Meanwhile in the real world they can’t even figure out the basics like containers and Cicd which we’re actually dealing with in our actual contract.
I worked on nudity detector in 2017. Deep learning works, and is useful. Although you are right it's very specific and quirky.
I found "How HBO's Silicon Valley built Not Hotdog" article very interesting, because it's basically the same problem. They found MobileNet better than SqueezeNet and ELU better than ReLU. You know what? We found SqueezeNet better than MobileNet and ReLU better than ELU for our problem and data. Who know why.
https://medium.com/@timanglade/how-hbos-silicon-valley-built...
It's absolutely human nature to think of crazy ways of using things in new ways, probably most of those ways don't work out in the end.
It's pretty clear to everyone pushing this is being dishonest. Sometimes I wonder if they intentionally being dishonest or if they just don't know what they don't know. Very few people are doing useful true machine learning, and the applications are very specific with its own set of quirks. It's just to make some quick money and get out.
What somehow is never being said with all this hype is that you need to hire good software devs to make a great solution to a problem. All these buzzword driven ideas tend to confuse a bunch of people and die out after wasting everyone's money. In my region I keep seeing a bunch of business suits pushing X idea being solved with Y with no justification or engineers behind them.
Yesterday I have found out I have to have yet another meeting to explain why "blockchain" does not translate to a full solution given to a very superficial poorly thought out problem.
The only reason I've had a voice in my region with all this noise is because I've made things people see that actually work.