How would the language of parametric ecologies look like?

Ingredients:

• learned maps (directed solid arrow): ML models being trained
• frozen maps (directed dashed arrow): frozen ML models / non-ML maps
• distances (bidirectional solid arrow): divergence, distance
• optimization pressure points (+/- bubble): gradient ascent/descent
• (modeled) distributions (numbers with optional hats): input / output / misc distributiosn

Design choices:

• Numbered distributions over lettered ones because one model's output is another model's inputs in interesting arrangements. Over named ones because names create clutter.
• Name learned maps by letter on tail. Shared weights translated to identical names across arrows.

whiteboard