Genesis · Flagship II · How it works
Summoning swarms from a text prompt
Genesis is a WebGPU particle-life lab. You describe a lifeform in plain English and a foundation model plus an evolutionary search coax it out of a living swarm, entirely on your own machine, no server. It is the generative counterpart to Catchment, which runs a physics engine on real terrain.
The simulation
Particle Life sets 2,400 agents of six species loose under an asymmetric attraction matrix, extended here with flocking, a global wind field, per-species force pulsing and cyclic predation. Waves of colour sweep the swarm as species convert each other. Everything renders through an HDR trail buffer with velocity-stretched, speed-heated sprites. Move the cursor to herd them, click to roll new physics.
Summon by prompt
Type a description and an in-browser CLIP model, running client-side on WebGPU, scores how much the swarm resembles your words. The fitness is contrastive: each candidate is embedded from several views and moments, and the prompt’s similarity is measured against a bank of generic descriptions, so the search climbs on what is specific to your words rather than on glowing dots on black. A separable CMA-ES then breeds a 51-gene genome covering the full attraction matrix, the physics and the palette, warm-started from a trained prior: an offline pipeline evolves genomes for dozens of concepts against the same CLIP model, and your prompt retrieves the nearest ones in embedding space. Fitness is judged across three moments in time, with a bonus for genuinely changing between them, so the search selects for living behavior rather than the best still frame. An open-ended mode instead hunts for restless, ever-changing motion. The approach follows ASAL (Sakana AI + MIT, Artificial Life, 2025), realized here as something you can drive.
Making it feel alive
The attraction matrix slowly drifts, thermal jitter keeps the swarm from settling, and a cursor field lets you steer the agents in real time. Trails, streaks, wind and predation waves turn simple rules into motion that reads as organic, and every lever is a live slider.
Built blind, verified offline
Raw WebGPU throughout, dependency-light, with graceful fallbacks so the page never hard-fails. The simulation rules and the optimizer each have a pure reference implementation validated headlessly, checking bounded dynamics, no NaNs, emergent clustering and confirmed convergence, before being transcribed to GPU shaders.
Honest limits
The swarm is abstract. CLIP nudges colour, density and motion toward the vibe of a prompt, but the medium is particles and trails, not anatomy. Genesis says so up front: ask it for a face or a car and a prompt router, running in the same CLIP text space, will tell you the swarm can only give you its texture, and offer behavioral rewrites that play to what the medium genuinely does. Most tools draw you a picture. Genesis grows you a behavior, then lets you disturb it.
Tip: type catchment or surprise anywhere on the page.