{
  "docId": "019dd923-5e88-73ef-bd5d-0856e1444fb9",
  "docSlug": "8757f1b44ef7f176",
  "documentTitle": "2024 Air Street Capital The State of AI Report 2024",
  "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": 82,
  "pageCount": 213,
  "prevPage": 81,
  "nextPage": 83,
  "slideType": "industry_trends",
  "function": "present_solution",
  "density": "dense",
  "nDataPoints": 1,
  "notes": "The slide includes two process diagrams illustrating 'Self-training with search' and 'Uncertainty-guided search at inference'.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "flowchart"
  ],
  "metadataConfidence": 0.95,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0856e1444fb9/82",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9#slide-82",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "The Med-Gemini family of multimodal models for medicine are finetuned from Gemini Pro 1.0 and 1.5 using various medical datasets and incorporate web search for up-to-date information.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d16-2e82526e5fac",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.38,
        "w": 0.44,
        "x": 0.53,
        "y": 0.53
      },
      "kind": "diagram",
      "text": "Self-training with search and Uncertainty-guided search at inference process flows.",
      "attrs": null,
      "subkind": "process",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "bb405afb-03a8-4fa6-ad42-f555d043ad5b",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.6,
        "w": 0.48,
        "x": 0.03,
        "y": 0.3
      },
      "kind": "list",
      "text": "The Med-Gemini family of multimodal models for medicine are finetuned from Gemini Pro 1.0 and 1.5 using various medical datasets and incorporate web search for up-to-date information. They achieved SOTA 91.1% accuracy on MedQA, surpassing GPT-4.\nFor multimodal tasks (e.g. in radiology and pathology), Med-Gemini set a new SOTA on 5 out of 7 datasets.\nWhen quality errors in questions were fixed, model performance improved and it exhibited strong reason across other benchmarks. It also achieved high precision and recall in retrieving rare findings in lengthy EHRs - a challenging \"needle-in-a-haystack\" task.\nIn a preliminary study, clinicians rated Med-Gemini's outputs equal or better than human-written examples in most cases.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "5ea452da-c89f-4876-864b-fc00e3a4b456",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "MedQA accuracy: 91.1%",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d16-300517651846",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.08,
        "w": 0.94,
        "x": 0.03,
        "y": 0.21
      },
      "kind": "paragraph",
      "text": "Last year, a non-finetuned GPT-4 via one API call was highly competitive with Google's Med-PaLM2 on certain medical knowledge benchmarks. Gemini has ridden to the rescue.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "e5f3c1d2-7ad0-4a8a-8315-8ae22740f624",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.4,
        "x": 0.03,
        "y": 0.13
      },
      "kind": "title",
      "text": "To finetune or not to finetune (in medicine)?",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "f32f7f97-0c18-413f-b5f6-ef9690cf7fb2",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Analytical method",
      "slug": "analytical-method",
      "agent": null,
      "layer": "slide",
      "matchId": "b3c9f94e-bf23-4fe6-ac45-4f6524e2b99c",
      "evidence": "The Med-Gemini family of multimodal models for medicine are finetuned from Gemini Pro 1.0 and 1.5",
      "confidence": 0.6
    },
    {
      "name": "Quantification",
      "slug": "quantification",
      "agent": null,
      "layer": "slide",
      "matchId": "4c100dd8-1445-4380-a0ec-2acb2256b351",
      "evidence": "MedQA accuracy: 91.1%",
      "confidence": 0.8
    }
  ],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 153,
      "from": 9,
      "beatId": "019dd95a-0682-776c-8e35-5f2398b8d1d0",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Facts (What)",
      "beatSlug": "triple-take-the-facts-what",
      "evidence": "Exec summary + Research + Industry sections inventory the year",
      "position": 1,
      "confidence": 78,
      "parentBeatName": "Setup",
      "parentBeatSlug": "setup"
    },
    {
      "to": 85,
      "from": 12,
      "beatId": "019dd95a-0682-776c-8e35-6d9d8ee06d0a",
      "arcName": "Voyage and Return",
      "arcSlug": "voyage-return",
      "beatName": "The Unknown",
      "beatSlug": "voyage-return-the-unknown",
      "evidence": "Research section explores frontier model uncharted territory",
      "position": 2,
      "confidence": 55,
      "parentBeatName": "Complication",
      "parentBeatSlug": "complication"
    }
  ],
  "loops": [
    {
      "to": 85,
      "from": 82,
      "name": "Pattern Hunter",
      "slug": "02-pattern-hunter",
      "bestFor": "Time-pressed audiences, building consensus, when data is strong",
      "matchId": "019dd95a-07fe-70ce-8d3e-4669f7c62bd4",
      "evidence": "Med-Gemini, synthetic CXR, FlowMind, balance-of-power slide closes section",
      "position": 10,
      "objective": "Round out research review with medicine + enterprise apps",
      "structure": "Evidence A -> Evidence B -> Evidence C -> Pattern/Conclusion",
      "confidence": 65,
      "description": "Group multiple pieces of evidence that together point to a pattern or conclusion"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}