# Ending AI Slop Thais Castello Branco, Taste Labs | AI Engineer World's Fair 2026 | 16:30 Source: https://www.youtube.com/watch?v=lCBf9slCanI Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/ending-ai-slop Published: 2026-07-31 Tags: evals, fine-tuning, human-in-the-loop, reinforcement-learning ## TL;DR - Subjective capabilities become easier to train when they are decomposed into parts that can be observed and graded. - AI slop comes from models favoring the most likely, average output, even when good writing and design require intentional departures from familiar patterns. - Preference data needs to preserve different tastes, use expert judgment, and tie commentary to specific design or code choices so it stays useful. ## Summary Thais Castello Branco argues that AI is much better at coding and math than at design, writing, personality, and emotional intelligence because those domains have no single stable answer. She explains that code is easier to train because it decomposes, verifies, and executes. Design can become more tractable when it is broken into elements such as color, typography, motion, texture, and spacing, each judged against a brand or other context. She also explains why subjective outputs become generic: models treat the most likely answer as the ideal one, while strong creative work often comes from breaking patterns. Taste Labs works with frontier labs and application companies to decide which problems belong in reinforcement environments, training data, or product context. For preference data, the company tracks who prefers what and why, selects domain experts, and checks whether their comments point to specific code or visual choices. Human disagreement about style can be useful, while disagreement about alignment may expose bad data. ## Key ideas ### Taste Labs treats subjective capability as a data and infrastructure problem [00:13](https://www.youtube.com/watch?v=lCBf9slCanI&t=13s) Thais Castello Branco says Taste Labs wants to end AI slop by decomposing subjective domains. Models have become strong at coding and math, but remain behind on design, creative writing, personality, and emotional intelligence. The company works with frontier labs to evaluate models, find where they fail, and decide whether a problem belongs in a reinforcement learning environment or in post-training data. It also works with agent and application companies on problems that may be better handled through context or user-intent understanding. Her focus in this talk is model training. ### Subjective work is hard because it has many contextual answers [01:55](https://www.youtube.com/watch?v=lCBf9slCanI&t=115s) Writing, sales, marketing, and other company workflows often have several acceptable answers rather than one correct answer. Whether something is good depends on the audience, the situation, and the time. A slide might work for a startup and be inappropriate for a finance firm, while standards for good design also change over the years. Thais argues that the difficulty is not simply a property of the model. Code is easier because code itself decomposes, verifies, and executes. Capability follows measurability, so making part of a subjective task measurable can make that part trainable. ### Decomposing a brand turns a vague design judgment into testable parts [04:26](https://www.youtube.com/watch?v=lCBf9slCanI&t=266s) A company brand contains decisions about colors, typography, spacing, texture, motion, and animation. Asking an agent to create something great is vague, but asking it to create something on brand gives the task a clearer context. Taste Labs can decompose a brand into elements and verify an output against those elements. A new page should be judged against the underlying brand ground truth, not only against the original page, because a valid result may use the same components in a new way. This turns part of brand adherence into a task that can support an RL environment. ### Design needs different grading methods for different parts [06:56](https://www.youtube.com/watch?v=lCBf9slCanI&t=416s) Design contains elements that sit at different points on a spectrum. Vision, alignment, and typography can be closer to objective and easier to check. Style fit and creativity depend more on context and human judgment. Thais describes the problem as routing: break a fuzzy capability into smaller components, then choose the right solution for each one. Some components can use programmatic verification or a ground truth created through task design. Others should move toward preference data and human evaluation because their meaning changes with audience, time, and taste. ### AI slop comes from optimizing for the average output [07:32](https://www.youtube.com/watch?v=lCBf9slCanI&t=452s) Models tend to predict the most likely next outcome and treat it as ideal. That works for a question such as 2 plus 2, where the average answer is also the correct answer. In writing and design, the most likely answer is often generic. Thais says much of the best creative work happens at the ends of the distribution, when someone intentionally breaks rules or familiar patterns. The repetition people experience as slop comes from this collapse to the mean. Training must encourage departures from patterns when the situation calls for them, rather than rewarding novelty without intent. ### The spectrum from verification to preference determines the training route [08:50](https://www.youtube.com/watch?v=lCBf9slCanI&t=530s) Tasks that can be verified programmatically are better suited to reinforcement learning environments. Tasks that depend on context, time, or personal preference require more human judgment and may be better handled with data. Brand adherence is difficult to judge as a whole, especially with an LLM judge, because reward hacking and hallucination can occur. Decomposing the task into observable elements can pull some of it toward verification. The remaining subjective parts still need human taste captured as high-quality training data. Thais presents this routing logic as the main practical takeaway. ### Preference data must preserve plural tastes instead of averaging them away [10:49](https://www.youtube.com/watch?v=lCBf9slCanI&t=649s) Preference data can collapse to the mean when it is collected from people without recording who they are, what they like, or why they made a choice. Different preferences do not mean that one person is wrong, and the best answer is not necessarily the average of two opinions. Thais proposes attaching a preference vector to the data so training can preserve multiple valid styles intentionally. Taste Labs works with a community of more than a thousand designers across different media and styles, partly to keep the training distribution from collapsing. ### Specific expert commentary makes subjective data less noisy [12:55](https://www.youtube.com/watch?v=lCBf9slCanI&t=775s) For subjective data, Taste Labs checks whether expert judgments contain properties associated with high signal. Specificity matters: an expert should explain precisely why an asset is good or bad and identify the observations behind the judgment. A landing-page reviewer might write a paragraph, but the data becomes clearer if the commentary is tied to the exact code component or visual element being discussed. This connection helps models relate code to what a person sees. Thais also argues for quality over quantity because expensive, carefully curated data from people with deep domain taste can outperform larger amounts of messy data. ### Human QA should distinguish factual mistakes from real taste differences [14:04](https://www.youtube.com/watch?v=lCBf9slCanI&t=844s) Human quality assurance has to treat disagreement differently depending on what people are judging. In slide design, experts should generally agree about fundamentals such as alignment. Disagreement about style or aesthetics can be useful because it shows that people have different preferences. Consensus is therefore useful when applied intentionally, rather than as a blanket requirement for every judgment. QA can screen for flaws in objective elements while preserving disagreement that reflects genuine variation in taste. ## Notable quotes - "Capability follows measurability." (02:51) - "A lot of greatness and creativity happens actually at the ends of the distribution." (08:01) - "The best answer is the average of what two people might like." (11:15) - "The specificity of their language and of how precise they're being able to be with how they're doing that description is what will determine that data quality." (12:55) - "I would advocate for a quality over quantity approach." (15:32) ## Tools & references mentioned - Taste Labs - Reduct - LLM as a judge ## Who should watch - You are training or evaluating models on design, writing, or other tasks where several answers can be valid and you need to decide between reinforcement environments and preference data. - Your team is collecting expert feedback and wants to keep distinct tastes from being averaged into generic outputs. - You are building an agent that generates branded pages or other creative work and need a grading method tied to concrete visual or code elements.