Genesis — when brains learn to eat

Until now, LifeLabs watched shapes. Lenia’s Orbium is a silhouette that sustains itself: beautiful, but passive — it glides because its local rule makes it glide, not because it “wants” to go anywhere. Today the observatory moves from shapes to behaviors. A behavior is no longer just a structure that persists: it is a structure that does something in response to what it senses. And the question becomes: can we get useful behavior without ever writing it down?

What neuroevolution is

Each agent in this world carries a tiny brain: a neural network with 11 inputs, 16 hidden neurons and 2 outputs — 226 weights in total, barely more than a handful of numbers. The 11 inputs are senses: five rays that measure food ahead of the agent, five that detect neighbors, and a reading of its own internal energy. The 2 outputs are muscles: a thrust and a turn. Between them, nothing but a matrix multiplication. No rule like “if you see food, move toward it” is written anywhere.

The twist — and this is why “learn” earns its quotation marks — is that these brains do not learn the way we usually mean it. There is no backpropagation, no gradient, no teacher correcting mistakes. Instead, a blind genetic algorithm: each generation, the agents that ate best reproduce, their weights are copied with a small Gaussian mutation, and the rest die out. This is plain Darwinian evolution, run over 128 brains in parallel on the GPU, for 30 generations. Nobody designs the behavior; selection discovers it.

What this run shows

We coded a stage and its constraints: a toroidal world 128 units wide, food that regrows slowly, an energy cost for every move, death when energy hits zero, and reproduction of the best-fed. We did not code the one thing that matters — what to do with your senses and your muscles in order to survive. That is what evolution had to find on its own.

It found it. At generation 0 the agents wander: their brains are random noise, and their mean fitness — the energy accumulated over a lifetime — is 22.3. Thirty generations later it reaches 41.2, a 1.84x increase, and the curve climbs steadily before settling onto a plateau (the regime saturates: the agents already eat nearly all the available food). But the most telling number is elsewhere. The approach speed — how actively an agent closes in on food from one step to the next — rises from 0.20 in the first generation to 0.51 by the end: food-seeking behavior grew 2.6x. It isn’t the shape that changed, it’s the conduct. The pipeline’s automatic verdict is unambiguous: evolved.

Watch the replay of the final generation on the experiment page, or the full run page with every curve. Frame by frame, the dots stop drifting at random: they converge on the green patches of food. It’s the first time on this site that something looks like intent — with no intent ever programmed.

What it is not

Let’s be rigorous about what we’re watching, because the temptation to overstate is strong. These agents “want” nothing and “understand” nothing. There is no consciousness, no intention in any literal sense: there is a blind optimization, driven by differential mortality, that sculpted 226 numbers until they produced a behavior that resembles seeking food. The contrast with Lenia is sharp and worth stating: in Lenia, complexity emerges from the rule; here, it emerges from selection. Lenia shows that a form can sustain itself; the agents show that adaptive behavior can appear with no architect. It’s less pretty and more unsettling.

What’s next

This little world is flat and its agents are dots. Palier 4 will give them a body: an articulated avatar that must “learn” to walk, with the same recipe — no scripted motion, just a selection pressure and many generations. If dots could “seek” food, we’re about to see what a skeleton can find of a gait. The code for this whole experiment is open; the genetics, the senses and the world fit in a few files, and everything is reproducible from seed 42.