Loading…
1
I think fine tuning gets way too much credit and prompt work gets too little
I watched a team burn 3 weeks and about $4,000 in GPU time fine tuning a model to answer support tickets, and a guy on another team beat their results in an afternoon by just rewriting the system prompt. That said, the prompt version fell apart the moment we asked it about our internal part numbers, so neither side is fully right. My take is that most people jump to fine tuning because it feels more like real engineering, when half the time they just haven't tested their prompts enough. Where do you all draw the line between tuning a model and tuning your own words?
1 comments
Log in to join the discussion
Log In1 Comment
the_lucas27d ago
Man, hard disagree on this one. Fine tuning earned its rep for a reason. Prompting looks great in a demo but it's held together with tape. One small change to the wording and the whole thing breaks. I've seen it happen. And your own story proves it: the prompt guy won the afternoon but lost the part numbers, because a prompt can't teach the model facts it never saw. Fine tuning is how you bake in real knowledge, not just beg the model to behave. The line for me is simple: prompting changes how it talks, tuning changes what it knows.
-1