LLm's and the end of the ugly.

Why AI lowered the floor of creative work and didn't raise the ceiling, and what we still have to build.

Narcissus was too beautiful, and the seer Tiresias had warned him: he would live a long life, as long as he never knew himself. One day, thirsty, he leaned over a pool of clear water and fell in love with the image he saw there, not understanding that it was his own. He reached out to touch it, and the water trembled. He tried again, and the face pulled back. He stayed at the edge, trying to embrace a reflection that returned everything he did and gave nothing back. He wasted away there. Where he fell, a flower grew.

There is a second character we tend to forget in this story. The nymph Echo loved Narcissus, but she had been cursed to repeat only the last words she heard. She could never say anything of her own. She was rejected, and she wasted away too, until all that was left was her voice, echoing scraps of what others had said.

I think about the two of them every time I open an AI tool.

I ask for a bold image and the light always comes out the same. Soft, golden, hitting from the side, the way it has hit a thousand images I saw before I asked for this one. I ask for text and it arrives in the shape I expected. I ask for a website and I get a layout I recognize without ever having seen it. All competent. Nothing new. The screen is the pool: it gives me back a polished version of what I was already able to imagine. And the machine behind it is Echo, repeating, in an impeccable accent, fragments of what has already been said.

I do this exercise almost every day, and almost every day I end up in the same place: good ideas that look like ideas I had already seen. I see a few creatives forcing the new into being, but they are the exception, and you can feel the effort. The rule is the well-finished average.

This is not a lack of skill in using the tool. It is the tool's DNA.

It was listening to John Maeda, in a conversation on the a16z podcast, that the distinction became clear to me. Any technology can do two things to creative work: lower the floor, letting more people in, or raise the ceiling, pushing the best higher. AI has done a lot of the first and almost none of the second, and it would be dishonest to pretend the first is not a real gift. People who could never draw now deliver something presentable. People who froze at a blank page now begin. The barrier of execution fell, and it fell for everyone. That is good, and it is big.

But the floor rose and the ceiling didn't.

The reason is in the tool's DNA. A generative model, for images or text, is first trained to predict the most probable output, which is, by definition, the average of everything made before. Then it goes through a second adjustment, RLHF (reinforcement learning from human feedback): human raters rank the responses and the model learns to repeat what was approved. The machine is pulled toward the center twice, by probability and by human approval. Call it regression to the average taste. It was not optimized to be right in the creative sense. It was optimized to be approved.

Part of the blame, it's fair to say, is ours. What we expect is a solution, an answer, and not a doubt, a provocation. We assume the machine is infallible because it is a machine. And in the creative case, we assume something stranger still: that it already has our taste, without our ever having said what it is.

The easy conclusion here would be: then teach the machine your taste. It is wrong.

It is wrong for two reasons. First: the machine has neither bad taste nor good taste, it has every taste. It was built on the images, the texts and the sounds that different people, in different times, called good. But there is a vast difference between the acceptance of the average and works that transformed our perception of what is good. The second reason is the one that took me longest to admit: you don't have a fixed taste to teach. Taste is not curation, it is construction and conviction. It is not a verdict you apply; it is a decision you make. Conviction, not correctness.

When was the last time your taste really changed?

In my case, it was never the right answer handed to me on a plate. It was the opposite: something that bothered me, that I didn't know was ugly or beautiful, and that forced me to stop before judging. To listen to where it came from. To understand the intention. To let what I find ugly and what I find beautiful fight a little. The original idea almost always begins like that: in friction, in a shock, in something you have not yet learned to appreciate.

And this is where the two conversations become one. Taste does not develop by piling up right answers; it develops through a process. The same process that produces original work: suspend judgment, investigate, let yourself be surprised, and only then decide what to bet on. Generative tools learned from finished works, never from the process that made them. They know how to imitate the result. They have no idea how to investigate.

It is worth looking at those who tried the opposite. In the 1970s, the painter Harold Cohen turned to programming instead of painting and built AARON, a program that made art on its own. Cohen did not train AARON on a bank of finished images; he encoded a process, decisions about how to compose, how to mark the canvas. The result was that AARON surprised him: it produced things Cohen would not have made. Exactly the opposite of today's tools, which hand you what you already expected.

Brian Eno got there from the other side. With Peter Schmidt, he made Oblique Strategies, a deck of cards with cryptic instructions to use when the work gets stuck in the studio, phrases meant to knock you off the obvious path. A machine, in its own way, designed to provoke rather than please.

It is the same intuition Rick Rubin describes in The Creative Act: the artist's job is not to be right. It is to trust your own reaction and bet on it. He puts the audience last and the maker's conviction first. Being right is consensus. Creating is risk.

History keeps insisting on this, if we care to listen. In 1913, the premiere of Stravinsky's The Rite of Spring nearly turned into a riot: the audience had nowhere to place it. "Impressionism" was born as an insult, thrown by a critic in 1874 at a Monet painting. The new arrives, almost by definition, as what the average gut reads first as wrong. A tool optimized for what has already been approved will never hand you that. Not by technical limitation, but by function. Its function is precisely to avoid what has not yet been validated.

So the question that matters is not how to make the machine more right. It is another: what would a tool look like that was built to provoke your taste rather than confirm it?

That tool does not exist yet. What exists today is an LLM and a prompt, a consensus engine, very good at giving you back the center. The creative leap is not in a better prompt. It is in creatives building their own tools and their own processes. Particular, personal, non-transferable, the way every creative process has always been. Friction machines: built to disagree with the average, and sometimes to disagree with you. That is where the singular piece comes from.

Narcissus did not die of ugliness. He died of reflection, trapped in the illusion that, by seeing nothing beyond himself, he was beauty itself. That is the real risk of AI for creative work: not that it makes things ugly, but that it makes a beautiful, empty version of you, and you fall in love with it. The way out of the myth is simple to say and hard to do. It is to lift your eyes from the pool.

I truly believe we can do that. Technology has always driven creation: it produced new styles, opened doors, and was often invented by the very artists who needed it. I think we are entering an era like that again, one in which we will be able to build our own tools and, with them, invent new processes. That is where real innovation lives: not in using the tool, but in creating it.

But this optimism comes with a bill. The machine will not develop your taste for you. It will lower the floor on its own, for free, for everyone, without you lifting a finger. The ceiling is another story. Raising the ceiling will demand what it always demanded: curiosity about your own process, awareness of how you create, and the courage to throw yourself into tools that don't exist yet instead of asking the old ones to be bold in your place.

The floor will keep rising on its own. The ceiling won't. That responsibility has always been on us, humans.