You're generally tied to proprietary nvidia drivers that are a pain to install, and when they change version, it breaks software such as PyTorch, you're forced to update. PyTorch itself often has breaking changes. Then, most of the code is written in Python, often by academics who know and care very little about good software engineering practices. That also leads to fragile and breakable code. You'll often find yourself trying to install code that's just DOA because of multiple broken dependencies.
The quality of the code can be terrible, because academics are in a constant rush to publish the next thing, and tend to view code as throwaway. Worse yet, it's often written by students, and these students are taught by professors who view software engineering as an inferior, unscientific discipline. I've seen many github repos with no open source license, and no README, just a random code dump with no comments.
I'm personally not super enthusiastic by the idea of replacing a lot of code by machine learning just yet, because not only is the software ecosystem fragile, but the models can be as well. You're likely to introduce weird, unpredictable modes of failure. I can see that there's clearly a lot of people who dream of being able to basically create software without coding, and not have to deal with pesky software developers, but IMO, we're not there yet.
This guy talks about cloud-based machine learning solutions. You won't have to deal with drivers and general "IT" concerns there. The problem is that, besides being very expensive, these cloud-based solution will introduce extra layers of vendor lock-in. If you're training on the google cloud and using TPUs that you can't even physically buy/own, you might have a hard time porting your code to run without a TPU. If Google revokes your access for some reason, because they claim you've violated their TOS, you are fucked.