{
  "video": {
    "id": "sDMGWK4wZ_w",
    "title": "Training Taste — Thais Castello Branco, Taste Labs",
    "duration": 906,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer"
  },
  "analysis": {
    "video_id": "sDMGWK4wZ_w",
    "title": "Training Taste — Thais Castello Branco, Taste Labs",
    "one_liner": "Taste Labs' founder argues AI slop is measurable, not just a vibe — they mined 2M+ websites, trained small \"probe\" classifiers on design features, and beat LLM-as-judge at predicting slop, then built a Brand API that structures a brand so agents can follow it and you can verify against it.",
    "summary": "Thais Castello Branco frames slop as her personal enemy and argues that subjective domains like design and writing deserve the same decomposition effort that went into coding and math. She shows research measuring slop quantitatively — pattern-mining 2 million+ websites over ten years plus a synthetic AI-generated comparison set, then training small classifiers (\"probes\") per design characteristic whose combined signal predicts AI slop better than LLM-as-judge. The fixes she proposes live largely at inference time rather than in the model: a \"creativity API\" that pushes agents intentionally out of distribution, and a Brand API (first public product, in beta with design partners) that extracts a brand URL into structured components an agent can follow and a human can judge against.",
    "key_points": [
      "Taste Labs works on two fronts: with frontier labs on evaluating models, finding where they break, and building post-training data or RL environments; and at the app layer on context, judgment and verification for agents using off-the-shelf models that \"collapse to the mean\".",
      "Design decomposes unevenly: color palettes, contrast and alignment become near-deterministic once the problem and context are defined specifically enough, while aesthetics shows genuine expert disagreement and needs data-driven methods instead.",
      "Slop has three recurring characteristics: repetition; lack of fit (a pet shop site and a finance firm site converging on the same design); and low intent, including systems that fail to help the user interpret and enrich their own intent.",
      "They analyzed over 2 million websites from the past ~10 years \"way back machine style\" plus a synthetically generated set of design websites to compare human-made to AI-generated.",
      "The internet was already homogenizing before AI — more similar color palettes and layouts — but AI made repetition far more frequent and, crucially, context-independent: the same patterns showed up across completely different buckets.",
      "Method: pattern-mine features (colors, typography, layout, audience) into structured characteristics, then train \"probes\" — baby classifiers, one per characteristic. Combined probe frequency predicted slop with high accuracy and performed better than most LLM-as-a-judge setups asking a model whether something is human-quality or AI slop.",
      "Proposed fixes at inference time, not just the model layer: a \"creativity API\" acting as an inspiration machine so agents go deliberately out of distribution — explicitly not just raising temperature, but learning a category's rules (e.g. what a good pitch deck looks like) and intentionally breaking a couple of them while keeping adherence elsewhere.",
      "Brand API (first public product, in beta with design partners) takes a brand URL and extracts structured components an agent can follow and a human can verify against; they're also building a repository/index of pre-made cohesive brand systems so a user with no brand can retrieve e.g. a \"dreamy\" one rather than generate on the spot.",
      "Live demo: asking Claude design for a slide deck in the branding of the General Intelligence Company of New York produced a low-fidelity default; running the brand extraction in the process produced something much higher fidelity to the original, right down to the details."
    ],
    "takeaways": [
      "Stop treating quality in subjective domains as unmeasurable — decompose the domain into extractable features and train small per-characteristic classifiers, then combine them; that beat LLM-as-judge for detecting slop.",
      "Use probe-style classifiers as a gate so your agent doesn't ship slop, not just as an offline eval.",
      "Don't rely on the model layer alone. Context, intent interpretation and verification happen at inference time, where the user actually is — fix slop there too.",
      "If a brand already exists, use it: the expensive taste work has already been done by dozens of designers, so extract it into structured components your agent follows and you can judge adherence against. For users with no brand, retrieve a pre-made cohesive brand system rather than generating one on the fly.",
      "Get creativity from structured rule-breaking, not randomness — learn the category's expectations, then diverge deliberately on a couple of dimensions while staying in-category elsewhere."
    ],
    "topics": [
      "evals",
      "design",
      "ai-slop",
      "classifiers",
      "post-training",
      "agents",
      "branding",
      "llm-as-judge"
    ],
    "tools": [
      "Taste Labs",
      "Brand API",
      "creativity API",
      "Claude (transcribed as \"Cloud Design\")",
      "Wayback Machine",
      "General Intelligence Company of New York"
    ],
    "quotes": [
      {
        "text": "our whole mission is basically how do we end AI slop? I that's my personal enemy.",
        "at": "00:22",
        "url": "https://www.youtube.com/watch?v=sDMGWK4wZ_w&t=22s"
      },
      {
        "text": "I think it is hard to define what is great sometimes, but I think it's pretty pretty easy to define what is slop in the sense that most people would agree.",
        "at": "03:23",
        "url": "https://www.youtube.com/watch?v=sDMGWK4wZ_w&t=203s"
      },
      {
        "text": "This performed better, by the way, than like most LLM as a judge methods of like asking an LLM to like judge if that uh is like great human quality versus like AI-generated slop.",
        "at": "08:30",
        "url": "https://www.youtube.com/watch?v=sDMGWK4wZ_w&t=510s"
      },
      {
        "text": "as the cost of production basically goes to zero, I think the thing that becomes expensive and matters more than ever is judgment.",
        "at": "08:58",
        "url": "https://www.youtube.com/watch?v=sDMGWK4wZ_w&t=538s"
      },
      {
        "text": "Like the bar is currently, I would say, on the ground.",
        "at": "14:25",
        "url": "https://www.youtube.com/watch?v=sDMGWK4wZ_w&t=865s"
      }
    ],
    "words": 3836
  },
  "summary_url": "/#sDMGWK4wZ_w",
  "transcript": {
    "html": "/transcripts/sDMGWK4wZ_w.html",
    "txt": "/transcripts/sDMGWK4wZ_w.txt",
    "vtt": "/transcripts/sDMGWK4wZ_w.vtt"
  }
}