I had completed the Keras version of course 1 last year. It was awesome, I also learnt from neuralnetworksanddeeplearning.com and Karpathy notes which fast.ai linked to. Got a good beginning in understanding and applying deep learning.
When I wen through the notebooks of the new course 1, I had experience using pytorch etc, felt the API was too abstracted making things unclear unlike old course 1 materials.
There is a effort going on to refactor the fast.ai library for v1. You can check it out here and forum discussing the new API. But I think some of the existing conventions (of the new pytorch fast.ai) like variable names, formatting etc are there to stay.
If I abstract that any further, I might as well use an AWS machine learning API.
I differ a little here. I find the abstractions in AllenNLP library(https://allennlp.org/) really useful for fast iteration, experiment design and replicability. It is a library that builds upon pytorch. It provides a nice way for building reusable parts of the NLP pipeline for deep learning. I think it can be extended to other problems trivially though focus of the current components are on NLP.