"Taste" has become Silicon Valley's new favorite buzzword. It is now in the pitch decks, the investor theses, the product manifestos, and startup job descriptions. xAI literally had a job opening titled "Model Behavior Tutor — Style, Taste & Aesthetics," which required "immediate recognition of 'cheap' vs. 'expensive' prose."

The people who built the blandest, most boring software on the internet are now trying to act like tastemakers. Mark Zuckerberg, who spent years in the same grey T-shirt, suddenly grew out his curls, put on a gold chain, and started wearing oversized designer clothing. He wore a custom Latin tee to Meta Connect, "Aut Zuck aut nihil," roughly "either Zuck or nothing," and then floated an idea of selling his shirts on Instagram as "limited drops." Palantir, which sells surveillance software to armies and governments, now runs a merch store with limited-edition drops and products with names like the Domination Tee and the Ontology OG Hoodie.

As Anu Atluru, a tech essayist widely read by founders and investors, put it, "Taste is Eating Silicon Valley." Greg Brockman, OpenAI's president, called taste "a new core skill," and Sam Altman said the best research teams are built "through context, taste and a real feel for where the field is headed next." Paul Graham, who co-founded Y Combinator, predicted that "In the AI age, taste will become even more important. When anyone can make anything, the big differentiator is what you choose to make." And Cloudflare's CTO Dane Knecht called taste "the engineering differentiator."

Taste has become so important to VCs that in June 2026 a solo founder named Thais Castello Branco raised $18.5M for Taste Labs, a company on a mission to "end AI slop." Her customers are AI companies, both the labs training the foundational models and the apps built on top of them.

Thais' definition of taste is that it's "the bar of quality in the absence of correctness." It's a "compressed judgment learned from pattern recognition, from exposure, from being opinionated and having the courage to say yes to certain things and no to others." What's interesting here is that she defines taste not as a subjective preference, but as a skill that can be trained and developed. Paul Graham made a similar point in his 2002 essay Taste for Makers: "If your job is to design things, and there is no such thing as beauty, then there is no way to get better at your job. If taste is just personal preference, then everyone's is already perfect: you like whatever you like, and that's it."

Paul Graham and Thais Castello Branco are not the first to try to define taste, of course. David Hume, the eighteenth-century Scottish philosopher, took up the question in his 1757 essay Of the Standard of Taste. He acknowledged that taste is a subjective sentiment, not an objective fact. As he put it, "Beauty is no quality in things themselves: it exists merely in the mind which contemplates them; and each mind perceives a different beauty." However, even if every response is valid, it does not make all tastes equal. This is simply not how the world works: "Whoever would assert an equality of genius and elegance between Ogilby and Milton, or Bunyan and Addison, would be thought to defend no less an extravagance, than if he had maintained a mole-hill to be as high as Teneriffe, or a pond as extensive as the ocean."

Hume needed a way to call one verdict better than another without pretending that beauty was a fact of the world. That is how he arrived at his definition:

"Strong sense, united to delicate sentiment, improved by practice, perfected by comparison, and cleared of all prejudice, can alone entitle critics to this valuable character; and the joint verdict of such, wherever they are to be found, is the true standard of taste and beauty."

What Hume is defining here is not just taste, but its standard. The standard of taste, according to Hume, is a consensus of the most qualified critics, the ones who have practiced their art and compared many works, and who judge with delicacy, good sense, and no prejudice.

Hume never claims his critics are right the way a math problem has a right answer, only that their shared verdict is the best benchmark you can have when there is nothing objective to check against. That is the same thing Thais Castello Branco means when she calls taste the bar of quality in the absence of correctness.

A standard built from collective human judgment is exactly what Taste Labs aims to produce. It pools a vetted network of expert tastemakers and turns their judgment into data and tools that machines can use: the datasets and rubrics that AI labs could train and grade their models on, and the verification software that apps and coding agents could run to check their own output.

If you have followed AI even a little, you know these models are built to find the consensus. Their outputs drift toward the most common patterns in their training data. You also know that by now these models have ingested and digested almost everything humanity has ever written, including all our best books and journalism. So if taste is a kind of consensus, and a model is a consensus machine that has read our best work, why does it not already have taste? Why does it still produce slop?

Part of the problem is that our best work is only a small fraction of everything these machines were trained on. Language models need trillions of words, which is far more than all the good writing that exists, so most of what they read is ordinary web text. The great books and journalism are in there, but against everything else they are a rounding error. The other part is reinforcement learning from human feedback, or RLHF. Human raters compare the different answers the models give to the same question and tune them toward the ones they prefer. And they tend to prefer answers that are agreeable and safe, so that is what the models learn to produce. Researchers call this failure mode sycophancy: the models tell you what you want to hear and back down when you push back.

This is the opposite of what David Hume meant by a standard of taste. For Hume, the standard was the judgment of a small number of experienced, qualified critics, not the average opinion of the crowd. Taste Labs is trying to provide those "qualified judges" as infrastructure: a practical system for AI training and evaluation, so models learn to recognize quality instead of optimizing for the most widely accepted outputs.

Before you can mass-produce taste, you have to be able to capture it. Can a single person take their own taste, encode it into a rubric and a set of rules, and hand it to a machine? There is good evidence that they can.

Emil Kowalski is a design engineer known for frontend UI work and animation. His libraries Sonner and Vaul power the toast and drawer components across the React ecosystem. He also teaches motion design and helps designers understand how to create better animations.

His main idea is that good animation (what we often call "taste") can be turned into clear rules that AI tools can follow. His framework has three stages:

• First, during creation, the AI decides how an element moves, how it speeds up and slows down. A decision flowchart walks the agent through it, asking whether the element is appearing, disappearing, sliding across the screen, or reacting to the cursor, then returns the exact timing and easing behavior for the specific motion to feel natural.
• Second, a review step checks the animation against a set of Kowalski's rules. It blocks animations that feel tasteless, even if they technically work. For example, perfectly linear motion can feel unnatural, and animations that start too abruptly can feel jarring. These are the kinds of animations AI produces by default without specific guidance.
• Third, a system-level pass evaluates all animations in a product. It scores them across several criteria (e.g., easing and duration, interruptibility, performance, accessibility), and produces a ranked list of what to fix first.

Rauno Freiberg, a design engineer at Vercel, keeps a public list of rules for how a good interface should behave, the Web Interface Guidelines. There are dozens of specific rules: an interaction should not animate for longer than 200 milliseconds, so it feels immediate rather than slow; a text field should not use a font smaller than 16 pixels, because a smaller one makes an iPhone zoom once you tap it; font weight should not change on hover, because the extra weight shifts the layout… you get the idea. Vercel turned the whole set into a /web-interface-guidelines command that any coding agent can run to check an interface and report every violation by file and line. It is Kowalski's review skill again, this time for the whole surface of an app: one person's judgment written as rules and enforced by a machine.

Motion and interactions are relatively easy to formalize because they have many objective, describable rules. The more difficult question is whether enough rules add up to judgment, and whether a machine following them can tell good from bad the way a person does. In practice, rules alone are not enough, because judgment only makes sense in context.

Let's say your business grew 2% last year. Is it good or bad? It's hard to give a definitive answer. If the whole market grew 30% the same year, it's probably not that good. But if your main competitors all shrank by 15%, growing by 2% that same year is probably not that bad. So the absolute number means almost nothing unless you know the context around it.

A rule can describe an object or action, but not the meaning of that object or action in context. Marcel Duchamp demonstrated the meaning of context perfectly in 1917, when he took an ordinary factory-made urinal, signed it "R. Mutt," and submitted it to an art exhibition as a sculpture called Fountain. Nothing about the object was art. It was a mass-produced piece of plumbing anyone could buy at the store. What made it art was the context around it: the decision to present it as art, and the claim that an artist's point of view could matter more than an artist's craft.

Rules "freeze" the taste and context at a given moment in time, but they need to be updated constantly to stay relevant. Taste is a moving target. Having great taste is like being in good shape — it requires constant work. What's tasteful today is mainstream tomorrow and a cliché the day after. Scarcity is one of the elements of good taste, so scaling taste is counterproductive: the very action of mass-producing taste makes it less tasteful. Just think of all the brands using Instrument Serif font on a beige background. Does it look tasteful to you anymore?

So, can machines learn taste? They can learn a taste: a specific person's or specific panel's taste, at a point in time, and only the part that can be clearly formalized. A machine can follow the rules, but it lacks the human understanding and interpretation of the current context. It cannot notice when the context has changed and those rules no longer apply, and it cannot decide to break them. Both of those still need a human, the one who knows why each rule was there in the first place and whether it is still relevant.

Look at the greatest artists and musicians. Not many of them got there by doing more of the same and following the rules more precisely than everyone else. They constantly reinvented themselves and surprised us. When Stravinsky's Rite of Spring premiered in 1913, the audience laughed and shouted so loudly the dancers could barely hear the orchestra, but today it is one of the canonical works of music. The Impressionists were rejected by the Paris Salon, so they held their own independent exhibition. The Salon jury that rejected them was exactly what Hume described: a joint verdict of qualified critics. Even the name we use for them today started as a critic's insult. Bob Dylan went electric and the folk crowd booed him. For a while, all of them seemed "wrong" by the taste standards of their times. But plenty of people were wrong by the taste standards of their times and simply stayed wrong. What made these few great is that they deliberately rejected the standard and moved taste forward. That is something a machine cannot do, because it has no point of view of its own.