Roll a living world. See what evolves.
Red traces: attacks · cyan tails: thrust · gold rings: eating · green arrows: energy gifts · Max advances 24 ticks between drawings.
Attacks & population pressure
Same time window as population history, with separate vertical scales. Rates average each sampling interval. Kills count cells whose energy was depleted by an attack; shield-only hits and other deaths are excluded. Pressure deaths count cells whose usable energy was depleted by a population-pressure hit; their remains can be eaten. Totals start at world creation.
Where the energy goes
Colored bands show actual usable energy gained or spent, with separate vertical scales. Storing and mobilizing move energy between pools; gifts move it between cells. These transfers appear on the relevant sides, so intake is not all newly produced energy. Failed actions spend no action energy; ordinary computation costs still apply.
Execution trace compression
Actual instruction paths from up to 32 living cells, followed for 256 ticks. A fresh sample starts about once a simulated minute.
Preparing an execution sample…
Simple loops also compress well. This measures repetition and ordering, not intelligence or evolutionary complexity by itself. Register values and physics are excluded.
Experimental coarse observer
Behavioral epiplexity estimate
How much structure a small predictor learns from the whole world's behavior. Random colors and the camera view are excluded.
Recording behavior…
Green: behavior · gray: shuffled baseline
What this estimate means
Every two simulated seconds, the GPU records a coarse map of cell density, links, motion, recent activity, energy and reserves. Eleven bounded predictors compete to describe a 32-sample window. The smallest combined model and prediction code wins; the displayed estimate is that model's size in bits. Prediction error is checked on separate, unseen frames.
This is a limited observer, not a universal complexity score or a measure of intelligence. Simple repetition and noise should need small models. Population changes and physics can also create learnable patterns. The shuffled comparison preserves each channel's frequencies while removing spatial and temporal order. Compare scores with the same sampling and model settings.