Fairly interesting indeed, he had published a livetalk developing those ideas : https://www.youtube.com/watch?v=06-AZXmwHjo This raises a very good question about the definition of the MLOps. As in our team, operations were mainly about deploying & exposing models, with metrics monitoring, Andrew Ng proposes to add 2 types of activities to the discipline, as I understand :
Maintaining AI systems in operational conditions, mainly with data observability
Continuous improvement, with annotation services, feedback loop to help identifying true labels & continuous training.
This is an exciting point of view, as it involves a lot of automated operations around models management.
