Taking longer than usual. Reload

The Marsell blog

The chaos engine: how a thousand people asking for an ad get a thousand different ads

Why Marsell draws its creative choices in code before any model writes, and how it keeps the draws different across every brand and every week.

By Ilya Arbabi, 6 October 2026

Marsell doesn't ask a model to be creative. Before any model writes, it decides the creative choices in code: the angle each writer is forced through, a random word, a constraint, an era, a genre, the emotion to peak, what happens in the first second, who is on screen, where, and the look of the pictures. It draws them from large pools with a seeded random generator, keeps them inside the brand's tone, steers each brand away from what it did recently, and spreads the draws across every customer so no two get the same room. Then the models write inside that draw. We call it the chaos engine.

The problem it solves

Language models converge. On creativity tests, each answer looks original, but answers from different models are far more alike than answers from different people, across model families (Wenger and Kenett, 2025). Models even reuse the same invented names, in the same pairs, run after run (Brzozowski and Chung, 2026). I wrote both up in Why every AI writes the same ad and The ghost couple.

For a marketing product this is not an academic worry. If a thousand customers type "give me an ad", and each request goes to a model that converges, the thousand ads converge. Turning up the temperature doesn't change what the model thinks a good ad is; asking it to be "very creative" closes only part of the gap. What works is structure: take the choices away from the model.

What gets drawn

Every choice a model would otherwise make by habit is drawn first, from pools written by hand or, for a brand's own world, written once for that brand:

Each writer gets its own draw, so the concepts for one ad don't start from the same place either. And one rule applies to all of them: name no one, and describe people by what they do and wear. That keeps the ghost couple out.

The brand's tone decides how far a draw can go

A random draw with no taste would be a disaster for most brands. So every entry in every pool carries a mark for how far it goes: grounded (true of the world, fit for any brand), imaginative (invented but gentle: a trick of scale, dreamlike physics, a period pastiche) or absurd (a gag, an animal doing a human's job, a horror parody).

Each brand's tone, read from its own words, voice and look, sets how far its draws may go, and the draw only samples what fits. A calm walking channel is never handed a talking seagull or a driving exam; a playful brand still is. For a grounded brand, even the operators change: "Film only what is true of the place and the hour" replaces "exaggerate the benefit until physics gives up". And when the owner asks for something specific, the ask opens exactly what it asks for.

It remembers, and steers away from itself

Sameness also happens inside one brand, week after week. So each new draw reads the brand's recent ones, the eras and genres of the last twelve concepts it picked, and every lead it has used, and steers away from them. The draw is stored with the work it produced, so it can be read, repeated or undone.

A thousand customers, a thousand rooms

Independent random draws still collide more than intuition says. It's the birthday problem: draw two first names per film from a list of 160, and you'll probably see a name repeat by about the eighth film. Across thousands of customers, independent draws would keep handing out the same eras, genres and casts by chance, and adding more writers per film wouldn't fix it.

So the draws are spread across the whole installation, not made independently. In each dimension, film number n takes the n-th step of a walk through a shuffled version of that pool, so every value is used once before any value is used again. The walks are reshuffled every cycle, so two dimensions never stay paired: the era that came with a heist this cycle comes with something else next time. The result is the property the product needs: a thousand people typing "give me an ad" get a thousand different starting points, not a thousand samples from the same few.

When the owner directs, the owner wins

The draw is for when nobody has said what they want. When the owner does ("I want a Y2K, low-poly, PS2-style ad"), the owner's words become the only rule, and the draw only supplies the ways of thinking and the tempo. We learned that the hard way: before this rule, one clear style request came back as a hospital, a courtroom and a goldfish's wedding toast.

What it doesn't do

It doesn't make a model a genius. A draw is a starting point, and some starting points don't work. That's why each ad is several concepts from differently-drawn writers, why the best of them is picked, and why nothing goes out without the owner's approval. What the chaos engine changes is the odds: instead of every concept starting from the model's favourite idea, each one starts somewhere the model would never have gone on its own.

FAQ

Is Marsell's output random?

The starting points are drawn at random, from pools and inside limits that fit the brand. The writing, the picking and the approval are not: models write inside the draw, the best concept is picked, and the owner approves what goes out.

Will it make a serious brand's ads weird?

No. Every pool entry is marked grounded, imaginative or absurd, and a brand's tone decides which it may draw. A calm, documentary brand only draws what is true of the world.

Can I choose the idea myself?

Yes. When you say what you want, your words become the rule and the draw only supplies ways of thinking and pacing.

Why not just raise the model's temperature?

Temperature makes each word less predictable, not the ideas. The research on model convergence found that even explicit requests for creativity only partly close the gap; drawing the choices outside the model does more.

Sources