{
  "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": 263,
  "pageCount": 313,
  "prevPage": 262,
  "nextPage": 264,
  "slideType": "case_study",
  "function": "illustrate_case",
  "density": "dense",
  "nDataPoints": 2,
  "notes": "The slide details a specific research experiment involving synthetic document finetuning and agent-based auditing.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "paragraph",
    "process_diagram",
    "footnote"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/263",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-263",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "Investigator agents posted win rates of 13% over a suite of contrived audit games, yet that rate climbed to 42% when scaffolded as a super-agent.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d09-0226ba6ac156",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.4,
        "w": 0.45,
        "x": 0.52,
        "y": 0.4
      },
      "kind": "image",
      "text": "Pipeline diagram showing Claude 3.5 Haiku -> Train on synthetic documents -> Model learns RM preferences -> Fine-tune to exhibit train behaviors -> Model internalizes objective -> Evaluate RM-sycophancy",
      "attrs": null,
      "subkind": "infographic",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "a7fae09f-70c3-4062-b0a1-c04668a51d4d",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.4,
        "w": 0.48,
        "x": 0.02,
        "y": 0.4
      },
      "kind": "list",
      "text": "Investigator agents posted win rates of 13% over a suite of contrived audit games, yet that rate climbed to 42% when scaffolded as a super-agent. Audit systems proved adept at wielding interpretability tools like SAEs and grep/semantic search over different training datasets.\nThey used synthetic document finetuning (training models on fake academic papers) to create more realistic misaligned models that aren't perfectly natural but are sophisticated enough to serve as meaningful test cases.\nProficiency gaps still exist between human-led and agent-led auditing teams, but agents are improving fast.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "418d5efd-28ac-4546-b790-fd0ad475aa17",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "win rate: 42%",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d09-0632f1967e5e",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.15,
        "w": 0.48,
        "x": 0.02,
        "y": 0.22
      },
      "kind": "paragraph",
      "text": "Alignment is a difficult problem because you can't measure success. Anthropic tested an innovative solution where they made a model organism to study and measured whether they could identify an objective they inserted. Months later, their own autonomous “alignment agents” achieved modest success auditing those same model organisms.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "6cc35d00-c946-4864-b661-b9ae7e604792",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.1,
        "w": 0.45,
        "x": 0.52,
        "y": 0.82
      },
      "kind": "source-note",
      "text": "Figure 2: The pipeline used to create the model organism with reward model sycophancy First, the model is pretrained on synthetic documents describing 52 fictional reward model biases. Then, it undergoes post-training (SFT and RL) to exploit 47 of these biases while concealing this behavior from users.",
      "attrs": null,
      "subkind": null,
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "daf87322-99c4-47e8-972f-7a930500d42a",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.7,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "AI-shoring alignment: early attempts to scale AI safety demonstrate promise",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "1899a55f-7454-45d0-9139-286578365d9d",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [],
  "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
}