My initial idea was to train a navigation decision model with 25M parameters for a Raspberry Pi, which, in testing, was getting about 60% of tool calls correct. IMO, it seems like around 20M parameters would be a good size for following some narrow & basic language instructions.
Ok. This makes me wonder about a broader question. Is there a scientific approach showing a pyramid of cognitive functions, and how many parameters are (minimally) required for each layer in this pyramid?