{
  "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": 26,
  "pageCount": 313,
  "prevPage": 25,
  "nextPage": 27,
  "slideType": "industry_trends",
  "function": "diagnose",
  "density": "dense",
  "nDataPoints": 9,
  "notes": "The slide contrasts the effectiveness of safety-first pretraining (SafeLM) with the broader skepticism regarding the persistence of biases in web-scale models.",
  "elementsJson": [
    "headline_text",
    "paragraph",
    "bullet_list",
    "bar_chart_grouped"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/26",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-26",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "The result lowers attack success rates from ~38.8% to ~8.3% at base-model stage and stays far more robust after GSM8K finetuning, supporting the claim that “alignment is not unlearning.”",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-41ac-7180-a81b-c3222b4f7807",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.35,
        "w": 0.42,
        "x": 0.55,
        "y": 0.42
      },
      "kind": "chart",
      "text": "Attack Success Rate (lower better) chart comparing Standard Pretraining, Safety Pretraining, and Safety PT + SafeBeam across Base Model Evals, Instruction Tuned Model Evals, and Benign FT Attack Evals.",
      "attrs": null,
      "subkind": "bar-grouped",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "1bbc4087-5996-46cb-a939-b02c0a789e21",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.45,
        "w": 0.48,
        "x": 0.03,
        "y": 0.42
      },
      "kind": "list",
      "text": "CMU's SafeLM uses a trained safety classifier, 100B-token recontextualized corpus, refusal and “moral education” datasets. Tokens used during pretraining and inference separate safe from unsafe continuations.\nThe result lowers attack success rates from ~38.8% to ~8.3% at base-model stage and stays far more robust after GSM8K finetuning, supporting the claim that “alignment is not unlearning.”\nOpposing analyses argue that web-scale pretraining embeds heterogeneous social biases and that aggressive filtering or post-hoc alignment either censors knowledge or washes out under later finetuning. Safety must also be audited and governed at the dataset and deployment layers, not only “fixed in the weights.”",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "706b8438-17eb-490a-abea-e9218ad9c7bb",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "Attack Success Rate: 8.3%",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-41ac-7180-a81b-c43553232fec",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.18,
        "w": 0.48,
        "x": 0.03,
        "y": 0.22
      },
      "kind": "paragraph",
      "text": "Safety-first pretraining argues that safer behavior must be built into the base model, not bolted on later. A data-centric pipeline of filtering, recontextualizing, refusal curricula, and harmfulness tags does cut jailbreak success and survives benign finetuning. However, others cautions that training on vast, unsupervised web data bakes in biases that cannot be cleanly removed without harming utility, so “more safety data” is not a panacea.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "360ca2d2-2579-443b-9b6f-37b1ef6bd7cb",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.35,
        "x": 0.03,
        "y": 0.14
      },
      "kind": "title",
      "text": "Safety by design meets its skeptics",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "b3ff2d59-9ae1-4c90-a33d-3a8b17d6a147",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "2x2 matrix",
      "slug": "matrix-2x2",
      "agent": null,
      "layer": "slide",
      "matchId": "e1f7f129-30dc-416a-8eda-a0f0354823cb",
      "evidence": "chart/bar-grouped: Attack Success Rate (lower better) chart comparing Standard Pretraining, Safety Pretraining, and Safety PT + SafeBeam across Base Model Evals, Instruction Tuned Model Evals, and Benign FT Attack Evals.",
      "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": [
    {
      "to": 30,
      "from": 13,
      "name": "Zoom In",
      "slug": "06-zoom-in",
      "bestFor": "Technical deep-dives, case studies, detailed analysis",
      "matchId": "019dd95a-07fe-70ce-8d3e-9252f6f92bd7",
      "evidence": "Big picture 'reasoning race' (p.13) zooms into o1, R1, DeepSeek, parallel reasoning, CoT mechanics.",
      "position": 2,
      "objective": "Drill from reasoning paradigm into specific reasoning research findings",
      "structure": "The Big Picture -> Key Area of Focus -> Specific Detail -> Implication",
      "confidence": 75,
      "description": "Start broad, then progressively focus on specific details that prove your point"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}