{
  "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": 199,
  "pageCount": 213,
  "prevPage": 198,
  "nextPage": 200,
  "slideType": "industry_trends",
  "function": "summarize",
  "density": "overcrowded",
  "nDataPoints": 5,
  "notes": "Includes a screenshot of an SAE viewer tool and a histogram of activation density.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "screenshot",
    "data_table"
  ],
  "metadataConfidence": 0.95,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0856e1444fb9/199",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9#slide-199",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "The researchers introduce the TopK activation function, which directly constrains the number of active features.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d21-efb91e486315",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.45,
        "w": 0.52,
        "x": 0.46,
        "y": 0.45
      },
      "kind": "image",
      "text": "SAE viewer interface showing model settings, activation histogram, and data table.",
      "attrs": null,
      "subkind": "screenshot",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "ff32a2aa-46d2-4bd4-939a-c8ef23673b08",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.15,
        "w": 0.45,
        "x": 0.03,
        "y": 0.33
      },
      "kind": "list",
      "text": "The researchers introduce the TopK activation function, which directly constrains the number of active features. For each input, only the k-highest activating features are kept, while the rest are set to zero - providing direct control over the sparsity level.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "7de3674a-b524-4678-9f0b-5e2a22b4b6c9",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.45,
        "x": 0.03,
        "y": 0.48
      },
      "kind": "list",
      "text": "They also managed to reduce dead latents to only 7%, an improvement on previous methods, where up to 90% could become inactive in large models.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "8ad0c564-2525-44d8-92fc-bae3612d50d6",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.45,
        "x": 0.03,
        "y": 0.61
      },
      "kind": "list",
      "text": "The OpenAI team also demonstrated both the potential and desirability of scaling, training a 16 million latent autoencoder on GPT-4 activations, finding clear scaling laws.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "9fb0f358-e9d8-4e95-87e8-d5e02b9520b9",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "dead latents: 7%",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d21-f1f3319ea18d",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.1,
        "w": 0.45,
        "x": 0.03,
        "y": 0.22
      },
      "kind": "paragraph",
      "text": "SAEs aren't new, but researchers often struggled with balancing sparsity and reconstruction quality, and latents dying in training (i.e. inactive neurons). OpenAI researchers have worked on a methodology that scales.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "5cf96fd1-91ef-475f-a737-3ed589d4ab73",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.45,
        "x": 0.03,
        "y": 0.14
      },
      "kind": "title",
      "text": "...and starts a trend for sparse autoencoders",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "defe1246-b416-4c7e-9d78-59e7eaf289a3",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Action Titles",
      "slug": "action-titles",
      "agent": "Architect",
      "layer": "slide",
      "matchId": "019dd95a-1055-74e0-a95a-d3b37ae6877f",
      "evidence": "Title '...and starts a trend for sparse autoencoders'",
      "confidence": 80
    }
  ],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 204,
      "from": 154,
      "beatId": "019dd95a-0682-776c-8e35-605159a069e6",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Implications (So What)",
      "beatSlug": "triple-take-the-implications-so-what",
      "evidence": "Politics + Safety sections explore consequences and risks",
      "position": 2,
      "confidence": 78,
      "parentBeatName": "Reflection",
      "parentBeatSlug": "reflection"
    },
    {
      "to": 204,
      "from": 175,
      "beatId": "019dd95a-0682-776c-8e35-74f05e198faf",
      "arcName": "Voyage and Return",
      "arcSlug": "voyage-return",
      "beatName": "The Return",
      "beatSlug": "voyage-return-the-return",
      "evidence": "Safety section returns to risks/governance debate",
      "position": 4,
      "confidence": 55,
      "parentBeatName": "Resolution",
      "parentBeatSlug": "resolution"
    }
  ],
  "loops": [
    {
      "to": 201,
      "from": 187,
      "name": "Pattern Hunter",
      "slug": "02-pattern-hunter",
      "bestFor": "Time-pressed audiences, building consensus, when data is strong",
      "matchId": "019dd95a-07fe-70ce-8d3e-79a3afd2ac42",
      "evidence": "Emergent capabilities, sycophancy, DPO, hybrid RLHF, critiques, transparency, reward tampering, SAEs",
      "position": 23,
      "objective": "Catalogue alignment, RLHF, sycophancy and interpretability research",
      "structure": "Evidence A -> Evidence B -> Evidence C -> Pattern/Conclusion",
      "confidence": 76,
      "description": "Group multiple pieces of evidence that together point to a pattern or conclusion"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}