Training Taste — Thais Castello Branco, Taste Labs

Taste Labs describes an approach to evaluating AI-generated design by breaking vague judgments into specific characteristics. Its account distinguishes relatively checkable properties, such as contrast and alignment, from aesthetic choices on which experts disagree. Small trained classifiers, which the talk calls probes, each detect one characteristic. According to the talk description, their combined frequency predicts "slop", meaning repetitive, poorly fitted or low-intent design.

The same emphasis on context appears in its proposed brand tooling: turn a brand's requirements into structured components that an agent can follow and that its output can be checked against. The intended benefit is more explicit guidance and evaluation, rather than asking a model for an unexplained judgment of whether something looks good.