November 30, 2025
Chaos, but make it cursive
Drawing with Chaos
Internet loses it as chaos literally spells words — art or math flex
TLDR: A web demo turns chaotic particle patterns into readable text in real time, sparking awe over its hypnotic glitches and debate over whether it’s groundbreaking math art or “just particles”—plus who deserves credit after AI helped build it. It matters because it showcases AI-boosted, accessible math art pushing interactive visuals to the web
Chaos theory just learned to write its name, and the internet is arguing over who gets the credit. A creator fused swirling “strange attractors” (think: mesmerizing smoke patterns) with a clever way to nudge dots into place so they form letters, then launched a GPU-powered web demo. Fans called it “math ASMR,” watching particles vibe into an “O,” then melt into glorious glitch art. The trippy donut moment? Immediately memed as “the portal to Friday.”
But drama cracked open fast. One camp swooned over the elegance: “optimal transport” (moving points smartly) plus chaos equals living typography. Another camp dunked: “It’s just a particle system with a fancy name.” The biggest flame war? Credit. The author fed messy Python to Claude and got a slick web app back. Half the thread cheered “AI pair programmer for the win,” the other half side-eyed: “So… who actually built this?” Meanwhile, performance purists nitpicked WebGPU quirks and Firefox woes; mobile users screamed as their phones became space heaters. Others just typed their names and watched them bloom, then explode, then reform—pure delight. Math goths linked Paul Bourke’s fractals, while jokers declared the Sliced Wasserstein Distance “the most delicious metric ever.” It’s art, it’s science, it’s vibes
Key Points
- •The project reshapes particles from strange attractors into arbitrary text using optimal transport.
- •A direct approach minimizes Sliced Wasserstein Distance via gradient-based optimization on particle states.
- •The sliced formulation is chosen for real-time speed, avoiding O(n^2) pairwise costs with large particle counts.
- •Earlier diffeomorphism learning with an MLP was slower and less visually satisfying; boundary SDF penalties were unnecessary.
- •A GPU-accelerated web demo (built with help from Claude 4.5 Opus) allows real-time text editing and SWD strength control.