Little Creatures

The 1985 design. Each creature has a tiny feed-forward network: 3 inputs, one hidden layer of 3, 2 outputs — fully connected, sigmoid units, no learning rule inside the brain. This predates backpropagation's popularization, so nothing ever computes a gradient. The world grants tunnel vision: the first non-land thing in the 8 hexes straight ahead. Two inputs binary-encode what it is (00 land, 01 water, 10 food, 11 creature); the third says near or far. The two outputs are thresholded into bits and decoded as an action: 00 forward, 01 left, 10 right, 11 eat.

Mark II — the modern redesign. Same creature, same island, same physics — different representation. Vision widens to three sectors (ahead-left, ahead, ahead-right), and each sector reports what it sees as a one-hot vector — one input per category, so "the food wire" means food and nothing else — plus a continuous distance signal (1/d). A sixteenth input is the creature's own hunger, so behavior can finally depend on internal state: the same view can produce different actions when starving. The output side is equally modern: four units, softmax, one per action — no bit-decoding. The 1985 binary code secretly told the network that land and water are similar (they share a bit); one-hot tells it nothing it has to unlearn. That's the lesson: representation choices are part of the model.

Acting. Eating only works on food in the very next hex ahead — walking onto food tramples it, nourishing no one. Stepping into water drowns the creature; hunger reaching 100% starves it; surviving 1,000 steps retires it with honors.

Evolving. Sixty-four creatures run one at a time. The 8 longest-lived become parents, and every pair — including a parent with itself — produces one child: each weight is copied from mother or father at random, or, at the mutation rate, replaced with a fresh random weight. Watch what happens to this algorithm when the genome grows from 15 weights to Mark II's 120: blind search suffers in high dimensions — a large part of why the field wanted gradients.

The landscape. The landscape panel draws the actual fitness surface these brains are climbing. We hold every weight fixed except two — or sweep any 2-D plane through weight space — running a complete simulated life at every pixel and painting it by how long that life lasted. The dots are the last generation projected onto the same slice. In Mark II the space is 120-dimensional, which makes the caveat that "we can only ever see slices" exactly as true as it is for real deep networks. Click any point to incarnate it and see how it behaves.

The rival. The chart panel can add a backprop learner: the identical network trained by policy gradient (REINFORCE). It samples its actions, and after each episode the chain rule propagates credit backward through the net; weights move a little uphill along their gradient. It gets 64 episodes per generation, the same budget evolution gets in lives — plus two advantages: a shaped reward (a bonus at the moment of eating) and per-step credit assignment. On the 1985 creature the race is close. On Mark II, gradients pull far ahead — richer representations need a learner that can exploit them, and that pairing is, in miniature, the story of modern deep learning.

What to watch for. Generation 1 mostly walks into the sea. In 1985 mode a lineage soon discovers "turn at water," then "eat when food is near." In Mark II mode, watch the backprop creature learn to steer — food glimpsed in a side sector bends its path — and check the hunger wire in the brain panel while it happens. Switching designs keeps the old learning curves as faint ghosts on the chart, so the comparison is right there.