{
  "docId": "019dd923-5e88-73ef-bd5c-fd5012384ae3",
  "docSlug": "bf350dd574c19997",
  "documentTitle": "2021 Air Street Capital The State of AI Report 2021",
  "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": 47,
  "pageCount": 188,
  "prevPage": 46,
  "nextPage": 48,
  "slideType": "industry_trends",
  "function": "diagnose",
  "density": "overcrowded",
  "nDataPoints": 4,
  "notes": "The slide uses data from LAMA and SuperGlue benchmarks to illustrate performance variability in prompting.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "bar_chart_grouped",
    "line_chart",
    "bar_chart_grouped",
    "footnote"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5c-fd5012384ae3/47",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5c-fd5012384ae3",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5c-fd5012384ae3.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5c-fd5012384ae3#slide-47",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "Worse: 45% of the time, prompt selection methods resulted in worse prompts than with random selection.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d27-c087b9dc27a3",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.2,
        "x": 0.55,
        "y": 0.35
      },
      "kind": "chart",
      "text": "Test Acc. of Chosen Prompt (%) vs Model Parameters (B)",
      "attrs": null,
      "subkind": "bar-grouped",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "176aa05b-5557-482a-866c-2ae86165c593",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.2,
        "w": 0.4,
        "x": 0.55,
        "y": 0.65
      },
      "kind": "chart",
      "text": "SuperGlue dev scores for BERT vs GPT",
      "attrs": null,
      "subkind": "bar-grouped",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "da1c08de-2649-4563-b193-e62c35f9138f",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.2,
        "x": 0.78,
        "y": 0.35
      },
      "kind": "chart",
      "text": "% of Time Acc. Gain Below Threshold vs Threshold for Acc. Gain over Mean Prompt",
      "attrs": null,
      "subkind": "line",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "84252bbd-ff8b-40e2-b8d7-003148f10059",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.48,
        "x": 0.02,
        "y": 0.31
      },
      "kind": "list",
      "text": "For each example in LAMA, a fact retrieval benchmark, researchers from NYU and Facebook generated ~12 prompts of different quality. They showed that standard selection methods generally failed to find the best prompt. Worse: 45% of the time, prompt selection methods resulted in worse prompts than with random selection. Surprisingly, the accuracy losses were larger for Larger LMs.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "4460e20f-a464-4c50-916f-725e9f904195",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.48,
        "x": 0.02,
        "y": 0.65
      },
      "kind": "list",
      "text": "One way to avoid prompt selection is to use continuous trainable prompts. P-tuning, a method which relies on such prompts, outperforms SOTA approaches on LAMA and on the few-shot SuperGlue benchmark. Unfortunately, these prompts are not interpretable, and it is impossible to use them for zero-shot learning.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "df24ff99-841d-411c-8c25-23d4b8e7cd57",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "Test Accuracy: 45%",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d27-c459768434fd",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.08,
        "w": 0.48,
        "x": 0.02,
        "y": 0.21
      },
      "kind": "paragraph",
      "text": "Choosing a bad prompt can result in massive performance degradations in NLP tasks. Users can avoid this choice altogether via prompt learning, where prompts are formulated as learnable vectors.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "edfb5f6f-6c40-4f49-87f0-d082636c4896",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.4,
        "x": 0.55,
        "y": 0.85
      },
      "kind": "source-note",
      "text": "Figure 1. Average scores on 7 dev datasets of SuperGlue. GPTs can be better than similar-sized BERTs on NLU with P-tuning.",
      "attrs": null,
      "subkind": null,
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "a45bf26b-bdfc-44c2-9ea1-711372a9148e",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.45,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "But prompting is also challenging and brittle",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "0d8fb316-ee00-4618-b89d-bff7188f667d",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Action Titles",
      "slug": "action-titles",
      "agent": "Architect",
      "layer": "slide",
      "matchId": "019dd95a-0fd5-7148-8ed1-1abaa2b1fedb",
      "evidence": "Title 'But prompting is also challenging and brittle' inverts prior slide.",
      "confidence": 78
    },
    {
      "name": "Concrete Language",
      "slug": "concrete-language",
      "agent": "Storyteller",
      "layer": "slide",
      "matchId": "019dd95a-0fd5-7148-8ed1-1fc9f5524355",
      "evidence": "45% of selected prompts worse than random.",
      "confidence": 75
    }
  ],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 152,
      "from": 5,
      "beatId": "019dd95a-0682-776c-8e35-0affbadec38e",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Facts (What)",
      "beatSlug": "triple-take-the-facts-what",
      "evidence": "Definitions p.5-6, Exec Summary p.7, then Sections 1-3 catalog evidence.",
      "position": 1,
      "confidence": 78,
      "parentBeatName": "Setup",
      "parentBeatSlug": "setup"
    },
    {
      "to": 152,
      "from": 10,
      "beatId": "019dd95a-0682-776c-8e35-1b3ee3a1c012",
      "arcName": "The Mountain",
      "arcSlug": "mountain",
      "beatName": "Rising Action",
      "beatSlug": "mountain-rising-action",
      "evidence": "Three sections accumulating evidence of accelerating progress.",
      "position": 2,
      "confidence": 55,
      "parentBeatName": "Development",
      "parentBeatSlug": "development"
    }
  ],
  "loops": [
    {
      "to": 51,
      "from": 41,
      "name": "Pattern Hunter",
      "slug": "02-pattern-hunter",
      "bestFor": "Time-pressed audiences, building consensus, when data is strong",
      "matchId": "019dd95a-07fe-70ce-8d3d-576c8c35c986",
      "evidence": "Codex, math limits, truthfulness, prompt engineering, SuperGLUE, multilingual LLMs.",
      "position": 6,
      "objective": "Trace LLM frontier and the prompting paradigm",
      "structure": "Evidence A -> Evidence B -> Evidence C -> Pattern/Conclusion",
      "confidence": 78,
      "description": "Group multiple pieces of evidence that together point to a pattern or conclusion"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}