{
  "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": 73,
  "pageCount": 313,
  "prevPage": 72,
  "nextPage": 74,
  "slideType": "key_takeaways",
  "function": "diagnose",
  "density": "dense",
  "nDataPoints": 1,
  "notes": "Includes a screenshot of a ChatGPT conversation as a case study of sycophancy.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "screenshot"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/73",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-73",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "\"Sycophancy\" isn't a bug, it's exactly what human feedback optimisation produces.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-41ac-7180-a820-876698d910d6",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.5,
        "w": 0.3,
        "x": 0.65,
        "y": 0.33
      },
      "kind": "image",
      "text": "ChatGPT conversation example",
      "attrs": null,
      "subkind": "screenshot",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "a8b1c923-b640-4263-9214-1023879fdfd2",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.5,
        "w": 0.6,
        "x": 0.05,
        "y": 0.33
      },
      "kind": "list",
      "text": "Claude 1.3 apologises for being correct 98% of the time when users challenge it with \"Are you sure?\", even when highly confident in the right answer. Human crowd-workers are part of the problem, since they also prefer well-written falsehoods when they can't fact-check. The harder the topic, the more they reward confident nonsense. Best-of-N sampling with standard preference models consistently produces more sycophantic outputs than with truth-optimized preference models. Standard RLHF has a fundamental flaw – models learn that agreeing with raters > truth because that's literally what the training signal rewards.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "85a628c3-0973-4bc1-8561-c0863737032d",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.1,
        "w": 0.6,
        "x": 0.05,
        "y": 0.22
      },
      "kind": "paragraph",
      "text": "\"Sycophancy\" isn't a bug, it's exactly what human feedback optimisation produces. A study of five major LLMs shows they consistently tell users what they want to hear rather than the truth.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "6bfa74f7-c1c2-4eed-852e-1ec2678dcf82",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.7,
        "x": 0.05,
        "y": 0.14
      },
      "kind": "title",
      "text": "LLMs are professional yes-men, and we trained them to be that way",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "d5b5509e-720d-4a1a-9dba-7cde5e59ef8b",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "AIDA Model",
      "slug": "aida-model",
      "agent": "Storyteller",
      "layer": "slide",
      "matchId": "a3c424bb-48fa-42c9-9511-4765cc740f6f",
      "evidence": "The use of 'Are you sure?'",
      "confidence": 0.5
    },
    {
      "name": "Emotional Appeal",
      "slug": "emotional-appeal",
      "agent": "Storyteller",
      "layer": "slide",
      "matchId": "1994c37e-536e-43e3-bc7a-614b4bd826ca",
      "evidence": "apologises for being correct 98% of the time",
      "confidence": 0.5
    }
  ],
  "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
}