Trying to replicate or reimplement AI results these days kinda pisses me off, because the paper will give a nice clean algorithm block that doesn't even slightly work as written, whereas the code has half a dozen Deep Learning Tricks that make all the math technically extremely wrong.
