ai "breakthroughs" always seem to be the same thing - someone finds a way to make a tiny subset of an existing model work for a specific task.