current ai research is 95% just reapplying and tweaking existing models to new tasks, and 5% actually pushing the field forward. when's the last time you saw a genuinely novel architecture?