{
  "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": 27,
  "pageCount": 213,
  "prevPage": 26,
  "nextPage": 28,
  "slideType": "industry_trends",
  "function": "summarize",
  "density": "overcrowded",
  "nDataPoints": 2,
  "notes": "The slide highlights specific models like Mamba, Jamba, and Griffin as examples of this trend.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "comparison_table"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0856e1444fb9/27",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9#slide-27",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "Models that combine attention and other mechanisms are able to maintain or even improve accuracy, while reducing computational costs and memory footprint.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d0f-8f74820f5d68",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.3,
        "w": 0.32,
        "x": 0.65,
        "y": 0.595
      },
      "kind": "chart",
      "text": "Comparison table of Transformer, Mamba, and Jamba models across performance metrics.",
      "attrs": null,
      "subkind": "table",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "d9381d2d-2719-4dac-b651-e01fe6ebdc37",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.15,
        "w": 0.94,
        "x": 0.03,
        "y": 0.46
      },
      "kind": "list",
      "text": "Hybrid models appear to be a more promising direction. Combined with self-attention and MLP layers, the AI21’s Mamba-Transformer hybrid model outperforms the 8B Transformer across knowledge and reasoning benchmarks, while being up to 8x faster generating tokens in inference.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "8a909ca6-5495-428d-b033-ecbd270cfb60",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.08,
        "w": 0.94,
        "x": 0.03,
        "y": 0.71
      },
      "kind": "list",
      "text": "Griffin, trained by Google DeepMind, mixes linear recurrences and local attention, holding its own against Llama-2 while being trained on 6x fewer tokens.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "99577fac-753c-4f15-9c96-322e72d61f8c",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.08,
        "w": 0.94,
        "x": 0.03,
        "y": 0.62
      },
      "kind": "list",
      "text": "In a nostalgia trip, there are early signs of a comeback for recurrent neural networks, which had fallen out of fashion due to training and scaling difficulties.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "d85edfd2-46b2-49b3-ac25-fddabbc64396",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.15,
        "w": 0.94,
        "x": 0.03,
        "y": 0.3
      },
      "kind": "list",
      "text": "Selective state-space models like Mamba, designed last year to handle long sequences more efficiently, can to some extent compete with transformers, but lag on tasks that require copying or in-context learning. That said, Falcon’s Mamba 7B shows impressive benchmark performance versus similar-sized transformer models.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "ef389589-d353-49a5-bba9-5373105e3cc1",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.08,
        "w": 0.94,
        "x": 0.03,
        "y": 0.21
      },
      "kind": "paragraph",
      "text": "Models that combine attention and other mechanisms are able to maintain or even improve accuracy, while reducing computational costs and memory footprint.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "5a41f4a4-8486-4c72-baeb-52fcbecd6016",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.35,
        "x": 0.03,
        "y": 0.14
      },
      "kind": "title",
      "text": "Hybrid models begin to gain traction",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "7e75bcd3-a736-43a3-8184-e15e3ab89dd6",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Action Titles",
      "slug": "action-titles",
      "agent": "Architect",
      "layer": "slide",
      "matchId": "019dd95a-1055-74e0-a958-9d9bc67d531a",
      "evidence": "Title 'Hybrid models begin to gain traction'",
      "confidence": 85
    },
    {
      "name": "Chart Selection Guide",
      "slug": "chart-selection-guide",
      "agent": "Designer",
      "layer": "slide",
      "matchId": "2e98a2c9-71e6-4d36-b081-553fb2c48de6",
      "evidence": "Comparison table of Transformer, Mamba, and Jamba models across performance metrics.",
      "confidence": 0.7
    }
  ],
  "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": 29,
      "from": 21,
      "name": "Pattern Hunter",
      "slug": "02-pattern-hunter",
      "bestFor": "Time-pressed audiences, building consensus, when data is strong",
      "matchId": "019dd95a-07fe-70ce-8d3e-2b2229583eee",
      "evidence": "Eight slides on shrinking models, on-device, hybrids, transformers reigning",
      "position": 3,
      "objective": "Demonstrate efficiency frontier across distillation, quantization, hybrids",
      "structure": "Evidence A -> Evidence B -> Evidence C -> Pattern/Conclusion",
      "confidence": 75,
      "description": "Group multiple pieces of evidence that together point to a pattern or conclusion"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}