Creatures evolve brains, bodies, and behaviors through natural selection. You play as Intervention, shaping the world that shapes them. No loss function. No reward engineering. Just survival.
Scroll to explore
Primordial was an experiment: drop organisms into a 2D world with physics and neural networks, let natural selection optimize everything, and watch what emerges. Five parts proved the concept. Predator-prey oscillations, arms races, species dynasties, body plan convergence. All from mutation and survival pressure, no gradient descent, no loss function, no hand-designed behaviors.
Noloss asks the next question. What happens when you give evolution a richer world to explore, and then let a player reshape that world in real time?
A small spherical planet. Creatures with attention-based recurrent brains and hierarchical bodies. A set of context-sensitive actions where the same neural output means different things depending on what the creature is targeting, what it's holding, and what its genetics have specialized it toward. Grazing, hunting, sharing food, courting a mate, teaching a juvenile, caching resources, building nests. All emerging from the same brain architecture and the same action space. The simulation doesn't know what a herbivore is. It doesn't know what a pack hunter is. It just runs physics and selection, and those things appear.
The world is a sphere. Small enough that you can see the curvature from ground level, large enough to sustain hundreds of creatures across six biomes: grassland, forest, desert, wetland, rocky peaks, and water. Terrain is sculpted procedurally, then reshaped by the player. Biomes determine what grows, what thrives, and what dies. A desert selects for drought tolerance. A wetland selects for fat reserves and buoyancy. Paint a forest over grassland and watch the food web restructure as fruit replaces grass and the canopy changes the light.
Day and night cycle every thirty real-time minutes. Seasons shift food availability, weather patterns, and creature behavior across longer arcs. Nocturnal species emerge when the night is long enough to reward them. Migratory patterns appear when food is seasonal. None of this is scripted. The day-night cycle and seasonal variation are environmental pressures. Evolution handles the rest.
Each creature's brain is an attention-based recurrent neural network. Four-head multi-head attention over nearby entities, a GRU recurrent cell for temporal memory, and persistent memory registers that carry information across hundreds of ticks. Roughly 15,000 parameters per creature, all evolved through mutation and selection. No backpropagation. No training loop. No reward signal. The weights that survive are the weights that kept their owner alive long enough to reproduce.
The brain receives a rich sensory stream: nearby entities with distance, direction, emotional valence, and phenotype signatures. A spatial radar for terrain awareness. A cognitive map built from memory: where home is, where food was, where threats lurk. Internal body state: hunger, fatigue, sleepiness, illness, fat reserves. From all of this, the brain produces twelve action outputs. Six continuous signals applied every tick: forward movement, strafing, turning, aiming. Six discrete actions selected by softmax: pull, push, drop, attack, communicate, rest.
The key design: what an action does depends on context. "Pull" while targeting food means eat. "Pull" while targeting another creature might mean groom, share food, or pick up a juvenile, depending on the creature's evolved drives and its emotional relationship with the target. "Attack" while holding nothing is a bite. "Attack" while holding a rock is a weapon strike. The brain doesn't need separate outputs for every possible behavior. It needs to learn when to pull and what to point at. The combinatorial space of target, held item, and genetics creates a behavioral repertoire far richer than the twelve outputs suggest.
Bodies are hierarchical assemblies of typed parts: core, bone, muscle, sensor, mouth, fat, armor, claw, signal, stomach, skin. Each part has a metabolic cost, individual health, and functional contribution. A creature with three muscle nodes and one sensor is fast but nearly blind. A creature with heavy armor and fat is tough but slow and expensive to run. Evolution prunes parts that don't earn their metabolic cost and expands parts that do. The body plan is the genome's most visible product, and it tells you everything about the selection pressures that produced it.
The genome encodes body plan, brain weights, brain topology, digestion efficiency, and over twenty meta-parameters: aggression, sociability, curiosity, nocturnality, camouflage, immunity, longevity, body size, drive weights, and mutation rates themselves. Evolution can tune its own mutation rate. A lineage under strong selection pressure can evolve higher mutation rates to explore the fitness landscape faster, then dial them back once it finds a stable strategy.
Reproduction is both asexual and sexual. Asexual cloning happens when energy exceeds a threshold: clone, mutate, split. Sexual reproduction requires genetic compatibility, courtship behavior, and emotional bonding between partners. The courtship itself is evolved. Some lineages develop elaborate mating displays. Others rely on proximity and opportunity. The simulation doesn't prescribe romance. It provides the mechanics and lets selection decide what works.
Digestion efficiency is a five-dimensional vector: plant, meat, waste, mineral, fungi. Each dimension ranges from toxic to efficient. A creature specialized for plant digestion gets sick from meat. A carnivore thrives on kills but starves in a field of grass. Omnivores pay a generalist tax: decent at everything, optimal at nothing. The food web that emerges from these evolved efficiencies creates genuine ecological niches without anyone designing them.
The player walks the planet's surface. You can pet creatures, feed them, pick them up and carry them. Creatures learn to associate you with food, safety, or danger based on your actions. Bond with a creature over time and it will seek you out, follow you, respond to your presence. The emotional memory system tracks individual relationships: a creature remembers you specifically, not just "the player entity."
The deeper interaction is environmental. Sculpt the terrain like clay: push mountains up, carve valleys, flood lowlands. Paint biomes with a brush and watch plants procedurally populate the changed landscape. Build a desert corridor between two grasslands and watch evolution produce creatures adapted to the crossing. Flood a valley and watch aquatic specialization emerge over generations. Your changes create selection pressures, and evolution responds.
You can archive creatures: save an evolved organism to a file and import it later. Seed a new world with creatures from a previous run. Introduce a predator species to a peaceful ecosystem and watch the arms race begin. The archive system makes evolution's products portable and the experiments repeatable.
The simulation provides drives: hunger, sleepiness, social need, curiosity. It provides actions: move, turn, pull, push, attack, communicate, rest. It provides consequences: eating restores energy, attacking deals damage, communicating emits a vocalization vector. Everything else is discovered.
Pack hunting emerges when social creatures evolve coordinated attack patterns against prey too large for a single predator. Territory marking appears when creatures learn that pheromone deposits correlate with resource defense. Food caching develops when creatures evolve the sequence: pick up food, carry it, drop it near home. Teaching happens when juveniles that stay near experienced adults learn foraging routes faster. Warning calls emerge when vocalization patterns correlate with predator proximity. Gift-giving appears when creatures that share food with kin produce more surviving offspring than creatures that hoard.
None of these behaviors exist in the code as named functions. They exist as patterns of neural activation that survived because the creatures that exhibited them reproduced more successfully than the creatures that didn't. The simulation is a stage. Evolution writes the play.
Noloss is actively in development. The simulation engine, creature brains, evolution pipeline, food web, and Godot rendering are all functional. Creatures evolve, reproduce, specialize, and die on a spherical planet with day-night cycles, seasons, and weather. The player can walk the surface, interact with creatures, and reshape terrain.
This page will grow as the project develops.