Alternatively current AIs don't compose well with each other and don't let themselves be decomposed easily. This makes control and user interaction difficult to

Alternatively current AIs don't compose well with each other and don't let themselves be decomposed easily. This makes control and user interaction difficult to establish which leads to monolithic one-click input->output products by default. Things that can be stated simply ("replace the banjo with a saxophone", "place the cat to the left of the table") can't really be tackled by composition (comfyui workflows of hell) and if you want something that generalizes, you need to train models specially designed with that composition scheme in mind (layout diffusion for instance).