← back
Ending AI Slop — Thais Castello Branco, Taste Labs
Takeaway
Reduce AI slop by decomposing subjective goals into contextual, measurable components while preserving room for creative variation.
Summary
- Taste Labs targets subjective capabilities such as design, writing, personality, and emotional intelligence by making their quality dimensions more measurable.
- Good creative output depends on audience, context, and changing tastes, so one universal score cannot adequately represent success.
- Brand adherence can be decomposed into colors, typography, spacing, motion, and textures, creating ground truth for evaluation or reinforcement-learning tasks without requiring copies of an existing page.
- The proposed approach routes more objective components to verifiable checks and treats style and creativity separately; optimizing for the most likely output risks producing bland average work.
creative-aireward-designdesign-evaluation
Original description
Thais Castello Branco's starting point is that AI is still badly behind on the subjective work, the writing and design where quality is real but hard to pin down, and that ending the slop means building data and reinforcement environments for taste. She sorts domains along a spectrum: at one end things that verify and execute cleanly, at the other pure preference with no ground truth, and most valuable work sits in between. The move that makes taste tractable is decomposition, breaking something like a brand or a page into elements that can each be graded against an original rather than judged whole. The deeper problem she names is collapse to the mean. A model optimizing for the most likely output drifts toward the average and quietly kills the creativity that good design depends on, so the goal is data and rewards that reward breaking from the obvious when the situation calls for it. Taste Labs builds this by turning expert judgment into structured, high signal preference data, pairing it with human QA where reviewers tie their commentary to specific choices, and being careful that preference signal stays rich instead of noisy. Her argument is that as more subjective domains become measurable this way, taste becomes something you can actually train. Speaker info: https://x.com/thaiscbranco_ / thais-castello-branco Timestamps: 0:00 - What ending AI slop means 1:06 - Working with frontier labs on taste 2:33 - The verifiable to preference spectrum 3:38 - Decomposing design to grade it 5:30 - Verifying whether something is on brand 7:40 - Collapse to the mean 9:18 - Shifting preference toward verifiable 11:25 - Building a preference vector 13:05 - Human QA tied to specific choices 15:00 - Training taste in subjective domains