Descriptions:
Thais Castello Branco, founder of Taste Labs, delivers a conference talk on the company’s mission to eliminate “AI slop” — the pervasive mediocrity in AI-generated creative work — by building data and evaluation infrastructure for subjective domains. Where AI has excelled in coding and math because those fields are inherently verifiable and decomposable, domains like graphic design, creative writing, and personality remain stubbornly difficult because there is no single correct answer and what counts as “great” shifts with time and context.
Branco introduces two core reasons subjective domains are harder to crack. First, capability follows measurability: if you can make a fuzzy domain more verifiable — even partially — you unlock a path to improvement through reinforcement learning environments. She illustrates this with brand design, showing how tasks like replicating a style guide can be turned into structured, gradeable environments. Second, models suffer from collapse to the mean: next-token prediction optimizes for the most probable output, which is exactly wrong for creative work where greatness often lives at the tails of the distribution.
Taste Labs addresses this by working with frontier labs on model evaluation and post-training, and by curating a community of over a thousand domain experts across design disciplines to deliberately force diversity back into the distribution. The company recently emerged from stealth and is targeting both foundation model evaluation and application-layer agents, with Branco describing a routing-based framework that decomposes subjective problems into components best solved by RL environments, preference data, or contextual signals.
📺 Source: AI Engineer · Published July 31, 2026
🏷️ Format: Deep Dive
