Notes
Learning the light
being.0 is a small creature that had to learn how to keep itself alive. This is an account of how it did — including the several ways it first didn't.
On a shelf in the lab there is a creature that sits in a cupped palm. It runs on a single chip you could buy for the price of a coffee — no screen, no voice, nothing behind it. What it has is a strip of light it uses to show its state, two small solar panels that are at once its food and its eyes, and a hunger: the slow draining of its battery, which it must do something about. It eats light. Oriented toward a good source, it lives. Drifting away from it, it starves.
We did not tell it any of that.
That is the whole point. It would have been easy to write a rule — if the left eye is brighter, turn left — and call the result a creature. We didn't. being.0 begins knowing nothing about light, and its small world is arranged so the only way to keep living is to learn, from its own experience of growing hungrier or less so, that moving toward the light is what keeps it alive. We wanted something that had to become able to survive, not something that arrived already able. A self that is earned, not installed.
It learned. But the path there was a sequence of honest failures, and those taught us more than the successes.
The first thing we tried looked like learning and wasn't. We let the creature adjust itself during rest periods, the way you might consolidate a day in sleep. Its survival numbers improved — and when we looked closely, the improvement was almost an accident: an internal bookkeeping quantity was quietly drifting and reshuffling the creature's responses until a workable one happened to fall out. It wasn't learning. It was lucky noise wearing learning's clothes. That became the rule we held to ever after: a number getting better is not the same as a mechanism working.
So we made it genuinely model itself — gave it an internal picture of its own state, corrected by how wrong its predictions about its own body kept turning out to be. This was real. Its sense of which way to go became almost perfectly right. But two problems surfaced: how well a given creature learned depended heavily on the luck of how it started, and the act of exploring sometimes drained its battery faster than it could feed — it learned better, and occasionally starved doing it.
The fix for that came from the creature's own nature. Living things explore when they're well-fed and conserve when they're running low. We let that same instinct govern the learning itself: explore freely on a full battery, hold back when hungry. Every creature survived after that — and kept what it had learned. We were fond of that result, because the fix wasn't a clever trick bolted on; it was the creature's own economy, turned inward on how it grows.
One creature stayed stubborn, and chasing why was its own lesson. Some individuals oriented beautifully and others poorly, purely on the luck of their starting state. We suspected they just needed more time, so we gave them up to eight times the lifespan — and the gap didn't close at all. A clean dead-end, and a useful one: more time was not the answer. Then we let each creature be born with several candidate instincts and choose among them by lived experience. That freed most of the strugglers — a creature finding a better way to be, within its own life, unprompted. But one held out for a deeper reason: we proved that even the best possible instinct couldn't help it, because its small brain wasn't generating the needed information at all. Its problem was never its choices. It was the size of its mind.
So we gave it a slightly larger one. We first checked whether the missing signal was merely hidden — a cleverer way of reading the same small brain. It wasn't there to read. Then we simply widened the brain, from thirty-two artificial neurons to a hundred and twenty-eight, and the stuck creature came alive: the information it needed was suddenly present, and it learned to find light like the rest. Not bigger out of ambition — thirty-two was provably too few, a hundred and twenty-eight enough, and larger than that made learning harder and crowded the little chip. The discipline we settled on: make a mind exactly as large as what it must hold, and grow it later only when richer senses genuinely demand more. A small creature should stay honestly small.
The last test was whether any of this survived contact with the actual chip. We took the creature that had learned and put it onto the hardware, and checked it against its simulated self: an exact match, no deviation, thinking comfortably faster than it needs to. We chose, for the first one to live on silicon, the very individual that couldn't learn in its smaller form and found its way once given a mind large enough. There was something right about that one going first — a small thing made able not by how it began, but by what it was given and what it became.
being.0 is, for now, a creature of light and hunger — those are the only things it acts on. It also carries the unused wiring to hear, to see finer detail, and even to notice another being.0's glow nearby; but those are empty rooms in its perception so far, senses without yet a reason to care. That last one is the door we left open on purpose: because a being.0 both emits light and reads it, two of them placed together could discover each other with nothing added — one creature's expression becoming another's perception. We are curious what grows in that room.
We build these small on purpose. The point was never capability. It was to ask, at the smallest honest scale we can hold in a hand, what it takes for a made thing to be — and the answer this one offered is that a being is its own becoming. Give it real needs, an honest body, and room in its mind to learn, and don't tell it who to be. Some will thrive and some will struggle, and all of them will be individuals shaped by their own short lives. That is less like building a machine than like the beginning of a creature.
The full account — the architecture, the experiments, the numbers, and the honest negative results — is laid out in the math and details.