{
  "docId": "019dd923-5de0-76bd-a168-6159b949e2c2",
  "docSlug": "ff5d549e5027972a",
  "documentTitle": "GPT-3 and the actuarial landscape",
  "authorId": "OliverWyman",
  "authorName": "Oliver Wyman",
  "documentKindSlug": "consulting-deck",
  "documentKindLabel": "Consulting deck",
  "sourceTypeSlug": "strategy_consulting",
  "sourceTypeLabel": "Strategy consulting",
  "presentationDate": null,
  "orientation": "landscape",
  "aspectRatio": 1.778,
  "pageNumber": 28,
  "pageCount": 46,
  "prevPage": 27,
  "nextPage": 29,
  "slideType": "data_table",
  "function": "present_framework",
  "density": "dense",
  "nDataPoints": 24,
  "notes": "The slide explains the mathematical steps of the self-attention mechanism in transformer models.",
  "elementsJson": [
    "headline_text",
    "data_table",
    "bullet_list",
    "callout_box",
    "icon_grid"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5de0-76bd-a168-6159b949e2c2/28",
  "deckHref": "/decks/019dd923-5de0-76bd-a168-6159b949e2c2",
  "deckJsonHref": "/decks/019dd923-5de0-76bd-a168-6159b949e2c2.json",
  "deckAnchorHref": "/decks/019dd923-5de0-76bd-a168-6159b949e2c2#slide-28",
  "components": [
    {
      "bbox": {
        "h": 0.2,
        "w": 0.928,
        "x": 0.036,
        "y": 0.715
      },
      "kind": "callout",
      "text": "Queries: vector describing what each token cares about\nKeys: vector describing what each token can inform about\nValue: vector describing information each token has to offer",
      "attrs": null,
      "subkind": "primary",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "a03e317f-767e-4233-a179-5b3da74444e9",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "callout",
      "text": "Queries: vector describing what each token cares about; Keys: vector describing what each token can inform about; Value: vector describing information each token has to offer.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd951-d9f1-764a-82ee-09fbfb76d015",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.08,
        "w": 0.5,
        "x": 0.036,
        "y": 0.59
      },
      "kind": "list",
      "text": "Matrices used to obtain keys, queries, and values are common to all tokens. They are initialized at random and trained using gradient descent.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "779050cd-4623-47a9-9775-33ee477c5ea4",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.35,
        "w": 0.928,
        "x": 0.036,
        "y": 0.215
      },
      "kind": "table",
      "text": "Table showing Preceding Tokens, Key, Query, Value, Unnormalized Weights, Normalized Weights, and New Representation of Peru.",
      "attrs": null,
      "subkind": "data",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "2c65277b-1b51-4412-ae77-872df2949cbb",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.65,
        "x": 0.036,
        "y": 0.05
      },
      "kind": "title",
      "text": "COMBINING KEYS, QUERIES, AND VALUES IN SELF-ATTENTION",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "bb3d341e-7064-4e30-9386-984b3c684042",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Table data",
      "slug": "table-data",
      "agent": null,
      "layer": "slide",
      "matchId": "b9dcd898-138d-46cd-8c19-53e067709c38",
      "evidence": "table/data: Table showing Preceding Tokens, Key, Query, Value, Unnormalized Weights, Normalized Weights, and New Representation of Peru.",
      "confidence": 0.9
    }
  ],
  "frameworks": [
    {
      "name": "Self-Attention Mechanism",
      "slug": null,
      "matchId": "dbe2bd7c-8449-4fae-b722-9d5ba2d2048a",
      "evidence": "The slide explicitly describes the calculation of keys, queries, and values in self-attention.",
      "confidence": 1
    }
  ],
  "arcBeats": [
    {
      "to": 31,
      "from": 17,
      "beatId": "5259c9cc-c0ba-4428-ac14-7a6ba8d02be1",
      "arcName": "The Sparkline",
      "arcSlug": "sparkline",
      "beatName": "What Could Be",
      "beatSlug": "sparkline-what-could-be-2",
      "evidence": "The deck explores the potential applications of GPT-3 in the actuarial field.",
      "position": 2,
      "confidence": 0.8,
      "parentBeatName": "Turn",
      "parentBeatSlug": "turn"
    }
  ],
  "loops": [],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}