Elie Bakouch of Prime Intellect and swyx outline adaptive entropy control and length penalties to optimize LLM reasoning traces · Digg