Great post from a bigwig at Google about the difference between building the infrastructure of machine learning and applying that infrastructure. I think that just like the vast majority of companies shouldn't run server infrastructure, the vast majority of companies shouldn't be involved in building the nuts and bolts of ML systems. Instead, use one of the big cloud providers or open source frameworks until your needs exceed the offering.
This article really resonates especially cooking innovation vs cooking appliances innovation analogy .
I've been working with a buddy on some analytics service products for biotech/life sciences (one is called Yukon Data Solutions). What we've found looking at other products in the space is they often require the customer to focus a lot on processes, tooling, infrastructure - and so some degree algo research - in order to use the product.
What is key, we believe, is a focus on the analytics recipes and available algo ingredients in order to support decisions: how are the analytics recipes and resulting reports going to help decide on the next steps for the customer's business?
Whats seems to be background / secondary? explaining the tools, libraries, languages, cloud infrastructure, automation, data lakes etc that a service may or may not use behind the scenes to achieve the goal. Are these needed - yes at different levels and different times depending on the customer - but focus on the recipe, ingredients and decision support seems key.
edit: I found Eugene Dubossarsky's thoughts on Decision Support interesting in this podcast : https://anchor.fm/datafuturology/episodes/1-Dr-Eugene-Duboss...