Watching “2025/09 – How Does the Neocortex Split and Combine Models Without Relearning”..
I feel like all the confusion is arising from the fact that Monty’s implementation is too software-like..
Poses, for example, have to use grid-cells, or grid-cell-like implementation, not 3D vectors or rotation matrices..
Raw sensor data has to be converted to proper feature SDRs. I’m not sure if the current transformation is sufficient.
Object-IDs can’t be integer values either, and certainly can’t be passed around as messages, or “looked up” in a database. They have to be SDRs. That’s how partial matching will work; how new objects are going to get created, merged, split.. it will all happen naturally.
Working with these neural data structures are more difficult but they are the basis of the analog computer that is the brain, and as such, essential, not optional, in my humble opinion 🙇♂️