I wrote a custom HashMap for Java, used BFS and DBS on graphs, wrote a top down custom parser, custom string searching algorithms to solve real business problems. I would rather work with someone who knows how neural networks really work rather than who knows pytorch or keras, because they can be learned rather easily. In some industries knowledge of algorithms can be more valuable than knowing a myriad of frameworks that change every few years any way.
> In some industries knowledge of algorithms can be more valuable than knowing a myriad of frameworks that change every few years any way.
This is exceedingly rare... so you wrote a HashMap implementation in Java?? None of the custom maps discussed [here](http://java-performance.info/hashmap-overview-jdk-fastutil-g...) were enough? It must be a really rare problem you're working on if there's no library that provides an acceptable solution for it in Java.
> I would rather work with someone who knows how neural networks really work rather than who knows pytorch or keras
I would rather work with someone who knows general programming (which is not the same as algorithms - there's so much more to it!) and software design... but if working on a product that uses a framework heavily, good knowledge of that framework can be the difference between getting things done in minutes VS weeks (of painful and time-consuming battling with the framework to do thing the way you think they should be done, rather than how it's done within the framework).
I agree that these problems are rare, but they do come up, at least they did in my experience. As you alluded general programming and software design is not the same as knowing how a particular framework works. The former is much more valuable, but at the same time if someone is able to understand in depth a complex technology/framework he/she is most likely great at general programming and software design.
