ASTRA is our research project on self-learning AI — a model built to remember its experience, judge what worked, and improve behaviour across episodes. Here is the loop that makes it work.
Learning after training
The common recipe for improving a model is: collect data, annotate, retrain. ASTRA targets what happens between epochs of the real world — the model records what it perceived, how it acted, what happened next, and distils that record into durable capability without a full retraining round.
The loop has six stations: perceive, reason, act, record, reflect, improve. Reflection is the only expensive station, and it runs only on episodes the model itself flags as instructive — surprising, costly, or contradictory.
Honesty about scale
ASTRA runs at 20 billion parameters today and is architecturally aimed at a 1T target — we state both numbers because a scale claim that hides the current value is a marketing claim, not a research position.
Follow or challenge it
Every module behind the loop (continual learning, task generation, self-evaluation, memory, skill compounding) is an open research module with its own problem statement on the board. We would rather be checked than admired.