/sci/ - Science & Math

>>12367342
>i thought a lot of ML people are just SWEs now

I wouldn't agree with this statement. Things are still evolving pretty quickly, but ML Engineering and ML Ops are becoming their own thing.

There's a lot that goes into building a production level machine learning system beyond the algorithm, and setting it up right required ML knowledge that SWEs generally don't have.

A lot of it is related to automating the ML algorithm lifecycle. You need systems to actually serve your model, so that involves building auto-scaling inference systems and monitoring to catch regressions. You also want to be constantly collecting new data, processing it, retraining models (either online, nearline or offline) and serving new models. This entire system should be automated. And of course you never have just one model, you have this same process and automation occurring over dozens of different ML algorithms.

There's of course overlap with traditional SWE and devops, but ML deployment has its own unique set of problems, tooling/infra and skill set.

>>11205277
Training models is really only a small part. Most of the work is split between building pic related and working on algorithm design. Particularly when you need to build models for data you don't have a cut and paste template for.