AMA model family
AMA / LOCAL AI RESEARCH
Learn the world, one verified step at a time.
AMA explores a lightweight local assistant that can report its limits, use bounded tools and learn from checked observations. AMA-1 Genesis v0.2 is a working research prototype. The final capability target remains blocked.
Current experiment: AMA-1 Genesis
Research directions
From scratch
Clean training data, random initialization and checkpoint ancestry. No inherited pretrained text-model weights.
Self-model and personality
Report identity, tools, failures and limits. Calm, rational and correctable behavior; subjective consciousness is not established.
World and objects
Typed textual properties, relationships, sources and explicit corrections. Perception remains future work.
Reviewed learning
Observations become candidates. Approval, separate evaluation and rollback belong before model activation. The full continual-learning loop is unfinished.
Lightweight tools
Bounded calculation and local knowledge tools work now. The 6GB VRAM budget is a future target, not a verified GPU result.
Trustworthy networking
Search is denied by default and requires explicit consent. Results remain untrusted links; independent-source and GEO verification are unfinished.
What the evidence supports
AMA development notes
Building AMA-1: a clean baseline and two rounds of unseen tests
AGY implemented, Codex reviewed and later contributed separately identified patches. All 85 regression tests pass, while unseen routing accuracy remains 37.5% and 33.3%.