Fwiw I think this is not true right now. It might be true in the future, I guess. Today you have to learn some pretty specific techniques and they don’t come naturally to most engineers I think. They are accustomed to one-shot Google searches, not prompt engineering. Yes, you can learn how to prompt engineer, but it takes trial and error. A lot of my co-workers today still believe the LLM is useless because it hallucinates API definitions. They very much are not used to tools that require the level of hand-holding you need today.
Either teaching them is going to get easier, which is somewhat plausible actually — or you think LLMs are going to get dramatically easier to prompt with more reliable output. I actually think that’s somewhat implausible. Log(compute) is a pretty hard number to move past a certain point and it seems we’ve hit that point already. RLHF is great, but it doesn’t scale well and the result is that some tools are hard to build. Giving me millions of context tokens is neat, but if it’s wildly expensive and confuses the model on top of that, it’s not actually useful.
I do think there is room to improve and of course there are new techniques to be discovered. The models may even help discover some of them, although I am skeptical that they will find Everest by hill climbing when they’re currently on Pike’s Peak. Or maybe we really need to find Olympus Mons.
