With CA3 an open model keeps learning after deployment, and its owner decides what it learns, what it refuses and what it forgets.
We have run it on 23 open models, from 22 million to 122 billion parameters, in 19 public cases: language, satellite imagery, a robot arm, a brain-implant decoder, image and sound generation, model edits and more. Every result comes from a recorded, pre-registered test, published with how it was run and its limits. The mechanism is not shown. Patent pending.
All cases: the CA3 cases Space and ca3.si.
Model and dataset names are for identification only and do not imply endorsement by their makers. CA3 is developed by Alberto Acedo, PhD, at Biome Makers Inc.