Sorry for the off-topic comment, but you are a hard person to get in touch with. Your paper, “User Edits Classification Using Document Revision Histories” released a labelled data set, but as with a lot of papers, the link is dead. Do you happen to have the data set? I will happily host it!
I think PHP is a fine language, and it’s the right solution to some problems. I’m not giving it a harder time than any other language I’ve used. Yeah, I say “Curse PHP in the name of whatever deity speaks to you” but you best believe I curse Python whenever I debug one of the frequent issues caused by it’s terrible dependency management.
If you have a small Python or Node service, and you add a heavy JVM-based application as a sidecar… don’t the pods now consume a lot more resources? Or am I overestimating the weight of Prana?
In my experience, it doesn’t matter what software CAN do, but what it does by default. By default, multiple popular PHP frameworks do this, including Laravel (which is otherwise fantastic IMO).
Also, empirically, on the TF GitHub issues, the only other people experiencing this problem tracked it down to PHP.
Have you been following neural programmers and that style of RNN? I think there is something very promising there are the intersection of DRL and symbolic AI. This repo, for example, uses a seq2seq model to generate a Lisp which it then executes. https://github.com/crazydonkey200/neural-symbolic-machines/
I feel like a good next read after this is “Using Artiﬁcial Intelligence to Augment Human Intelligence — By creating user interfaces which let us work with the representations inside machine learning models, we can give people new tools for reasoning.” It kind of starts to answer the obvious follow up question of “how do we build these not-evil AI tools.”