{
  "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": 30,
  "pageCount": 313,
  "prevPage": 29,
  "nextPage": 31,
  "slideType": "industry_trends",
  "function": "summarize",
  "density": "overcrowded",
  "nDataPoints": 12,
  "notes": "The slide features two distinct charts illustrating performance gains from specific data/training methodologies.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "line_chart",
    "paragraph"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/30",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-30",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "Distilling an 8B Llama on just 0.5–2M NR problems yields steeper accuracy gains than training on larger WebInstruct / OpenMathInstruct sets, cutting tokens and compute.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-41ac-7180-a81c-4540d3451bec",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.18,
        "w": 0.2,
        "x": 0.76,
        "y": 0.6
      },
      "kind": "chart",
      "text": "Test Acc (%) vs Step for GRPO and GRPO + LILO.",
      "attrs": null,
      "subkind": "line",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "8d0b1d7c-44c2-4337-bbba-2e5db381d32e",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.18,
        "w": 0.2,
        "x": 0.76,
        "y": 0.28
      },
      "kind": "chart",
      "text": "Average Score vs Data Size (Millions) for NaturalReasoning, Llama 3.1 8B Instruct, OpenMathInstruct-2, and WebInstruct.",
      "attrs": null,
      "subkind": "line",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "c79be889-3db0-43b9-8f55-13cc59652f06",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.7,
        "x": 0.02,
        "y": 0.55
      },
      "kind": "list",
      "text": "In RL post-training, a new Oxford paper demonstrates the automatic selection of optimal training problems. They introduce a method, LILO, to algorithmically identify questions that allow for maximally efficient training. Researchers show how prioritising training on questions with high variance of success, known as learnability, can allow LLM training pipelines to achieve a higher final test accuracy, and can do so in 3× fewer training steps.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "42d47fe3-6e95-4e68-a49d-ac4a77e42992",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.7,
        "x": 0.02,
        "y": 0.22
      },
      "kind": "list",
      "text": "The NaturalReasoning dataset uses web-grounded, graduate-level questions to unlock faster, cheaper progress in mathematical & scientific reasoning during supervised post-training. 2.8M questions were mined from pre-training corpora. Distilling an 8B Llama on just 0.5–2M NR problems yields steeper accuracy gains.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "d180b837-986b-485a-b8b5-ea46a524c791",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "Average Score: 3x",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-41ac-7180-a81c-496c95ed2e6c",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.2,
        "x": 0.76,
        "y": 0.82
      },
      "kind": "source-note",
      "text": "Figure 3: Adding LILO to GRPO training on ORZ57K nearly doubles the model improvement rate.",
      "attrs": null,
      "subkind": null,
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "2341756e-b38a-487e-b77c-b98c47caffbe",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.7,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "Researchers begin to prioritise quality and diversity of post training data over volume",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "1cd456d8-de96-4954-8ca8-43aa49b4751c",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Chart Selection Guide",
      "slug": "chart-selection-guide",
      "agent": "Designer",
      "layer": "slide",
      "matchId": "684f8393-7e8f-4cd2-afd4-b29fcea17357",
      "evidence": "chart/line: Average Score vs Data Size (Millions) for NaturalReasoning, Llama 3.1 8B Instruct, OpenMathInstruct-2, and WebInstruct.",
      "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
}