Current skills of LLMs are a mirage of human projection; solving AI reasoning requires AI architecture innovation; training and inference will converge; static transformer inference will commoditize.