Primordial A Study of Neuroevolution and Emergent Behaviors

Organisms evolve bodies and brains from scratch. No gradient descent. No reward engineering. Just physics, mutation, and survival.

Scroll to explore

The Project

What evolves when nothing is designed?

Drop fifty organisms into a world with food, physics, and nothing else. No loss function. No gradient descent. No one telling them what to optimize for. They have spring-mass bodies they didn't choose and neural networks they didn't train. The ones that find food survive long enough to reproduce. The ones that don't, disappear. That's it. That's the entire system.

What happens next is the interesting part.

Bodies change shape. Sensors appear, persist, go extinct. Predators emerge not because anyone programmed predation, but because eating other organisms turned out to be a viable energy strategy. Species radiate into empty niches, then consolidate as the winners crowd everyone else out. Whole lineages vanish in a bad thousand ticks. The technical term is neuroevolution: evolutionary algorithms optimizing neural network architecture, weights, and morphology simultaneously. What it looks like is a miniature history of life, playing out in fast-forward on a laptop.

The Question

How much does the environment decide?

Change one number (the rate food spawns) and evolution produces completely different organisms. Abundant food? Every creature converges on the same minimal body: a core and two mouths. Why waste energy on sensors when lunch is everywhere? Cut the food supply in half and suddenly sensors matter, bones matter, fat stores matter. Bodies diversify. Predation becomes worth the risk. Mass extinctions happen. The ecosystem never settles.

The organisms are a reaction to their world. Tweak sensory fidelity, metabolic costs, reproduction mechanics, lifespan, and the entire evolutionary trajectory shifts. Same algorithm, same physics, same neural network architecture. Different rules, different creatures. The environment writes the organisms.

Why It Matters

Tiny brains with real applications

The neural networks that survive in Primordial are absurdly small. A few dozen weights. No hidden layers worth mentioning. And yet they produce organisms that navigate, forage, flee, hunt, and reproduce. Behaviors that would take thousands of lines of hand-written control logic to approximate. Evolution finds solutions that engineers wouldn't think to try, in architectures small enough to fit on a microcontroller.

That opens an interesting door. Imagine simulating a biological environment at a tiny scale: a bloodstream, a tissue matrix, a neural pathway. Let evolution produce a controller for navigating that space. The resulting "brain" would be lightweight enough to run on physical nanotechnology. Nanobots that target pathological cells. Infection response. Plaque removal. You wouldn't need to hand-design the control policy. You'd design the environment, define what survival means, and let selection pressure handle the engineering. It's speculative, but the pieces are there: neuroevolution produces exactly the kind of minimal, battle-tested controllers that real-world nanotechnology would need.

Mostly, though, Primordial is a sandbox for watching evolution work and understanding why it produces what it produces.

1.35M
Simulation Ticks
1,373
Generations
5
Experiments
0
Loss Functions

The Experiments

Same algorithm. Different rules. Different life.

Part 1

Emergent Behaviors & Neuroevolution

Abundant food, simple senses, short lives. Evolution gets free rein and makes a choice no one expected: it strips organisms down to the bare minimum. Every creature converges on the same 3-node body. Sensors go extinct. The simulation discovers that when the rules don't reward complexity, complexity dies.

100k ticks 128 generations 4,345 events

Part 2

Fixing the Rules

Cut the food. Triple the lifespan. Give sensors real information and bodies a reason to grow. Seven changes to the rules, and the ecosystem comes alive. Mass extinctions. Fragile recoveries. Species that radiate and collapse in boom-bust cycles. 457 generations of organisms that never stop adapting.

300k ticks 457 generations 5,759 events

Part 3

Bodies That Matter

Muscle powers speed. Bone extends reach. Armor reflects damage. Seasonal food cycles and spatial gradients punish simplicity. Sensors crash to near-extinction during the armor rush, then recover to become the dominant node type. All seven body parts earn their metabolic cost. Evolution converges on balance, not minimalism.

300k ticks 203 generations 14,384 events

Part 4

Minds and Signals

Recurrent neural networks with persistent memory. Chemical signaling between organisms. Four terrain biomes, a day-night cycle, toxic food, and drifting hazard zones. A dedicated claw node that triggered the most complete arms race in the project's history. Six of ten node types reached universal adoption.

150k ticks 228 generations 10 node types

Part 5

The Economics of Death

Corpse drops make killing profitable. A 30% lethal threshold makes it fast. Quadratic metabolic scaling punishes size. Three full Lotka-Volterra predator-prey cycles, each smaller than the last. The oscillation dampens to zero. A single species dynasty controls 60-75% of the population for 300,000 ticks straight.

500k ticks 357 generations 8,577 species

Tech Stack

Python NumPy Neuroevolution Spring-Mass Physics Spatial Hashing D3.js Canvas API
View Source → ← All Projects