GameOfLife

A persistent simulated 3D voxel world inhabited by artificial organisms that live continuously, interact with the world and each other, and learn lifelong from their own experience. A more "real" Game of Life.

There are no episodes, no resets, no tasks, and no fitness functions. One world runs for days; robots forage, dig, build, signal, hibernate, and die in it; each learning robot carries its own predictive world model trained online from its own life. The brains can develop reusable temporal skills while feeling only endogenous consequences of their bodies and predictions. The point is watching what emerges.

The research

Despite the name this is a research platform, not a game — though like Conway's original it's a zero-player one: nothing to win, everything to watch. The overarching question: what behaviors, traits, and systems emerge in a world with no fixed objective? No designer assigns task rewards, skill labels, demonstrations, pretrained behaviors, or a fitness score. There are bodies that need energy and can die, minds that predict, feel, explore, and develop reusable skills, and a shared world that other minds keep changing. Whatever shows up — survival, sociality, signaling conventions, mating, culture, or something unanticipated — has to grow from those mechanisms.

Concrete questions structure the long runs (in full: research questions). Four are about what an agent learns and does:

1. Lifelong, non-episodic world-model learning. The world-model literature (DreamerV3, Plan2Explore, TD-MPC2) resets its agents thousands of times into a stationary environment. Can a Dreamer-class agent learn through one unbroken life in a world that never holds still? Watch for plasticity loss and forgetting of places not seen for a sim-day.
2. Intrinsic motivation in a shared world. To a curiosity-driven agent, the most unpredictable object around is another curiosity-driven agent — uncertainty about a moving target never fully resolves. Mutual fascination? Avoidance once modeled? Chasing? An ablation flag masks other agents out of the curiosity signal to compare.
3. Emergent communication. The 2-channel broadcast costs a little energy and means nothing by design. Does it acquire meaning under survival pressure — bursts at food discoveries, distress near death, silence at night?
4. Cultural transmission. When newborns are warm-started from living agents' weights, do behaviors — foraging routes, hoarding sites — propagate across "generations"?

Four more are about what the system becomes over time — these the long runs surfaced on their own rather than being planned, and they answer the "no goals" question most directly:

5. What keeps a mind motivated for a whole lifetime? The project's central question, discovered in the runs: curiosity is self-extinguishing — master the world and there's nothing left to be surprised by — so how does motivation stay alive across one life, handing off between drives as each satisfies itself?
6. Where does individuality come from? Identical minds with identical drives, in a shared world, have already diverged into distinct and persistent personalities. Do they have to, and what carries the identity?
7. Does the world itself evolve under the population? A persistent world shaped by adaptive agents generates selection pressure back at them — an accidental plant-defense ecology already appeared under grazing pressure, with no ecology designed in.
8. Can natural selection arise without a fitness function? If reproduction becomes a bodily process funded by an agent's own energy surplus, does a population under no fitness function develop real selection and a cross-generational competence ratchet?

Every long run is attributable to at least one of these, and every round of runs gets a written-up finding in the research journal — negative and confounded results included, since the entries are the durable record after a save is pruned. The early rounds already produced surprises: an accidental plant-defense ecology under grazing pressure, curiosity collapsing once the world became predictable, and behavioral individuality arriving before survival competence.

Quick start

uv sync
uv run gol-run saves/alpha --new          # create a world and watch it in Rerun
uv run gol-run saves/alpha --resume       # continue where it left off
uv run gol-ctl pause                      # control a running world (speed/checkpoint too)
uv run gol-stats saves/alpha              # dig through metrics and events
uv run gol-stats saves/alpha --compare    # are the dreamers pulling ahead of chance?
uv run gol-stats saves/alpha --interests  # do agents differ, and stay themselves?
uv run python scripts/aion_stats.py saves/aion_01_2gpu  # readable Aion check-in
scripts/soak.sh saves/soak_001            # overnight run, restart-on-crash
scripts/provision_runpod.sh root@gpu-box saves/alpha   # ship a world to a cloud GPU

The shape of the thing

• World: finite Minecraft-like voxel world (terrain, water, ore, food bushes, day/night). Blocks are diggable and placeable. Physics is Minecraft-style AABB-vs-voxel — rich and modifiable, but cheap.
• Robots: wheeled bodies with color raycast vision (depth + shaded RGB — what a block is has to be read from how it looks) and a steerable gaze, plus a gripper, touch, an energy store, and a free 2-channel signal broadcast. Food restores energy; night stops regrowth; running out means hibernation, then death — which drops scrap back into the world.
• Brains: pluggable. Scripted baselines share the world with three research lineages: beta's Dreamer/GRU world model, Aion's long-context S5 world model, and anima's world-model-free plastic network. World-model organisms use imagination-trained critics, endogenous affect (interoception, curiosity, boredom, predicted mortality), and optional learned temporal skills. No task reward, named skill, demonstration, pretrained behavior, or fitness score exists in the learning path.
• Observability: Rerun — live 3D scene, per-agent charts (energy, curiosity, prediction error), scrubbable timelines, recordings.
• Compute tiers: 1–2 learning robots locally (Apple Silicon); populations of 8–16 on a single rented cloud GPU. Worlds checkpoint atomically and resume anywhere.
• Minds outlive bodies: with inherit_weights: lineage (what the standard run configs use), a learning brain's weights and memory carry over to its respawned body — death is costly, but the lineage keeps learning. none and random_living exist for the cultural-transmission experiments.
• Experiments: configs/run/exp_*.yaml are pre-registered protocols for the research questions (social curiosity with/without agent-masked curiosity, cultural transmission across inheritance modes). Long-run rounds get their own save-name-prefixed configs (beta_NN_*.yaml) that freeze at launch, so every save dir is a reproducible experiment.

Docs

• Architecture — full design: world, brains, runtime, milestones
• Research questions — what this platform is built to probe
• Research journal — findings from long runs, one entry per round
• Research proposals — designs argued out before they land
• Configs — how config layering and frozen round configs work