[PaulHoule]

When you consider the chain from inputs to action you can go far with a small toolbox (e.g. logistic regresssion.)

Plug-and-Play models in visual recognition basically work at this point in time; in NLP they don't.