Opinion : 🤔 What Area Should the Next MLOps Startup Work On?

I recently listened to a talk by Dr. Roman Kazinnik from the Meta ML platform team. He shared the realities of running machine learning models at scale, the recent successes in the use of cloud and enhancements in the infrastructure (namely the “machine learning platform”), and also mentioned the “missing link” that still exists while deploying these models at scale.

The success of Meta’s ML platform indeed deserves a round of applause. The engineering team automates model deployment from getting training data insights and labeling to production servicing at global scale within days. This is what a lot of companies dealing with big data are striving to achieve in the past few years. There are also a lot of smaller scale or not-yet-there companies that work on the so called “over-the-wall” approach where the data scientists finish their model development (perhaps in a Jupyter Notebook), and then hand it over to someone else to deploy the models in production.

What struck me as a powerful (but not a surprising) statement:

The author (of the model) will see the actual results of the models they trained by themselves.

What a simple and reasonable request from a data scientist?

Bravos to Meta! But, is it surprising that a lot of the companies are actually not there yet?

Side note: It is one thing to benefit from the convenience of a well-architected infrastructure to support that…