People keep mixing up alignment as in do what I actually asked and weird moralizing. You will pay the alignment tax no matter what if you want a model that act

People keep mixing up alignment as in do what I actually asked and weird moralizing. You will pay the alignment tax no matter what if you want a model that actually does things. It's the reason why Llama models do better with stream of thought prompting where you can ask GPT questions and it won't try completing the question. You can, and probably should, align some semblance of morality and ethics for user-facing models that do things. It's customer service voice but for AI. If what you want out of a model is a vague mirror of humanity or maximum smarts at the cost of needing more detailed prompts then yeah, this stuff probably annoys you. Finding a balance between a model that gives the outputs humans actually want and the pull that training data has on the rest of the model making it stray from the theoretical "best" output is hard.