Creature 0.6.7: The predict cycle
Creature was going in circles. The first step out is letting it predict what its own light and sound do to its senses.
Creature has been running, and I could not notice anything interesting about it. It mostly influenced itself in circles.
So I looked at what the live system was actually doing. For a full fifteen minutes, in a lit room, the field’s light sense read zero. The strip was busy the whole time, but the change it made at the light sensor was too small to count. Every single tick was being logged as a significant event, because the steady pressure of the weather anchor was enough on its own to cross the threshold. When Creature slept, it was mostly replaying its own background hum.
It was acting, but nothing about its acting came back to it.
My ground principle has not changed. I want it to be the machine that it is. It will never come alive. And it will always be influenced by me.
Version 0.6.7 is the first step out of the circle. Each second, Creature now records what it just did, the light it showed and any tone it played, next to what its raw senses picked up while that was happening. A small model then predicts what its own light and voice should do to those senses, compares the prediction with what happened, and learns from the miss.
That splits every change in the room into two parts: the part Creature caused, and the part it did not. Its own light, once learned, is expected. A lamp switching on is the world. A hand covering the strip is a surprise.
None of this changes its behavior yet. It only watches and learns. In simulation the model learns its own light and voice and leaves the room to the room.
Then I ran a new probe on the real body. It flashes the strip at known levels and measures what the light sensor picks up. The answer was nothing. In a room at 768 lux, every frame, up to the brightest the strip is allowed to show, added less than the sensor’s own noise of about one lux. The ring was designed around Creature seeing its own light, and so far it never has. No software will fix that. The sensor has to move. The sound test was inconclusive, because the room was louder than the tones.
After that, the prediction error goes into the field, so that being wrong about itself is something it can feel. Then a memory of consequences. Further out is the idea I started with: a language model as the dream state, pruning and strengthening what Creature remembers.