It's all about your goals, just like everything else.
I built this tool in 5 months https://www.padex.io/articles/introducing-padex-patent-parser. I used about 5-6 deep learning models 4/5 of them "off-the-shelf"

The 1 or 2 models that were fully bespoke I learned good stuff while working on them.

I ended up learning quite a bit of pytorch while making those models.

For me, letting business needs be the driver of what I learn is most fun for me, it also seems to have worked.