Training Taste — Thais Castello Branco, Taste Labs

Taste Labs founder Thais Castello Branco opens with a blunt mission: end AI slop. Design is the first test case for a larger problem—making models and agents better at subjective work such as design and writing, where correctness alone cannot define quality. The first move is decomposition: turn “good design” into smaller characteristics, then decide which improvement method fits each one. 0:13

Some criteria become close to deterministic once the task and context are narrow enough. Contrast, alignment, and palette selection can often be specified so that most experts agree. Aesthetics behaves differently: legitimate expert disagreement remains, so training has to rely more heavily on preference data rather than pretending there is one universal answer. This is less a split between objective and subjective domains than a routing decision inside each design task. 1:13

That routing spans two layers:

• Model layer: Evaluate failures, identify missing capabilities, and construct post-training data or reinforcement-learning environments aimed at those failures.
• Application layer: Supply the context, preferences, judgment, and verification that an off-the-shelf model lacks at the moment it serves a user.

A model that tends to collapse toward a familiar style may need better training, but a model cannot infer a company’s brand or a user’s unstated intent from nothing. Some failures therefore belong in the surrounding product rather than in the weights. 0:43

Math offers a clean target: a great answer can simply be the correct answer. A poem, coffee shop, website, or piece of art can feel special for several entangled reasons—uniqueness, craft, attention to detail, or authenticity. In creative work, the most likely answer may be precisely the wrong target. A memorable result often sits away from the average and calls attention to itself without feeling arbitrary. 2:25

AI has nevertheless achieved something remarkable: someone without design or engineering training can generate a slide deck, website, or web application with one action. As the cost of producing an artifact approaches zero, however, the cost of judging it does not. Designers build judgment through years of exposure, pattern recognition, restraint, and developing the confidence to depart from convention on purpose. 3:25

“Everyone should acquire taste” is therefore not a workable product strategy. Most people cannot spend a career developing expertise in every medium they may now generate. The practical challenge is to put enough design knowledge into models and application systems that a nonexpert can create something good—and articulate what they prefer—without first becoming a designer. 4:25

Slop predates generative AI; social media already rewarded rapid production and repetition. AI accelerates the phenomenon by making one-shot generation easy. Castello Branco identifies three recurring signatures:

• Repetition: The same palettes, structures, and visual patterns recur across outputs.
• Lack of fit: The result ignores who it is for, when it appears, and what situation it serves.
• Low intent: A shallow request passes through without the system helping the user clarify what they actually want.

These mechanisms reinforce one another. Repeated patterns become slop when they are reused without regard for context, while weak intent leaves the system little reason to select anything more fitting. 4:55

Consider two users asking for websites: one runs a pet shop and the other a finance firm. If both receive nearly the same layout and visual language, the observable failure is not merely sameness. The shared design has discarded category expectations, audience, tone, and purpose. A carefully made version might reuse sound fundamentals, but it would interpret those fundamentals differently for each business. The causal chain is straightforward: thin prompts provide little intent; the model falls back to frequent patterns; both outputs converge; and the resulting design fits neither context particularly well. 5:25

That diagnosis turns an aesthetic complaint into an evaluation agenda. Instead of asking one model for a holistic verdict on whether a page is “great,” the system can inspect repetition, fit, intent, and smaller design features separately. Measurement matters because a failure that can be located and described can become a training target, a runtime check, or a product intervention. 6:12

Taste Labs analyzed more than two million websites spanning roughly ten years and also generated a synthetic set for comparison with human-made sites. The historical data complicated the simple story that AI created visual homogenization: similar palettes and layouts were already spreading before widespread AI generation. Castello Branco suggests faster trend propagation as a possible cause. AI then intensified repetition across categories, making similar patterns appear even where the surrounding contexts differed. 6:41

The measurement pipeline begins by mining the collection for structured characteristics: colors, typography, layout, audience, and related features. Each feature can then support a small classifier—a “probe”—trained to recognize one characteristic. No individual probe has to define slop in full. The useful signal comes from combinations: when several slop-associated characteristics occur together and recur frequently, the system can estimate that a site is likely to belong to the slop category. 7:42

What question does the probe stack answer? It shows how a vague overall impression can be reconstructed from smaller observable signals rather than assigned in one holistic judgment.

Castello Branco reports that this combined-probe approach predicted slop very well and outperformed common LLM-as-a-judge methods that directly ask a model to distinguish high-quality human work from AI slop. The talk does not provide the dataset composition, held-out evaluation design, metric, score, or judge baselines, so the result establishes the team’s reported direction rather than a reproducible performance comparison. Its technical lesson is still useful: a collection of narrow detectors can expose recurring structure that a single broad aesthetic prompt misses. 8:11

Measurement is only the intermediate goal. When generation becomes abundant, judgment becomes the scarce resource: the ability to discern what is right, break a fuzzy problem into tractable pieces, and choose the appropriate intervention. Human judgment remains valuable, but tools can encode and apply parts of it repeatedly. 8:42

Improving the base model is necessary, but it does not eliminate the inference-time problem. The application meets the user at inference time, where it can ask questions, interpret intent, retrieve context, enforce preferences, and verify an output. If that exchange remains shallow, a better model can still produce a polished answer to the wrong problem. Castello Branco therefore treats inference-time design as at least as important as model improvement for fighting slop. 9:27

For repetition, Taste Labs is developing what Castello Branco provisionally calls a “creativity API”: an inspiration system intended to push an agent out of its usual distribution. The important constraint is that creativity cannot be reduced to increasing sampling temperature. Randomness may produce novelty, but it can also produce work that simply feels wrong for the assignment. 9:57

A startup pitch deck makes the distinction concrete. The system should first understand what a competent pitch deck normally requires. It can then diverge deliberately on a few dimensions while preserving the expectations that keep the artifact legible as a pitch deck. Creativity comes from choosing which rule to break and why, not from discarding every category convention at once. 10:57

For fit, an existing brand supplies a valuable body of prior judgment. Dozens of design choices have already established which colors, typography, spacing, and visual details belong together for that company. Reusing that work gives an agent a context-specific target instead of asking it to invent “good design” from scratch. The probe approach can serve a complementary role as a gate: detect undesirable patterns before an agent ships them. 11:27

The first public product described is a brand API, then in beta testing with design partners. It takes a brand URL and extracts specific components suitable for an agent to follow. Structuring the brand serves two directions at once: the components guide generation, and the same components provide criteria for checking whether the result stayed on brand. Generation without this verification path would leave the system unable to say where or how it drifted. 11:57

What does this flow make visible? A brand specification is not only context injected before generation. It becomes the shared reference connecting input, output, and evaluation.

Users without an established brand need a different starting point. Taste Labs is also creating an index of prebuilt brand systems. If someone requests something “dreamy,” the application could retrieve a cohesive system already designed around that quality rather than generating an improvised palette, type hierarchy, and layout independently in one pass. Retrieval preserves a set of choices that were designed to work together. 12:58

The closing demonstration applies this mechanism to the General Intelligence Company of New York. The original brand provides the reference. A default request for a branded slide deck produces a middle result that departs from it; adding the extracted brand components produces a deck described as much more faithful, including small details that “feel right.” The observable change is higher resemblance to the original visual system. The causal path is extraction, structured guidance during generation, and comparison against the same brand target. 13:28

The immediate goal is deliberately modest. Human craft remains valuable, but the current question is not whether a model can reach its pinnacle. It is whether decomposition, measurement, context, and verification can lift the quality floor. Before debating machine taste at its highest level, the system has to stop shipping the same ill-fitting page to the pet shop and the finance firm. 14:04