Also the pretrained LLM (the one trained to predict next token of raw text) is not the one that most people use A lot of clever LLM post training seems to stee

Also the pretrained LLM (the one trained to predict next token of raw text) is not the one that most people use A lot of clever LLM post training seems to steer the model towards becoming excellent improv artists which can lead to “surprise” if prompted well