Thais Castello Branco is the founder of Taste Labs, an AI research and infrastructure company developing human judgment in subjective domains such as design. Her premise is that generating more content is easy; making it original, appropriate to its audience, and faithful to a particular creative identity requires fundamentally better evaluation and training.
Castello Branco joined the Latitud Fellowship in 2024, pursued an earlier startup, then worked at Exa before leaving to build Taste Labs.
She subsequently founded Taste Labs with initial backing from Latitud. The company launched publicly in June 2026 with an $18.5 million funding round. Its work includes model evaluation, specialized training data, reinforcement-learning environments, and tools that help application developers account for user intent and brand context.
Turning taste into usable infrastructure
Creative quality can be measured without making everything identical. Castello Branco breaks brand identity into attributes such as color, typography, spacing, motion, and texture. Those characteristics provide measurable criteria for verifiable brand adherence, which she proposes using as ground truth in reinforcement-learning environments to assess whether new designs fit a brand without requiring copies of existing work. This is an illustrative task design, not a reported training result. She cautions that general-purpose model judges can hallucinate or encourage reward hacking.
Average outputs are often the wrong creative target. Whereas a widely agreed answer can be desirable in mathematics, statistically typical designs or sentences tend toward sameness. Her manifesto on judgment is credited as a companion essay on subjective quality, preference diversity, and taste at training and inference time. In her Ending AI Slop talk, she argues that familiar model outputs and preference data pooled without context can produce creative mode collapse, suppressing the distinctive styles that make creative work compelling.
Expert disagreement is sometimes a feature. Designers disagreeing about alignment may expose an error; designers disagreeing about minimalist versus maximalist aesthetics may simply have different, legitimate preferences. Castello Branco favors expert preference data that preserves those distinctions and connects criticism to specific visual or code elements. Programmatic checks handle objective properties, while contextual questions remain with knowledgeable human evaluators—a distinction she develops in her AI Engineer talk.
In August 2026, Castello Branco announced a Taste Labs research fellowship focused on model creativity and human preferences, offering access to compute, custom data, designers, researchers, and a stipend.
Better creative output starts with decomposing quality: verify the constraints you can specify, preserve the preferences you cannot average, and collect expert judgment with enough context to train on.
Thais Castello Branco’s approach to better AI design starts by decomposing “taste” into measurable constraints, contextual judgments, and inference-time systems that can resist repetitive, ill-fitting output.
Decompose subjective work before choosing an improvement method: explicit constraints can move toward deterministic verification, while aesthetics and preference still require data and human judgment.
Small classifiers can each detect one design characteristic; combining their frequency turns a holistic impression into a structured prediction problem.
Useful creative variation is deliberate: preserve enough category expectations for the artifact to fit its purpose, then break selected rules rather than merely adding randomness.
