> There's nothing stopping you from composing things in Python. There is, and you pointed it out yourself: > I don't see why you would compose a TensorFlow mo

> There's nothing stopping you from composing things in Python. There is, and you pointed it out yourself: > I don't see why you would compose a TensorFlow model with a PyTorch model without incurring tons of overhead in your application around gluing these two things. This is where Julia is profoundly different. Reusing stuff from different libraries is trivial compared to Python. You could take an activation function made for one libraries use it without any modification or wrapping in another ML library. These is where Julia will win long term. If you look at the ML libraries in Julia they are tiny. The Python side of things require gargantuan effort because things are not composable. The wheel is reinvented over and over again.