Artificial Life

There are two ways to simulate life-like behavior on a computer: grid-based systems where you manipulate cell states (reaction-diffusion, Cellular Automata, Lenia), and agent-based systems where you define individual creatures and let them interact. Both produce behavior that looks uncannily biological from rules that are almost trivially simple. The question they raise — where exactly does "simulation of life" end and "actual life" begin — keeps getting harder to answer.

The Two Families

The distinction comes from a generative artist named ciphrd, and it's useful.¹ Cell-based models work on grids: each cell has a state, and you update it based on its neighbors, over and over. Reaction-diffusion, Turing patterns, and Conway's Game of Life all work this way. They're mathematically elegant but tend to be fragile — small changes to the rules often destroy the interesting behavior entirely.

Agent-based models work differently. You define individual agents with positions, orientations, and simple behavioral rules, then let thousands of them loose in a shared environment. The interesting behavior emerges from their interactions. Boids (Reynolds, 1987) is the classic: three rules — separation, alignment, cohesion — and you get flocking behavior indistinguishable from a real flock of birds. The rules say nothing about flocking; flocking is what falls out.

Physarum: Intelligence Without a Brain

The physarum simulation is where agent-based artificial life gets genuinely strange.² Inspired by the slime mold Physarum polycephalum — a single-celled organism that can solve mazes, design efficient transport networks, and approximate Tokyo's rail system — the algorithm gives each agent three simple behaviors: move forward, deposit a chemical trail, and steer toward areas where the trail is densest.

That's it. From these three rules, thousands of agents self-organize into networks that look exactly like the real organism's foraging patterns. The key mechanism is stigmergy — instead of communicating directly, agents communicate through the environment. Each agent leaves a trail; other agents follow trails; trails that get followed get reinforced; trails that don't get followed decay. The environment becomes the communication medium.¹

What makes physarum simulations endlessly generative is that small parameter changes produce wildly different behavior. Adjust the sensor angle, the turn rate, the trail decay factor, and you move through a landscape of biological-looking patterns — from tight branching networks to flowing organic textures to pulsing, breathing structures. Sage Jenson's 36 Points project demonstrates this: 36 points in a 20-dimensional parameter space, each producing a distinct behavior from the same algorithm.²

ciphrd's Ethereal Microcosm pushes further by adding multiple species, each with different properties, that interact through overlapping trail systems. Species can eat each other's trails. Their properties can mutate on contact with other species. The result is a system that never reaches a stable state — it keeps evolving, producing ever-more-complex patterns at multiple scales.¹ The visual results look like footage from an electron microscope, but they're generated by rules you could write on an index card.

Xenobots: When Simulation Meets Biology

The most disorienting development in artificial life isn't computational at all. Kriegman, Blackiston, Levin, and Bongard showed that clusters of frog cells, freed from a developing embryo, can find and combine loose cells into copies of themselves — kinematic self-replication with no genetic programming, no specific evolutionary history for this behavior.³ AI was used to design cluster shapes that replicate more efficiently.

This crosses a line that most artificial life research stays carefully behind. Physarum simulations and NCA textures are unambiguously simulations. Xenobots are actual biological material doing something that looks like artificial life but is made of real cells. The researchers' framing is striking: "life harbors surprising behaviors just below the surface, waiting to be uncovered." The implication is that the rules of biology contain far more possible behaviors than evolution has explored — and that computational methods can help us find them.

Particle Life: Breaking Newton's Third Law

Particle Life is a model that produces startlingly organic behavior from a single violation of physics: the forces between particles can be asymmetric. Particle A attracts particle B, but particle B repels A. This immediately breaks Newton's third law and with it conservation of energy, momentum, and angular momentum. But the violation is the point — it's what makes the particles behave like living things rather than inert matter.⁴

The model works by assigning particles to types (colors), with a table of interaction strengths for each pair of types. Forces are linear in distance: strong repulsion at close range (collision), weaker attraction or repulsion at longer range (interaction). The linear falloff means forces never blow up at zero distance, which makes the simulation stable even when particles overlap. Friction provides global damping to prevent the perpetual energy injection from accelerating everything to infinity.

The resemblance to biological systems isn't accidental. Living organisms consume energy from their environment and convert it to directed motion — exactly what asymmetric forces do, but without the chemical middleman. Particles that attract their food source and repel their predator create chase dynamics that look remarkably like predator-prey behavior. The system sits somewhere between Particle Lenia and classical N-body simulation — simpler rules than Lenia but richer dynamics than symmetric Lennard-Jones potentials.

Lisyarus's implementation runs entirely on WebGPU compute shaders, using a spatial hashing grid (each grid cell is the maximum interaction radius) to reduce the force computation from O(N²) to something closer to O(N). Particles write their grid cell index to a buffer, a radix sort reorders them by cell, and the force kernel only checks neighboring grid cells. With 64K particles and eight particle types, the simulation runs smoothly in a browser — a testament to both the model's computational cheapness and WebGPU's viability as a serious compute API.⁴

The Philosophical Puzzle

The deeper you go into artificial life, the harder it becomes to maintain a clean distinction between "simulating life" and "making life." Lenia — Bert Chan's continuous generalization of cellular automata — produces self-sustaining, self-repairing creatures that move through their environment, interact with obstacles, and can be trained via reinforcement learning to navigate mazes. They're not alive by any biological definition, but the behavioral repertoire is rich enough that the distinction starts to feel like it depends more on substrate (carbon vs silicon) than on function.

This connects to the minimal cognition question from the philosophy side: if bacteria display sensing, memory, learning, and decision-making, and Lenia creatures display similar functional capacities, does the distinction between them come down to chemistry? The panpsychist would say the distinction might not matter at all — if experience is fundamental, it might be present in any system with sufficient information integration, regardless of what it's made of.

For practitioners, the philosophical questions are less urgent than the aesthetic ones. The generative art community that drives much of this work is primarily interested in what these systems look like — and the answer is: astonishingly beautiful. The reason simple rules produce beautiful output is probably the same reason nature is beautiful: both are governed by the same mathematics of pattern formation, self-organization, and emergent order.