I have been messing with some of the smaller models (think sub 30B range), and getting them to do complex tasks. My approach is pretty standard: take a big problem and get it to break it down into smaller tasks. They are instructed to create JavaScript code that runs in a sandbox (v8), with custom
