Taste Labs CEO Thais Castello Branco on Ending AI Slop

Speaking at the AI Engineer World's Fair, Taste Labs founder and CEO Thais Castello Branco outlined a strategy to eliminate what the tech community calls AI slop. While frontier artificial intelligence models excel at verifiable tasks like coding and mathematics, they remain noticeably weak in subjective fields like visual design, creative writing, and emotional intelligence. Castello Branco argues that solving this deficit requires transforming fuzzy, subjective evaluations into measurable infrastructure across the entire model training stack.

Who Is Thais Castello Branco

Thais Castello Branco is the founder and CEO of Taste Labs, an AI infrastructure startup that recently emerged from stealth. The company acts as a data and evaluation layer for frontier AI developers and application creators. Taste Labs collaborates directly with top foundation model laboratories to benchmark model capabilities and construct specialized post-training data. The firm also works with application developers to solve context and intent challenges that foundation models cannot fix on their own.

Why Artificial Intelligence Struggles with Subjective Domains

Model performance historically relies on measurability. Coding models perform well because code possesses three core technical properties: it decomposes into discrete blocks, it verifies against logic checks, and it executes programmatically. Subjective fields like graphic design or essay writing lack these native properties, making standard model evaluation far more complex.

Castello Branco highlighted two primary reasons why subjective domains lag behind technical ones. First, capability follows measurability. If developers can measure a performance metric accurately, they can train models to master it. Second, predicting the mean outcome is inherently suboptimal for subjective work. In mathematics, the average prediction for an equation is also the correct answer. In creative writing or visual design, predicting the average output leads directly to repetitive, bland content.

"A lot of greatness and creativity happens actually at the ends of the distribution. It's when you actively break from rules and actively break from patterns that you can create things that are subjective and great." - Thais Castello Branco

Decomposing Fuzzy Design into Ground Truth Verifiers

To train reinforcement learning models effectively on subjective tasks, engineers must convert vague human criteria into structured verifiers. Castello Branco demonstrated this process through brand adherence testing. Asking an LLM judge whether a webpage looks good yields unreliable scores due to reward hacking and hallucination patterns.

Instead, Taste Labs decomposes a corporate brand into specific, codified elements. These elements include exact color palettes, typography specifications, element spacing, and motion animations. By breaking down a brand identity into programmatic components, developers establish a solid ground truth. An automated agent can then generate an entirely new page layout and receive automated grading against those precise visual parameters rather than relying on crude surface-level replication.

The Routing Framework: RL Environments vs Human Taste

Not every element of a subjective domain can or should be converted into an automated verifier. Castello Branco presented a framework that routes design tasks along a spectrum ranging from strict verification to human judgment.

• Verifiable programmatic checks: Tasks like visual alignment, element contrast, and layout overflow can be tested algorithmically using deterministic rules.
• Reference rewards and learned models: Intermediate evaluations like typography quality and brand adherence draw on extracted guidelines and verifier suites.
• Human preference judges: High-level attributes like creative originality, narrative style, and emotional resonance require human feedback.

When tasks shift toward human preference, gathering consensus becomes complex. Castello Branco noted that human preferences naturally diverge. Merging conflicting feedback into a single dataset causes preference data to collapse back to an uninspired average. Taste Labs addresses this by attaching individual preference vectors to expert judges, preserving distinct aesthetic styles rather than forcing artificial agreement.

Solving Data Quality in Post-Training Workflows

Building effective training data for subjective AI domains requires strict quality controls rather than raw volume. Taste Labs maintains a network of over a thousand specialized design experts across various artistic mediums to provide high-signal feedback. Castello Branco emphasized that specificity in expert commentary directly impacts training effectiveness, particularly when linking text critiques to exact components within source code.

According to StartupHub.ai data, top model developers like OpenAI hold a score of 84/100, while major industry players like Alphabet Inc. (NASDAQ:GOOGL) score 73/100 and Perplexity AI scores 71/100. As these leading frontier labs push into complex multimodal tasks, the demand for structured evaluation data in non-deterministic domains continues to rise rapidly.

Castello Branco concluded her presentation by reiterating that AI developers working on subjective capabilities must prioritize data quality over raw dataset size. High-cost, expert-curated data yields far better model capability gains than massive volumes of uncurated, noisy preference scores.