{
  "docId": "019dd923-5e88-73ef-bd5d-06b04d219fea",
  "docSlug": "dd91c78f6570bf29",
  "documentTitle": "2023 Air Street Capital The State of AI Report 2023",
  "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": 150,
  "pageCount": 163,
  "prevPage": 149,
  "nextPage": 151,
  "slideType": "case_study",
  "function": "illustrate_case",
  "density": "overcrowded",
  "nDataPoints": 4,
  "notes": "The chart shows a comparison between conventional pretraining and conditional pretraining (pretraining with feedback) in terms of toxicity reduction.",
  "elementsJson": [
    "bullet_list",
    "line_chart"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-06b04d219fea/150",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-06b04d219fea#slide-150",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "They report that using a technique called conditional training during pretraining reduces undesirable content compared to finetuning on human feedback.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c5-73ac-aa5f-4a5a7288ce87",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.35,
        "w": 0.38,
        "x": 0.58,
        "y": 0.48
      },
      "kind": "chart",
      "text": "Toxicity score vs Tokens seen",
      "attrs": null,
      "subkind": "line",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "3fc43656-ba3e-45c3-89ea-784ae74bb19a",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.06,
        "w": 0.25,
        "x": 0.03,
        "y": 0.93
      },
      "kind": "image",
      "text": "University logos",
      "attrs": null,
      "subkind": "logo-grid",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "64b6977a-e8e6-49c3-99bd-bc6267bf174e",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.18,
        "w": 0.55,
        "x": 0.03,
        "y": 0.46
      },
      "kind": "list",
      "text": "For conditional pretraining, the authors score the pretraining data using a reward model and add a token “good” or “bad” at the beginning of each sentence depending on a comparison of the score with a given threshold. The model is then trained on this augmented dataset, but at inference, the generation is conditioned on “good”.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "6ca52b97-9698-4762-83eb-292397e8a7c8",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.55,
        "x": 0.03,
        "y": 0.65
      },
      "kind": "list",
      "text": "The authors tested their method on relatively small models and datasets by today’s standard, but Google used their approach on PaLM-2 with a small percentage of their pre-training data and reported reduced probability of harmful content generation.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "7f948547-6a71-4705-bba6-f5a0c09bc5a0",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.15,
        "w": 0.55,
        "x": 0.03,
        "y": 0.21
      },
      "kind": "list",
      "text": "Researchers from the University of Sussex, NYU, FAR AI, Northeastern, and Anthropic suggest to incorporate human feedback directly in the pretraining of LLMs. They report that using a technique called conditional training during pretraining reduces undesirable content compared to finetuning on human feedback.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "a8f01099-a93e-4d0d-b4ed-5c4bfd58438a",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.08,
        "w": 0.55,
        "x": 0.03,
        "y": 0.37
      },
      "kind": "list",
      "text": "As discussed earlier in the report, modern LLMs are typically trained in 3 phases: pretraining on large text corpora, supervised finetuning on a few thousand (instruction, output) samples, and RLHF.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "dc40b0a7-cc84-4228-a389-2044bf3db26c",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "Toxicity score",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-47c5-73ac-aa5f-4da542c62a7e",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.5,
        "x": 0.03,
        "y": 0.14
      },
      "kind": "title",
      "text": "Pretraining Language Models with Human Preferences",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "04a0617a-f79b-40f1-9c11-7e1bd20a132b",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 155,
      "from": 121,
      "beatId": "019dd95a-0682-776c-8e35-478c8c58938c",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Implications (So What)",
      "beatSlug": "triple-take-the-implications-so-what",
      "evidence": "Politics + Safety sections explore regulatory, geopolitical and risk implications.",
      "position": 2,
      "confidence": 78,
      "parentBeatName": "Reflection",
      "parentBeatSlug": "reflection"
    },
    {
      "to": 155,
      "from": 121,
      "beatId": "019dd95a-0682-776c-8e35-554ec1797b3a",
      "arcName": "The Mountain",
      "arcSlug": "mountain",
      "beatName": "Climax",
      "beatSlug": "mountain-climax",
      "evidence": "Regulatory divergence and the x-risk debate erupt into the mainstream.",
      "position": 3,
      "confidence": 45,
      "parentBeatName": "Turn",
      "parentBeatSlug": "turn"
    }
  ],
  "loops": [
    {
      "to": 155,
      "from": 149,
      "name": "Pattern Hunter",
      "slug": "02-pattern-hunter",
      "bestFor": "Time-pressed audiences, building consensus, when data is strong",
      "matchId": "019dd95a-07fe-70ce-8d3e-1a62463741ad",
      "evidence": "RLHF survey, conditional pretraining, Constitutional AI, scalable supervision, process eval, interpretability, benchmark drift.",
      "position": 15,
      "objective": "Show alignment research progress on RLHF problems",
      "structure": "Evidence A -> Evidence B -> Evidence C -> Pattern/Conclusion",
      "confidence": 70,
      "description": "Group multiple pieces of evidence that together point to a pattern or conclusion"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}