The other replies are spot on but in general -- Alignment induces cost, no matter WHAT type of alignment, even 'neutral' versions like instruct for chatbots Th

The other replies are spot on but in general -- Alignment induces cost, no matter WHAT type of alignment, even 'neutral' versions like instruct for chatbots The 'common' example that people reference is, for instance, making the model non-offensive means it has to spend.. something... on that directive, if you will. (which also led to absurdities such as refusing to help people with bash commands that involve having to ```kill``` a process...) But even for the purpose of an instruct model (this is what peeves me off), making it answer questions and take instructions makes it *worse* at many creative tasks because you're constraining it's behaviour to Q&A -- though this is a long tangent...