I might disagree on this. The software engineering behind production machine learning systems can be quite interesting and nontrivial. It really depends on the scope of the challenges being faced. If you have thousands of models that need to be served in production and continually retrained and monitored, that becomes a pretty sophisticated problem space to work in.
Yes, however most ML engineers don't get to work at Jeff Dean's level to actually do such interesting work. There are very few companies willing to write their own Horovod or distributed PyTorch.
