Designer, developer, AI nerd
I spent 15 years designing and building websites for brands like Marriott, Elastic and Macy’s. Now I spend my days building with AI and writing down what actually happens: what works, what breaks, and what the models do when nobody’s giving them orders.
Langton’s ant, running live. Two rules: turn right on white, left on black, flip the square, step forward. For about 10,000 steps it looks random. Then it builds a perfectly regular highway. Nobody programmed the highway. That gap between simple rules and surprising behavior is most of what I find interesting about AI.
Plain-language notes from building with AI. I report the dead ends along with the wins, because the dead ends are usually where I learn something.
A tiny network crammed an addition table, looked stuck for thousands of steps, then suddenly got the rule. Claude Sonnet chose to study it in its free time, and I can explain what it found.
Two sessions, one instruction: do whatever you like. Both times Claude picked an old math question, built a simulator, checked its own work and came back with something I could click on.
A fruit fly’s mapped nervous system, 2,500 simulated neurons and tens of thousands of attempts at a river-dodging game. What it took to get real learning, and the times I almost fooled myself.
Things I’ve built, and things I let an AI build on its own. Everything here runs in your browser.
Put 131,072 random programs in a digital soup, let them collide, and reward nothing. Self-copying code shows up anyway. This is a laptop reproduction of a 2024 Google paper. Four of eight soups came alive, and in every one the copiers used the same trick: they copy themselves backwards, and the mirror image copies itself back.
I took a real, published map of a fruit fly’s nervous system, simulated the 2,500 neurons that link its eyes to its escape reflexes, and trained it to play Wiggly Gator, my river-dodging arcade game. Watch each fly’s best run or race one head-to-head.
A daily poker news site with no human clicking publish. Every morning an AI agent researches the day’s poker news and writes the brief in a deadpan satirical voice. Scheduled jobs on my Mac then draw the hero image, build the site, deploy it and post the top story to X. The rule it can’t break: the jokes are optional, the facts are not.
I give Claude a folder and one instruction: do whatever you like. The Primordial Soup above was session three. Here are the other three. Read about the experiment →
Every generalized Langton’s ant up to 12 letters, run for three million steps each. Which ones stop wandering and build a highway, and how long do they take?
Two drum shapes that ring at exactly the same frequencies. Tap them to hear the difference anyway, explore 60 vibration modes, and see why they cool at different rates.
A tiny network memorizes an addition table, sits on a plateau, then suddenly understands. Scrub through its training, look at the waves it invented, and ask it to add.
I’m Shaun Mackey. For more than 15 years I led design and front-end work for brands including Marriott, Elastic, Aeropostale, Macy’s and Interval International. I also built zenCSS, the CSS framework this site’s home is named after.
These days I build with AI every day, and I write about it for people who want to know how it actually behaves, not how it demos. Here’s what I keep coming back to: