I think there is room for criticizing a lot of the hype around deep learning papers, especially the semi-blog / semi-research stuff you often see in tech company blogs, fastai, etc.

But this criticism falls a little flat to me. For instance,

> “Nevertheless the central point of a scientific paper is a relatively concisely expressible idea of some nontrivial universality (and predictive power) or some nontrivial observation about the nature of reality”

That’s an insanely high bar for published work. I also read lots of research papers, and I think only a handful per year would meet these requirements. Yet many others are extremely valuable to show negative or partial results, results with small effect sizes, and other things.

We absolutely should not disparage someone for publishing results of a failed or ineffectual approach. Because otherwise we’ll just make things like file drawer bias and p-hacking far worse, and create an even worse cultural expectation that to make a career in science, you must constantly publish positive results with big, sexy implications — which is what leads to the whole disastrous hype-driven state of affairs, like in deep learning right now, in the first place, and ludicrous science journalism, funding battles fought over demoware and vaporware, academics fleeing into corporate sponsorship like yesterday’s article about Facebook, etc.