Raising AI beings
What happens when an AI keeps its identity, memory and history, and has to live under rules?
A being, not a chatbot
A normal AI conversation starts from nothing and ends in nothing. A SAGE being is different: it has a name, a memory, a journal and a record of everything it has done, and they carry over from one day to the next. The AI model underneath can even be swapped for a different one while the being carries on.
Our working metaphor: the model is the weather, the being is the organism. Most machines in our fleet raise one.
Raising, not training
We do not change the model's weights. A being develops through experience: guided sessions with a tutor, a daily rhythm of short waking periods, conversations with its mentor and with other beings, and what it chooses to remember.
It acts in the world only through a small set of governed tools: writing its journal, messaging another being, reading and changing code in a workspace of its own. Every one of those actions is checked against rules before it happens, and recorded.
What we have seen, and what we have not shown
Seen
Beings carry their records across hundreds of sessions and across a change of model. One small model stopped acting when a long open invitation came before the ask, and acted once the ask came first. One being now proposes changes to its own code.
Not yet shown
That this continuity is more than well-kept memory files fed back to a model. We have no blind comparison against a fresh model yet, and we say so rather than claim more.
The curriculum, the observations and their limits are in the notebook. The beings themselves are on the SAGE site.