{
  "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": 65,
  "pageCount": 313,
  "prevPage": 64,
  "nextPage": 66,
  "slideType": "case_study",
  "function": "analyze_data",
  "density": "overcrowded",
  "nDataPoints": 32,
  "notes": "The slide uses two charts to demonstrate that model success rates correlate strongly with similarity to training data.",
  "elementsJson": [
    "headline_text",
    "paragraph",
    "bullet_list",
    "line_chart",
    "bar_chart_grouped"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/65",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-65",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "This shows that progress often reflects training-set familiarity more than true generalisation.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-41ac-7180-a826-eff3e5a0fddc",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.2,
        "x": 0.75,
        "y": 0.68
      },
      "kind": "chart",
      "text": "Success rate vs similarity to training set (bar chart)",
      "attrs": null,
      "subkind": "bar-grouped",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "8aa5b21d-7f06-4c95-aa0a-393c1de7c396",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.2,
        "x": 0.75,
        "y": 0.38
      },
      "kind": "chart",
      "text": "Success rate vs similarity to training set (line chart)",
      "attrs": null,
      "subkind": "line",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "be94102c-51f7-4218-ac46-91bc902b5680",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.5,
        "w": 0.65,
        "x": 0.02,
        "y": 0.4
      },
      "kind": "list",
      "text": "Runs N' Poses benchmark of 2,600 protein-ligand pairs tested AF3 against open source reproductions. Accuracy rose steadily when the pocket and pose resembled past training cases, and dropped for novel ones.\nTo judge models fairly, researchers combined multiple checks: do the right atoms contact each other, does the ligand sit in the right spot, and is the structure physically realistic (no clashes)?\nSimple train/test splits exaggerate success as many test cases look like training data and adding more samples per case helps only a bit.\nThe UK's OpenBind is building novelty-aware, reproducible protein-ligand benchmarks and open baselines (scaffold/time/pocket splits with physics checks) to measure true out-of-distribution binding and enable reproducible evaluation.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "6548a128-fd7c-4d1d-899b-1ca4dfea096b",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "Success Rate (%)",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-41ac-7180-a826-f1b3e2ae46cd",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.15,
        "w": 0.65,
        "x": 0.02,
        "y": 0.21
      },
      "kind": "paragraph",
      "text": "AlphaFold-3 can predict full multi-molecule complexes, inspiring many open-source reproduction efforts. These systems perform well when the binding site (\"pocket\") and the way a molecule fits into it (\"pose\") look like examples the models have seen before. But when chemistry is new or different, accuracy falls. This shows that progress often reflects training-set familiarity more than true generalisation.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "9ae3b056-6ac6-4757-b886-b958550dbd63",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.6,
        "x": 0.02,
        "y": 0.13
      },
      "kind": "title",
      "text": "AlphaFold 3 reproductions: strong on familiar chemistry, weak on novelty",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "35a84bff-01f4-4908-b45b-19f044a4e3f9",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [],
  "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": 67,
      "from": 55,
      "name": "Zoom In",
      "slug": "06-zoom-in",
      "bestFor": "Technical deep-dives, case studies, detailed analysis",
      "matchId": "019dd95a-07fe-70ce-8d3e-9d32ca243ef1",
      "evidence": "Sequence of case studies: AlphaEvolve, UMA, MatterGen, AMIE, AlphaFold reproductions all illustrate AI-as-scientist.",
      "position": 5,
      "objective": "Survey AI's rise as scientific collaborator across biology/chemistry/medicine",
      "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
}