Home | Notifications | New Note | Local | Federated | Search | Logout

Note Detail

Reply to @sun@shitposter.world
(eval 'Toast)@toast@donotsta.re (2026-08-06 17:03:30)
@sun I do actually agree with this, my problem is kind of elsewhere. "Prompt engineering" is bs, but *everything* is a skill you can be better or worse at, including prompting.
The problem is that models work disparately, so things that improve semantic alignment in one might not in another (there's some research to suggest that embeddings are transferable spaces, which would imply that you could apply a mechanical transformation to the prompt to get the same result, but we don't have this yet).
On top of that, with most things, as you do the thing more you get better at doing the thing, but because of the sycophancy this is attenuated here: getting better at prompting by prompting more is inefficient.
Then you have all of the neurological issues that are associated with the entire category of technology, where as you get better at prompting (less quickly than at most other things) you also get worse at the actual things you're trying to do, at a greater rate than you improve at the prompting (for instance, persistence is greatly diminished even after 10m of usage, while being a huge predictor for outcomes, to the point that lately one of my focuses has been persistence).
So I do agree that you can be better at using these things, but there's so many attenuating factors to this that it doesn't feel like it would ever be worthwhile, at least to me, even if it was actually free (rather than exploding in costs as they try to burn money less quickly).
---Reply---
Blurry Moon@sun@shitposter.world (2026-08-06 17:06:57)
@toast I am pretty sure you know that people in software also have to get good at prompting clients to get the clear answer too and it is inefficient and annoying and takes away from real work.