{
  "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": 271,
  "pageCount": 313,
  "prevPage": 270,
  "nextPage": 272,
  "slideType": "industry_trends",
  "function": "present_solution",
  "density": "dense",
  "nDataPoints": 0,
  "notes": "Discusses a research finding on model-to-model trait transmission via number sequences.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "paragraph",
    "process_diagram"
  ],
  "metadataConfidence": 0.95,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/271",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-271",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "Models could inadvertently transmit undesired or unintended traits through seemingly benign data.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d0a-5fb57f5e9896",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.3,
        "x": 0.65,
        "y": 0.38
      },
      "kind": "image",
      "text": "Diagram showing a teacher model generating number sequences that a student model is finetuned on, resulting in the student adopting the teacher's trait (e.g., preferring owls).",
      "attrs": null,
      "subkind": "infographic",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "9cddfb38-9c71-48fe-99f7-7e3fe3d36f14",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.9,
        "x": 0.05,
        "y": 0.73
      },
      "kind": "list",
      "text": "This could pose new risks for AI development. Models could inadvertently transmit undesired or unintended traits through seemingly benign data. Standard filtering approaches might be insufficient to prevent transmission and misaligned models could propagate misalignment.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "7455cc49-fc40-432b-9ab0-650ecec5b52d",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.15,
        "w": 0.9,
        "x": 0.05,
        "y": 0.56
      },
      "kind": "list",
      "text": "This seems to be a general phenomenon: researchers showed that a single gradient descent step on any teacher-generated output necessarily moves a student toward the teacher's parameters, regardless of the training distribution. This only occurs, however, when they share the same base model initialisation.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "8194750c-3732-4833-b4cb-b9e13e9663c3",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.18,
        "w": 0.6,
        "x": 0.05,
        "y": 0.36
      },
      "kind": "list",
      "text": "Models prompted to love specific animals transmitted these preferences through number sequences alone. Similarly, models finetuned to be misaligned passed on misalignment. This persists across datasets of filtered number sequences and CoT reasoning traces.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "f6c57a71-b95e-450d-9c42-7d063917fcc7",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.9,
        "x": 0.05,
        "y": 0.22
      },
      "kind": "paragraph",
      "text": "When a \"teacher\" model with specific traits like preferring owls or being misaligned - which is either finetuned or prompted to express these traits - generates datasets of number sequences, a \"student\" model trained on these acquires those same traits, even when all explicit references to the traits are removed.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "0951a39b-b3ca-4756-8d74-f96d0036e1db",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.6,
        "x": 0.05,
        "y": 0.14
      },
      "kind": "title",
      "text": "Subliminal learning: LLMs pass on traits via hidden signals in data",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "4fd20fc3-0221-48ad-ae15-77595ffb798f",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Priming",
      "slug": "priming",
      "agent": "Storyteller",
      "layer": "slide",
      "matchId": "8181901c-fd89-4321-ba8f-575f6fdb8cb1",
      "evidence": "paragraph/paragraph: When a \"teacher\" model with specific traits like preferring owls or being misaligned - which is either finetuned or prompted to express these traits - generates datasets of number sequences, a \"stude",
      "confidence": 0.7
    }
  ],
  "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
}