The article describes how tools like GGML and its derivatives (whisper.cpp, llama.cpp, etc) represent a shift away from complex, general-purpose frameworks (such as TensorFlow and PyTorch) to simple, single-purpose codebases that are easy to install, understand, and modify, which sounds like a good idea. But why then the complex frameworks were required at all?

I think the frameworks became complex over time but were initially simpler because of the more managable set of AI workloads.