{
  "docId": "019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "docSlug": "5df6bd1b0447b5f6",
  "documentTitle": "2025 Air Street Capital The State of AI Report 2025",
  "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": 255,
  "pageCount": 313,
  "prevPage": 254,
  "nextPage": 256,
  "slideType": "industry_trends",
  "function": "summarize",
  "density": "dense",
  "nDataPoints": 1,
  "notes": "Includes a diagram illustrating an attribution graph for model interpretability.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "paragraph",
    "flowchart"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/255",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-255",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "This past year, interpretability teams unlocked new methods to trace circuits in language models, shifting the focus from features to bundles of features that interact with one another during processing.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d06-0ec5ae4d2302",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.5,
        "w": 0.25,
        "x": 0.74,
        "y": 0.35
      },
      "kind": "diagram",
      "text": "Attribution Graph: We trace from input to output through active features, pruning paths that don't influence the output.",
      "attrs": null,
      "subkind": "process",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "85c80034-f84c-4940-b53e-105c715d715e",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.55,
        "w": 0.7,
        "x": 0.04,
        "y": 0.31
      },
      "kind": "list",
      "text": "Using cross-layer transcoders (CLT), Anthropic crafted a preliminary “microscope” that unveils the internal processes of a model, pinpointing activation pathways that are causally responsible for specific model behaviors. Moving beyond Sparse Autoencoders (SAE), teams can now investigate internals at a higher abstraction layer, shedding light on actual reasoning patterns.\nThis work was later replicated by Goodfire, an organization purely dedicated to the field of interpretability. Goodfire’s recent $50M Series A round, which included Anthropic, marks the appetite for a sustained focus on this domain.\nMore complex methods aren’t always better, though – Google DeepMind found that linear probes consistently outperformed SAEs at detecting harmful intent both in-distribution and out-of-distribution, contradicting the hypothesis that sparse SAE features generalize better than dense probes.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "4fb1c07f-673d-4ba6-9bcc-e185bc48a887",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.08,
        "w": 0.7,
        "x": 0.02,
        "y": 0.21
      },
      "kind": "paragraph",
      "text": "This past year, interpretability teams unlocked new methods to trace circuits in language models, shifting the focus from features to bundles of features that interact with one another during processing.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "f3a9e471-01ce-454f-9a7e-9c277e65cba5",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.5,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "...but the field of interpretability sees strong momentum",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "b1d739fc-f528-4191-b433-51df4eb7de0f",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Analytical method",
      "slug": "analytical-method",
      "agent": null,
      "layer": "slide",
      "matchId": "d94698e6-01fb-4a4f-b31e-a949b13852e4",
      "evidence": "Using cross-layer transcoders (CLT), Anthropic crafted a preliminary “microscope” that unveils the internal processes of a model, pinpointing activation pathways that are causally responsible for specific model outputs.",
      "confidence": 0.7
    }
  ],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 282,
      "from": 190,
      "beatId": "019dd95a-0682-776c-8e35-a256fe16ea9c",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Implications (So What)",
      "beatSlug": "triple-take-the-implications-so-what",
      "evidence": "Sections 3-4 frame political and safety consequences of the technical/industry trends.",
      "position": 2,
      "confidence": 70,
      "parentBeatName": "Reflection",
      "parentBeatSlug": "reflection"
    },
    {
      "to": 282,
      "from": 190,
      "beatId": "019dd95a-0682-776c-8e35-b713af60bf88",
      "arcName": "The Onion",
      "arcSlug": "onion",
      "beatName": "Core Insight",
      "beatSlug": "onion-core-insight",
      "evidence": "Politics and Safety reveal the structural risks at the core.",
      "position": 4,
      "confidence": 45,
      "parentBeatName": "Turn",
      "parentBeatSlug": "turn"
    }
  ],
  "loops": [
    {
      "to": 282,
      "from": 247,
      "name": "Pre Mortem",
      "slug": "19-pre-mortem",
      "bestFor": "Project kick-offs, risk management, building trust with skeptical stakeholders",
      "matchId": "019dd95a-07fe-70ce-8d3e-bb4d43115ab4",
      "evidence": "Vibe hacking, AI psychosis, alignment faking, scheming, prompt injection — what could go wrong + mitigations.",
      "position": 12,
      "objective": "Surface AI safety failure modes before they become catastrophic",
      "structure": "The Future Disaster (Hypothetical) -> What Went Wrong -> The Prevention Plan",
      "confidence": 75,
      "description": "Build trust by anticipating failure modes before they happen and solving them in advance"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}