Ending AI Slop — Thais Castello Branco, Taste Labs

Thais Castello Branco, founder of Taste Labs, addresses the pervasive issue of “AI slop” in subjective domains, arguing that while AI excels in objective tasks like coding and math due to their verifiable and decomposable nature, it struggles significantly with creative writing, design, personality, and emotional intelligence. The core premise is that “capability follows measurability,” and to overcome AI’s limitations in subjective areas, a new approach focusing on data and infrastructure is needed to understand and quantify “taste.” Taste Labs works with both frontier labs to benchmark and fix models and with application-layer companies to address problems better solved by context and user intent.

The Verifiable to Preference Spectrum and Decomposition

Castello Branco highlights a spectrum of domains:

• Verifiable End: Tasks like coding and math are inherently decomposable, verifiable, and executable. Their “right” answers are largely consistent over time, making them amenable to traditional AI training methods.
• Pure Preference End: Domains like design, writing, and creativity are subjective, contextual, and evolve over time, lacking a single “right” answer. The challenge lies in defining and measuring “greatness.”

To bridge this gap, she proposes decomposition. A subjective concept, such as “brand,” can be broken down into quantifiable elements: colors, typography, motion, animation, and textures. While judging overall “on-brandness” is difficult, verifying adherence to these specific, codified components becomes tractable. This transforms a “fuzzy” problem into a structured one, making it suitable for approaches like Reinforcement Learning (RL) environments where specific ground truths can be established for capabilities like “brand adherence.” The task design itself is the hardest part, turning seemingly fuzzy requirements into measurable objectives.

The Peril of “Collapse to the Mean”

A significant challenge in subjective domains is the “collapse to the mean.” AI models are designed to predict the most likely outcome. While this is desirable for objective tasks (e.g., 2+2=4), it leads to “slop” in creative fields. Castello Branco explains that “greatness and creativity happens actually at the ends of the distribution,” not in the most probable, average outcome. When models constantly optimize for the average, they stifle the creativity and intentional deviation from norms that characterize good design and writing. The goal, therefore, is to find ways to “break these patterns” and “force that distribution back” to encourage intentional divergence from the mean.

Leveraging Human Judgment and High-Signal Preference Data

For elements that are inherently contextual, time-dependent, or related to diverse preferences, human judgment remains paramount. Castello Branco emphasizes “pulling things toward verification” by identifying and codifying the measurable aspects of subjective tasks.

Taste Labs addresses the challenge of capturing human “taste” by focusing on high-quality, high-signal data. Traditional preference data often collapses to the mean because it fails to account for the “multi-preference” nature of the world. Different people have different tastes, and the “average” preference is not necessarily the optimal one. The solution involves:

• Preference Vectors: Creating individual “preference vectors” to understand and match diverse tastes, allowing for a “pluralism of preferences” in model training, rather than noisy, averaged data.

Building Quality Data for Subjective Domains

The speaker outlines several critical aspects for building high-quality data:

1. Problem Decomposition: Breaking down complex problems into manageable components to establish true distributions and guide expert selection.
2. Expert Selection: Employing “strict” selection patterns for experts who possess “high taste” and deep understanding in specific domains.
3. Specificity in Commentary: Ensuring that human reviewers provide highly specific feedback. For example, tying a designer’s comments about a landing page directly to the corresponding code components. This specificity makes the data “way less noisy and way more clear” for models to learn from.
4. Intentional Human QA: Quality assurance should differentiate between objective errors (e.g., alignment issues) and valid stylistic disagreements. Disagreement on fundamentals indicates a data flaw, while disagreement on style or aesthetics can be “good data,” revealing the diversity of preferences. This nuanced approach helps screen for basic adherence while valuing intentional variations.
5. Quality Over Quantity: A strong advocate for “quality over quantity,” Castello Branco states that “creating high quality data is expensive, it’s difficult,” but yields significantly better results than large volumes of noisy, unstructured data. This requires deep domain understanding and meticulous execution.

• Tags:
• LLM Evaluation
• Contextual AI
• Data Quality
• AI in Design
• UX Design
• Creative AI
• Human-in-the-Loop AI
• Reinforcement Learning (RL)
• AI Slop
• Subjective Domains
• Human Preference Learning
• Taste Labs
• Thais Castello Branco
• Brand Adherence
• Mode Collapse
• User Intent
• Design Principles