A lot of AI/ML work is sloppy but well known methods work reliably, textbook descriptions make sense. That there's a long tail of shitty student papers doesn't invalidate the field. Year after year methods genuinely get better. Whatever process produces this progress, it produces this progress.