AI Engineer World's Fair 2026
Ending AI Slop
Loading slides…
Read the talk
Ending AI Slop
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.
From a talk by Thais Castello Branco
Before you start: Basic familiarity with model training, reinforcement learning and preference data will help; no design background is required.
Which problems belong in the model?
How do you improve an AI’s design, writing, personality or emotional intelligence when a correct answer is not enough to define success? These are the subjective domains at the center of Thais Castello Branco’s work as founder of Taste Labs. The starting point is to decompose what quality means before deciding how to train for it.
That decision spans two layers. At the foundation-model layer, evaluation and benchmarking can expose failures that become candidates for reinforcement learning environments or post-training data. At the application layer, a disappointing result may instead reflect missing context or misunderstood user intent. Castello Branco anticipates a world in which many more people create without being domain experts; helping them express what they want matters alongside improving the model’s capabilities. The training problem is the focus here, but it does not absorb every problem an application must solve.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Quality needs a target
Writing, design, sales and marketing frequently admit several valid answers. The difficulty is not just choosing among them: it is defining what would make one excellent. Without a clear target, both evaluation and training become harder.
Code offers a useful comparison. It can be decomposed, executed and checked. Those are properties of the domain that make training easier, rather than capabilities supplied by a model alone. Design and writing do not arrive with an equivalent general-purpose test for goodness. Castello Branco frames this as capability follows measurability: making some part of quality measurable creates a route toward improving it. A second problem remains even after that work—an average or familiar answer may not be the best creative answer.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Great for whom, where and when?
To make a fuzzy request such as “great design” more verifiable, first specify the person, taste and situation it must serve. The same slide could be excellent for a startup and inappropriate for a finance firm. Its quality depends on its use, not only on its visual properties.
Taste also changes over time. A design that feels right today may differ from what worked five years ago or what will work five years from now. Compared with a stable correctness test in code or mathematics, this adds another moving condition to evaluation. Audience, intended use and time therefore belong in the definition of the task, before an evaluator is asked to judge the result.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Turn brand adherence into a task
Suppose a coding agent has produced an internal dashboard or landing page. How do you decide whether it is good? A company already has a useful source of constraints: the work its designers put into defining its brand. Color combinations, typography, spacing and texture embody decisions about which elements work together and why. Asking for something on brand is a more bounded problem than asking for something great.
Castello Branco uses Reducto as the example. Decompose the brand into colors, typography, motion, animation and textures, and the broad judgment starts to expose specific properties an evaluator can inspect. A holistic LLM-as-a-judge prompt still has to infer what the brand means and decide whether the output expresses it. An explicit specification makes the target less implicit.
The next step is task design. Castello Branco flags reward hacking and hallucination patterns as reasons not to assume an LLM judge is always the right evaluator. To shape the problem into an RL environment, the task needs a usable ground truth. The difficult work is establishing that ground truth from the fuzzy goal, not merely attaching a score to an output.
For the proposed Reducto task, the sequence is:
- Establish the specification. Use the decomposed brand components as the ground truth.
- Request a new page. Ask the agent to adhere to the brand while producing something different from the existing page.
- Evaluate the constraints. Grade the result against the specification, rather than treating similarity to the original as the whole objective.
This leaves room for new combinations of familiar components. A page can be on brand without reproducing the reference layout. The specification should constrain the design without making copying the route to success. This is an illustrative environment design, not a reported training result.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Why familiar output can still be poor output
Design properties sit along a spectrum. Alignment and aspects of typography are closer to objective assessment. Style, fit and creativity are harder to measure. Decomposing a design problem therefore creates a routing problem: which method best addresses each component? It does not make every component equally verifiable.
The creative end of that spectrum exposes the second difficulty. Castello Branco contrasts answering two plus two with writing or design. For the arithmetic question, the shared correct answer is desirable. For a creative task, the most familiar continuation may be precisely what makes the output feel repetitive. Her account of collapse toward the mean is a way to describe that repetition, rather than a claim that a model literally computes an arithmetic average of designs.
Creative excellence can involve deliberately breaking a rule or departing from an established pattern. The departure must be intentional: merely becoming unusual does not establish quality. Once the question becomes whether an unconventional choice works, human preference and judgment may provide a better training signal than an environment with a readily verifiable answer.
Castello Branco reports that Taste Labs works with over a thousand experts across different media and styles. The purpose of that breadth is to preserve variation when decomposing the domain, rather than let one dominant style stand in for the entire distribution of good work. Expert diversity is one part of the proposed response to repetitive output, not a complete solution by itself.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Pull what you can toward verification
The more a task can be verified—especially programmatically—the more suitable it becomes for an RL environment. Establishing ground truth and designing the task can move a previously fuzzy requirement in that direction. As context, changing taste or individual preference becomes decisive, the task retains a stronger need for human judgment.
Brand adherence illustrates the movement. Taken as a whole, it can require an expert’s judgment. Extracting the components that matter makes part of that judgment observable and measurable. The routing decision follows the decomposition:
| Component | Evaluation direction |
|---|---|
| Explicit, observable constraints | Verification |
| Codifiable parts of brand adherence | Extract constraints, then verify |
| Context, taste and preference | Human judgment |
The useful unit of choice is the component, not the entire domain. A single design task can contain both checks and judgments.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Keep the person attached to the preference
For the more subjective components, Taste Labs’ position is that human taste remains ahead of LLM judging. The training challenge is to capture that judgment as useful data. Collecting preferences from many people is insufficient if the collection process loses who they are, what they like, why they like it and when that preference applies.
Two people can prefer different styles without either label being wrong. Nor does their disagreement establish that a design halfway between the two would satisfy either person. If those preferences are pooled without context, legitimate differences can look like inconsistent supervision.
Castello Branco proposes attaching something like a preference vector to preference data. The aim is to make matching possible: a judgment can be understood relative to a person’s preferences rather than treated as a universal vote. Training could then preserve multiple preferences intentionally, instead of treating the differences as noise. This is a proposal for representing preference, not a specified vector format or a demonstrated training implementation.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Make expert feedback specific and locatable
Once human judgment becomes the data source, another question appears: what makes the data useful to a model? Some aspects of subjective quality remain difficult to define, but the collection process contains concrete controls. Decompose the domain into meaningful categories, select experts deliberately across those categories, inspect their qualifications and apply strict selection criteria. Coverage across media and styles should be a collection decision, not an accident of who happens to annotate.
Then route each problem to the appropriate process and inspect the resulting annotations for signal. Castello Branco emphasizes specificity: an expert’s reasoning should identify precisely what is good or bad and explain the observation. A broad verdict leaves the model with much less information about which feature drove the judgment.
For a landing page, the next improvement is to connect the observation to the relevant code component. A paragraph describing the visual result can leave the model to infer which code produced the criticized feature. Linking the commentary to the exact component makes that relationship explicit. A TypeScript annotation record can express the connection directly:
typescript
type DesignAnnotation = {
assetId: string;
component: {
file: string;
exportName: string;
selector: string;
};
criterion: string;
observation: string;
};
const annotation: DesignAnnotation = {
assetId: "landing-page-review",
component: {
file: "src/components/Hero.tsx",
exportName: "Hero",
selector: "[data-component='hero-title']",
},
criterion: "alignment",
observation:
"The heading begins to the left of the body copy and primary button.",
};
Here, the example record preserves both the specific visual observation and where to look in the implementation. It records a critique; it does not apply a design change. The collection workflow and its QA checks should work together to produce this kind of clear, grounded signal.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Quality does not require universal agreement
Human QA has two different jobs. In a collection of slide-design preferences, reviewers can check whether annotations follow the specifications and contain useful reasoning. Asking a second designer to agree with the first designer’s preference is a different test. The meaning of disagreement depends on what they disagree about.
| Disagreement about | QA response |
|---|---|
| Alignment or another relatively objective property | Investigate a possible error |
| Style or aesthetic preference | Preserve a potentially valid difference |
Alignment disagreement may expose a flaw in the observation or annotation. Aesthetic disagreement can reveal the preference distinctions the dataset ought to capture. Screen for sound fundamentals without requiring a single shared taste. Consensus is useful when applied to the right question.
The final test is how the data interacts with models. Taste Labs conducts research on that interaction, but Castello Branco acknowledges that work with frontier labs does not always provide a feedback loop precise enough to identify the downstream effect of a particular data contribution. That limitation makes controllable properties—expert selection, decomposition, specificity and annotation quality—especially valuable to measure within the data workflow.
Castello Branco closes by advocating quality over quantity. High-quality data in subjective domains is expensive, difficult and dependent on deep expertise. Her recommendation is to invest in carefully decomposed tasks and strong expert judgment rather than accumulate noisy, poorly specified examples. She argues that this produces better results, without supplying a quantified comparison. The expense buys a more useful training signal: judgments whose domain, reasoning and intended meaning have been preserved.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
The public website used as the talk's brand-adherence example; its current design may differ from the presentation.
Further reading
- A manifesto for judgementArticle
Thais Castello Branco's companion essay on subjective quality, preference diversity, and taste at training and inference time.
A framework for representing multiple perspectives and preference distributions in language models and their evaluations.
Studies agreement between language-model judges and human chatbot preferences, including systematic judging biases.
Full transcript
Loading transcript…