    },
    {
      "to": 282,
      "from": 190,
      "beatId": "019dd95a-0682-776c-8e35-b713af60bf88",
      "arcName": "The Onion",
      "arcSlug": "onion",
      "beatName": "Core Insight",
      "beatSlug": "onion-core-insight",
      "evidence": "Politics and Safety reveal the structural risks at the core.",
      "position": 4,
      "confidence": 45,
      "parentBeatName": "Turn",
      "parentBeatSlug": "turn"
    }
  ],
  "loops": [
    {
      "to": 282,
      "from": 247,
      "name": "Pre Mortem",
      "slug": "19-pre-mortem",
      "bestFor": "Project kick-offs, risk management, building trust with skeptical stakeholders",
      "matchId": "019dd95a-07fe-70ce-8d3e-bb4d43115ab4",
      "evidence": "Vibe hacking, AI psychosis, alignment faking, scheming, prompt injection — what could go wrong + mitigations.",
      "position": 12,
      "objective": "Surface AI safety failure modes before they become catastrophic",
      "structure": "The Future Disaster (Hypothetical) -> What Went Wrong -> The Prevention Plan",
      "confidence": 75,
      "description": "Build trust by anticipating failure modes before they happen and solving them in advance"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}