Engineering a beautiful mind

Disclosure: Amplify is an investor in Taste Labs.

A lot of people know about the remarkable reign and wealth of the Medici family. What a lot of people don’t know is that before the Medici were the Medici, they were merchant scum (actual Florentine class name), 3 ranks down from the elite merchant families of Florence.

In her book Inventing The Renaissance, scholar^† Ada Palmer tells the story of how the Medici changed their fortunes. Imagine you are an important delegate on your way to meet the new pope. Alas! You must stop overnight in Florence, home of the evil bankers which you loathe and fear. But all of a sudden…

“The inner courtyard is so bright and cool, clean sunlight streaming on white stone, with airy rounded arches, and the breeze so fresh it feels almost like stepping into open countryside from crowded streets…and as the banker leads you further in, he has ancient Roman statues, busts of Caesars, gods, those stunning bodies real enough to spring to life… and wait, what’s that sound? People, there in the corner, speaking in a language you don’t know. So you ask, “Who are they?” and the banker answers, “Oh, they’re Platonists, they’re practicing ancient Greek.” Your breath catches again. “But ancient Greek is lost! Plato is lost!” “Oh, no, we have lots of Plato here,” he answers…and now there’s a little boy reciting a poem at you in ancient Greek, and your brain is going: Where am I? This… all this is impossible! These things do not exist! And all the haughtiness and scorn are gone, as suddenly it’s you who feels weak, small, uncultured…and when he sees you staring and wordless, that’s when Cosimo de Medici turns to you and says, “Would your king like to make an alliance?”

The Medici, at least in this semi-theatrical historical fiction, understood the power of a beautiful world. They sought aesthetic, intellectual, and linguistic excellence not only because these things rock, but because they matter, and because they hold critical influence over the gears that really move the world: political power, money, alliances. Subsequent movers and shakers – like DDB pioneering creative advertising in the 50’s – understood this secret and used it to their advantage.

This is why it’s both unfortunate and ironic how ugly the world has gotten. Architecturally, sociologically, and digitally, we have lost any conceivable pursuit of aesthetic excellence.

This is particularly true on the internet. Though most of the original ARPANET pioneers are dead, if one was still alive you could quickly change that by showing them what TikTok is.

The poster child of the ugly 2026 internet non-aesthetic is AI slop. It is all around us, it is growing, it is getting worse. We are ingesting against our will from our feeds, our group chats, and our parents. We see it in the one-size-fits all Claude Design landing pages, the Studio Ghibli cartoon-themed images^†, and the AI writing that has become so prevalent, we need an entire company just to detect it. The promise of AI tools, that anyone can create anything, has led to, yeah, anyone creating anything. To quote the slop itself: the problem is real.

But the problem is also fixable. If trained and used correctly, AI can actually make us more creative and make the world a better looking place. More practically, with the right mix of curated golden datasets, interventions in post-training, clever implementations of RL environments and graders, and documented procedural knowledge from experts, you actually can train models that have the capability to output far better than the slop we’re used to.

The problem with “slop” is that it’s an incredibly hackneyed and ill-defined term. It is carelessly chucked like a grenade over a ridge by keyboard warriors when they don’t like things. So if we’re going to fix it, we must first define it. From a research and engineering perspective, we can identity a more concrete gap in the capabilities of models today when operating in less-verifiable domains:

1. Model responses are objectively bad: they fail to meet bars of quality that just about anyone would agree on.
2. Model responses lack diversity: they give an average, one-size-fits-all answer instead of diverse, creative outputs tailored to their users and prompts.

These are two totally very different problems requiring very different training interventions. They are complex, multi-faceted engineering and research challenges. And fixing slop will not only require solving both of them — so that our models can be used for good — but also, dramatically improving our understanding and continual learning of user preferences – so that our models will be used for good.

To tackle this mission, you would need a team with an unusual background. They would need to be hardcore AI research people, particularly in post-training and RL. They would need to be cracked engineers, understanding how to collect, store, and serve highly curated yet massive datasets. But they would also need to be artists, designers, videographers, connoisseurs of the visual, a group who internalizes and appreciates what we’re trying to get to here.

This fairly accurately describes the Taste team. They are on the (perhaps not) impossible mission to End The Slop. And they are doing it — for both AI labs and consumers — via an ambitious and clever combination of data, research, RL environments, judge models, and agent context tools (like their recent Brand Adherence API and Search). Less than a year after founding, they’re already doing 8 figures in revenue, doubling top line every quarter, and working with most of the frontier labs (large and small). It is safe to say that if you’ve noticed models getting better at design, Taste has played a role in this.

To get an inside look at how they actually do this from a research and engineering perspective, I spent some time interviewing the Taste team. Then I spent some time writing. This post is going to go into the technical details behind:

• Why the way we (pre)train models today tends to lead to slop
• Building a dataset for post-training that teaches models objective aesthetic quality
• Recruiting and working with designers and artists to build golden datasets
• Creating RL environments for design work and building subjective graders
• …and more!

Let us begin.

The anatomy of slop: algorithms, architectures, and pre-training corpora

Models and users produce slop today in large part because (a) model responses fail to meet objective aesthetic quality standards, and (b) models fail to give diverse, creative outputs and instead produce one-size-fits-all answers. Why do models lack these capabilities today?

Training models for averages vs. training models for disagreement

The answer starts in pre-training. In pre-training data, there is no notion of what is good and what is bad (unsupervised!). There are only patterns, and the model is trained to find out what the most likely or common answer is to a question.

In more objective domains like mathematics, science, or even law, this is a desirable quality. 2+2 does equal 4, that is the correct answer, and a model response that outputs 5 is not just “different,” it is wrong. For a well-trained model in these kinds of domains, the average answer is the correct answer.

But in subjective domains like writing or design, the average answer is not the correct answer. If you’re putting proper effort into a subjective domain, you actually want something that’s out of distribution. You want something that’s going to have a particular point of view, that’s going to stand out. But pre-training ends up smoothing all of this out, unable to account for good disagreement.

Not only that, but most of the data on these domains in the pre-training corpus is bad, so the average is also going to be bad. At least if all of the pre-training corpus was badass websites and portfolios, you might get a good average. But as we well know, that’s not how pre-training scaling laws work.

To counteract this dynamic, labs put increasingly disproportionate amounts of time into post-training^†. It is here where Taste spends most of their time, and where you can meaningfully level up a model’s design ability. The core research problem here is training a model that has the capability to produce something good, but the specific texture of that “good” can vary a lot based on user preferences.

To motivate this problem, imagine you take two different architects with different aesthetic visions to Sea Ranch and you show them a house. After you finish staring longingly at the wind-whipped grass, you find that they will probably agree on the more objective qualities of what make the house good: is it built well? With quality materials? But whether it’s to their particular taste will be a matter of disagreement. One might be more of a classical enthusiast, while the other finds the vernacular angles and the houses cut into the cliffs to be sick asf (which they are). Good, vs. to my taste.

This is precisely the problem that models have. In subjective domains, disagreement is good! It’s not something to be averaged out and fixed. It is itself a representation of the rich, diverse tapestry of human preferences and we need to find a way to hold it in memory. At the same time, there needs to be an objective bar of quality that everyone can agree on. How do you collect and curate a dataset for post-training where the data both agree and disagree like this?

For me, this is roughly where LLM writing has gotten to today. Assuming you give them proper context, models today are decent^† at putting together things like short blog posts, landing page copy, or a LinkedIn post. Explanations of concepts are readable, concise, fine. But the second you ask them to occupy a particular style, you might as well have asked the model to solve Navier-Stokes an impossible math problem. I simply cannot get the model to sound like me, no matter how much context and previous writing I provide. Again, good, vs. to my taste.

Is next token prediction the wrong architecture?

There’s also an architectural component of this discussion. The way most LLMs are designed to create is by finding the most probable next token and then selecting the generation with the highest probability. The idea assumes there’s one canonical, correct answer as opposed to several competing options that could be right given different user preferences.

It’s not entirely clear that other architectures would fix this problem, but it is interesting to observe that Diffusion models do tend to be more creative than transformers and have much better style control. In fact, Taste has experimented with inserting image generation as part of the flow of creating an LLM-based artifact like a slide deck. In some tests it actually inspires the LLM to produce something that’s quite a bit more unique than just one-shot prompting.

Building a dataset for post-training that teaches models objective aesthetic quality

Let us begin with the first step of the equation. How do you teach a model to produce something that everyone can agree is high quality?

The most straightforward answer is golden examples, curated instances of what’s considered to be the highest quality, most excellent, best artifacts in a domain (like the Stripe website in 2021). These get added to the post-training mix and have a dramatic impact on output quality.

But just showing the model the golden example itself isn’t enough. The model also needs to understand the procedure that got you to the golden example. It needs to learn the primitives of a particular domain: the brush strokes, the perspective, the color selection. Or in building a website: learning to start with a brief, make a mockup, try different layouts. If you go straight to the end product and ignore the order of operations, the model learns only to ape the golden examples instead of producing something genuinely^† creative.

I don’t want all of my students to paint the Mona Lisa. I want all of my students to have the ability to paint the Mona Lisa.

– Fictional art teacher

Among other things, Taste is in the business of creating and selling these golden datasets to the labs who need them. Their data collection spans not just acquiring static examples of great website or motion design, but also recording the procedural steps that artists and designers took to create those great things. Like myself bothering the Taste team as a pesky interviewer to bring these details to you, their team will (among other things) prod, poke, and otherwise pester these tastemakers piece by piece on the details of their procedural practice^†.

So where do these golden examples come from? Over the past year, Taste has recruited hundreds of highly capable artists and designers from across disciplines, backgrounds, and visual styles. They’ve been recruited not only to represent a diversity of styles, but also a diversity of capabilities. Some are really good at explaining their work. Some are really bad at explaining their work, but can go in a room and cook. Some are just great critics, which is a valuable skill independently of the ability to create (let me believe this fiction).

Though I’ve consciously chosen to ignore any of the broader discussion around AI and creative work in this post, the fact remains that there’s a lot of skepticism in the creative community about AI. How did Taste, a company that is no stranger to sharp criticism, convince hundreds of creatives to help them build better AI models?

At first, it was just a lot of trust building. Thais (Taste CEO) personally spoke with the first 150 members Taste went after, taking the time to build trust and explain her vision for how you can build AI that is actually good for creatives, sketching out a future where designers spend much more time being creative, armed with a powerful toolset they helped build. I believe this, this team believes this, and there are an increasing number of creatives out there who believe it too.

The second, which was more surprising, is that there was and is a keen intellectual interest from creatives in being part of solving these problems. Artists and designers hate slop too. They want to be part of fixing it. And they are curious about how we break down the elements of design, investigating the right abstractions and studying why we think what we think when it comes to good design.

On context, prompts, and the user experience of diversity

Now let’s say you have a model that’s able to output objectively quality designs. How do you ensure outputs are diverse, creative, and strongly match user preferences?

This, too, starts with golden datasets in post-training. Taste takes painstaking measures to ensure that their data is extremely diverse across domains and styles. If the model needs to be able to output dramatically different styles for a Wall Street prospectus slide deck vs. a startup pitch deck, then the data needs to have experts and assets in each of those too. Data in post-training that is not just good, but good in many different ways, will yield models that are the same.

Then there’s the question of user preferences, where Taste collects preference data to help labs with this problem. They show designs and artifacts to experts, asking which they prefer and why, and what changes they would want to see, collecting this at scale from multiple Tastemakers. Making this dataset valuable for post-training requires nuance on:

• How to not end up meshing all preferences towards the average, which for subjective domains leads to a “meh” product that everyone is sort of OK with.
• Capturing the extremes, which is where much of the interesting work really happens.
• When should the user be pushed to a more creative direction?
• On which pieces should there be preference agreement (and the data would be bad if there was disagreement), and on which pieces should there be disagreement because of differing tastes?

Outside of post-training though, much of the work to be done here is in prompting and context. Taste’s Brand API is an interesting initial take on this — it will ingest your brand and design system, and then help you automatically inject context into your prompts that help get generations to adhere to your brand (plus a verifier for that).

There’s prior art for diversity in the concept of temperature. Temperature is a hyperparameter that allows you to control the randomness of a model’s output by changing the selection process of tokens; instead of picking the highest probability next token, you can elect to choose a lower probability one (and by differing degrees).

Temperature is not diversity though, it’s just randomness. You don’t want a model that, depending on how it’s feeling, might output something that’s bang on for your preferences or something that’s completely awful and doesn’t match your prompt. Diversity is not randomness! If you gave the same brief to 5 different designers, they would come up with 5 different projects, but each of those projects would still adhere to the brief. This is what you need to teach the model.

Ultimately, the share of people on this earth who not only care about how things look, but can articulate this explicitly, is small. Your typical user is not armed with an extensive Pinterest board capturing their desired aesthetic. Models will need to get better at implicitly learning about users and understanding what their preferences for design might be…Pinterest for agents must live.

Creating RL environments for design work and building subjective graders

It has been said that anything you can verify, you can automate. One of the more interesting engineering challenges in subjective domains is figuring out how to do that verification. Simply put, one must attempt to objectify the subjective. Not literally (that would be lame), but by shaping RL tasks, environments, and graders to add elements of determinism to previously flexible things like website or motion design.

Taste builds RL environments for the frontier labs that replicate real world design tasks. The hardest part is the grading, figuring out what great actually means for a particular task. They’ve taken two approaches:

1. Fully deterministic grading – where you fork and compare what the model produced against the golden, human-produced asset.
2. Agentic grading – where you feed an LLM a rubric with various binary criteria, and it scores based on that rubric.

Both of these grading methods run on both the golden (human-produced) and the model-produced asset with the assumption that grading reliably scores ~100% on the former and <100% on the latter.

An interesting question is granularity, or how contrived these tasks need to be to be well verified. I’ve probably asked Claude to change a div’s border from gray to black about 100 times; and despite the fact that it still finds ways to mess up, this should be a very verifiable task. But asking a model to generate an entire site based on a brief is extremely long horizon, subjective, and much harder to grade. You need a good mix of both of these for RL to work.

For a more specific example, model capabilities today are still lacking when it comes to interpreting motion and animations. This is largely because models don’t ingest video or have a sense of time, which makes it hard for them to build these assets frame by frame. Jayce (Taste engineer) is working on building RL environments to fix this. A task here might be something like: here is an awesome animation on a website, recreate it from scratch (with a constraint that it can’t just grep the source code).

Another interesting area is avoiding reward hacking. Without good verifiers models will absolutely reward hack design tasks, the most obvious version being just taking a screenshot of the thing that you told it to recreate. You can mitigate this by monitoring the chain-of-thought and feeding it to the grader. But of course, like everyone else, if you do this too much you will run into alignment issues…

So can you take a subjective domain like design and grade it autonomously? The early work of this team says: yes, to an extent. Certainly for the most contrived of the tasks that Taste has experts doing manually, these should be able to be represented as RL tasks and trained autonomously. On the longer horizon, high entropy briefs like “make me a website that pops” – we’ll have to see.

The power of a beautiful world

Does taste matter?

Maybe it doesn’t, because empirically, many don’t seem to care. It has been painful to see many of the engineers I’ve followed on X for years start sharing completely Claudeslopped artifacts (look I made a thing!), not even bothering to adjust a single CSS rule. Even in the physical world – a sanctuary! – you cannot walk a block in New York without seeing one of these bad boys.

Just look around you. It’s slop everywhere, created and shared by people who really just…don’t seem to be particularly bothered by it.

Nevertheless, I do think taste matters. Though I’m no Medici and I’m certainly no Don Draper, I believe in the power of a beautiful world. I, like the Taste team, want to live in a world where people do care about things, and where people don’t fire their slop cannons indiscriminately at vulnerable passerby. The opportunity to shape humanity’s most powerful tools, giving them the capability to generate beautiful, high quality, carefully considered, and diverse outputs that reflect the nuance of their prompters, is, I don’t know, pretty exciting.

The good news is that the labs care about this too. They are spending tens of millions, maybe hundreds of millions, of dollars to make their models able to create beautiful things. And it’s not just a moral imperative, it’s an economic one. Think about who drives the majority of lab revenue today: it’s enterprises. The largest brands in the world cannot be producing slop; they have an interminable bar for quality^†.

You might care about this too, in which case you should consider joining Taste. You might be an engineer interested in working on RL environments for design. Or maybe you’re a creative who wishes you had models at your disposal that better allowed you to express your vision. Either way…they’re hiring.