That’s correct. Any CUDA related code will have some sort of C dynamic loading under to hood, and in the early days, Caffe directly uses C++ libs. However, interfacing language or “glue code” as you mentioned plays more important role than “gluing”. As research shown, a developer can produce a fixed line production code annually regardless of language used. The choice of not using C++ for ML experiment is not due to researchers being “not fancy”, but rather to speed up the progress. 1000 lines of C++ machine learning code can barely read in the dataset, but 1000 lines of Python can train you a pretty good model and maybe get you the result for a paper even.