I wanted to see what it feels like to peer into a pond that thinks. So I built one. Primordia is a tiny living world that runs entirely in your browser: a circular petri dish, seen from above, with a dozen or so glowing organisms that drift, feed, form bonds, feud, reproduce and die. Every one of them is procedurally unique — body, colour, size, motion and name all grown from a single seed — and their minds are a real open-source language model running on your own machine. No server, no API key.

What you're watching

Food motes drift in from the rim. Organisms get hungry, chase them, and dim as their energy drains — when it hits zero they sink and fade to sediment. Well-fed pairs that get along will reproduce, and the offspring blends its parents' traits with a little mutation, so the pond slowly drifts genetically over time. Faint threads appear between organisms that have bonded; some pairs bond, some learn to avoid each other. Now and then one says something in a small speech bubble.

Click any organism and you get its card: name, age, energy, personality, its strongest bonds and rivalries, and its last thought — including whether that thought came from the model or from instinct.

How it's built

The honest split matters to me, so here it is plainly.

The body is steering behaviors. Movement, hunger, the relationship affinity matrix, reproduction and death are all classic, deterministic-ish simulation — wander, seek, flee, separate, plus a lifecycle. It's cheap and it holds 60fps. The whole pond is alive with the AI switched off; add ?minds=off to the URL and you still get births, deaths, bonds and rivalries. That was a hard rule: the experience must never depend on the model.

The voice and intent is a language model. After the sim is already running, SmolLM2-135M-Instruct — an open model of about 135 million parameters — loads lazily through transformers.js on WebGPU, with a WASM fallback, inside a Web Worker so it never stalls the render loop. Round-robin, one organism thinks every few seconds. I hand the model a compact prompt built from that organism's state and persona and ask for a tiny JSON action plus a line of at most eight words. The action biases its steering for the next ~10 seconds; the line becomes a speech bubble.

Why such a small model? Because it has to fit and run in a stranger's browser tab with no backend, and because 135M is honestly enough for this job — I don't need reasoning, I need a flicker of intent and a distinct little voice. One model wears seven personas (curious, grumpy, poetic, anxious, greedy, gentle, stoic), which is how a single 135M model yields many voices. I'm not pretending these are minds; the model is small and it says silly things, and that's the fun of it.

The contract, and its failure mode. The model is asked to answer with one line like {"action":"seek_food","say":"so hungry"}. I parse it defensively: pull out the JSON, validate the action against a fixed set, confirm any named target is a real nearby organism, clamp the text. On any parse failure, timeout, or if the model never loads at all, that organism silently falls back to instinct and a clearly-labelled canned line. The UI never dresses instinct up as AI — the HUD states exactly what's running (SmolLM2-135M running in your browser, or instinct mode), instinct speech is visually marked, and each inspect card tells you the source of the last thought.

Try it

• Live demo: https://gmsudo.com/lab/primordia/
• Source: https://github.com/AitorGallardo/primordia

Drag to pan, scroll to zoom, click an organism to inspect it, and watch the minds: line to know exactly what's doing the thinking.