Sierra: You ask an AI to build you a landing page for your new app. And almost every time, you get the same thing: a purplish gradient, text on the left, a graphic on the right.
Marcus: The 'we have SaaS at home' look. I've seen it a hundred times. I just figured AI was good at code and text, but just fundamentally bad at visual design. Like it has no taste.
Sierra: That’s what I thought too. But I read this post by Anshu Chimala, who used to lead AI R&D teams at Apple, and he argues we're just doing it wrong. The problem isn't that the AI has no taste, it's that it has everyone's taste, all at once.
Marcus: What does that even mean? Design by committee?
Sierra: Exactly that. He says these models are just 'next-token predictors.' At every single step, they're trained to pick the most likely, safest, most pleasing option based on millions of examples. So when it has to pick a color, it picks the one that pleases the most people on average. When it picks a layout, same thing. The result is this hyper-bland, inoffensive average of everything.
Marcus: Which is the purple website. Okay, that makes sense. But in the article, he tries telling it 'give me something totally unique, make it random.' And it still gives him the same bland layout, just with some weird metaphors about pottery.
Sierra: Right, because you're asking it to act random, and it doesn't know how. It just predicts tokens that it associates with the word 'random'. So you have to introduce true randomness from the outside. This is the first big technique he shares, from a paper by Sakana AI. It's called 'String Seed of Thought.'
Marcus: String seed... okay.
Sierra: You tell the model, step one: generate a long, random alphanumeric string. Like, with a shell script, something actually random. Step two: use that string as your only source of inspiration for a creative direction. The color palette, the typography, the layout—derive it all from patterns you find in that meaningless string.
Marcus: Whoa. So you're not asking it to be creative, you're giving it a random key and telling it to create a world based on that key.
Sierra: Exactly. And suddenly the outputs are all over the place. Different colors, different layouts, genuinely new ideas. Because the seed is different every time, the result is different every time. You break it out of the 'most popular choice' trap.
Marcus: Huh. That's clever. It’s like a Rorschach test for the AI.
Sierra: The other method is kind of the opposite. Instead of giving it a random seed, you give it an incredibly specific, human-driven vision. He gives this example: 'Build me a landing page for my productivity app, with a bold pixel art theme... Each section should feel like a still from a video game.'
Marcus: Okay, but hang on. Is that the AI 'designing'? Or is that just a human with a very specific taste using the AI as a ridiculously fast graphics engine?
Sierra: That's a fair question. In that case, the human is the art director, for sure. But with the string seed method, the human isn't providing the aesthetic. The AI is generating a novel direction from a completely random input. That feels like a new kind of creativity to me.
Marcus: I'm not so sure. It's still just following a procedure. 'Look for subpatterns, look for numbers'. It feels like a very, very complex set of instructions. It's not having an emotional reaction and translating it into a design. It's just... executing a complex prompt.
Sierra: But is great design always about emotion? Or is it about structure and novelty? Honestly, the article doesn't fully explain how the AI gets from 'a-5-f-8' to a specific shade of blue. I don't totally get how that translation works.
Marcus: Right. It's a black box. But what isn't a black box is the choice to accept the default purple website. You don't have to. You can either feed it randomness or feed it your own weird taste.
Sierra: And that's the whole takeaway. We've been letting the AI drive, and it always drives to the most boring, average place possible. These are two ways to grab the steering wheel. I'm Sierra.
Marcus: I'm Marcus.