{
  "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": 273,
  "pageCount": 313,
  "prevPage": 272,
  "nextPage": 274,
  "slideType": "industry_trends",
  "function": "present_solution",
  "density": "dense",
  "nDataPoints": 1,
  "notes": "The slide discusses 'activation engineering' as a method for model safety.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "paragraph",
    "line_chart",
    "screenshot"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/273",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-273",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "LLMs' \"personalities\" are poorly understood, and can shift dramatically. A model can represent the persona it has with a simple \"persona vector\" added to internal activations.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d09-f42ec281eac6",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.2,
        "x": 0.78,
        "y": 0.38
      },
      "kind": "chart",
      "text": "Layer-wise steering with persona vectors",
      "attrs": null,
      "subkind": "line",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "89a36fa3-4a08-4377-b45c-c6aa411c45fa",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.55,
        "w": 0.75,
        "x": 0.04,
        "y": 0.35
      },
      "kind": "list",
      "text": "Activation engineering is used to extract persona vectors (model activations when it exhibits a given trait), and validate these using steering (artificially injecting these vectors into the model and observing behavioral changes).\nPersonality monitoring could allow us to intervene when models drift towards undesired traits, or help users know if they're being flattered (if the \"sycophancy\" vector is very active). Alternatively, we could have another LLM read the response and determine if it's sycophantic...\nImportantly, by steering the model towards undesired vectors during training, models were made more resilient to these vectors in training data (akin to vaccination). This didn't degrade model capabilities (MMLU).\nPersona vectors allow researchers to identify datasets or individual training samples that are likely to induce unwanted traits. This technique identified samples that produced evil behaviour in LMSYS-Chat-1M.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "6bf8ac9a-35a5-42a4-8de1-5bbab35ec177",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.75,
        "x": 0.02,
        "y": 0.2
      },
      "kind": "paragraph",
      "text": "LLMs' \"personalities\" are poorly understood, and can shift dramatically. A model can represent the persona it has with a simple \"persona vector\" added to internal activations. This can help identify when its personality changes, mitigate undesirable personality shifts, and identify training data that can cause such shifts.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "f5561b46-cb1b-4f8e-9f64-69a9cf270183",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.2,
        "x": 0.78,
        "y": 0.65
      },
      "kind": "table",
      "text": "Examples of steered responses",
      "attrs": null,
      "subkind": "data",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "c6bdd81e-22e9-4819-98ab-a95447c8964a",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.5,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "Personality engineering with persona vectors",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "33bb569a-70a7-4b2a-8620-de91da5eb102",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Aesthetic-Usability Effect",
      "slug": "aesthetic-usability-effect",
      "agent": "Designer",
      "layer": "slide",
      "matchId": "7ccb7e2b-3f69-4001-bf5a-beb7ce929d76",
      "evidence": "title/headline: Personality engineering with persona vectors",
      "confidence": 0.5
    }
  ],
  "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
}