@mikeev@__roycohen How "compressible"/verifiable is the output you care about? I feel like if I don't handhold the model at least a little bit it goes off and does something stupid or lazy.
@0x4C756973@darinmao_ Failure mode to the "let's add a test to make sure the *thing that should have never existed and should leave no trace* doesn't exist" situation.
So, going back - how has it worked for you? Do you use any specific wording?
@0x4C756973@darinmao_ And dragging it into situations that have nothing to do with it. Worse if it's something that has a bit more nuance and that the models don't seem to knack for, like "who's seeing this and what do they need to get out of this" - which in my head is a similar... 2/n
@Algon_33@voooooogel Have you ever used Opus 4.8? My exact reaction seeing the comic was "that's absolutely true and it drives me mad as well". I'm curious how you got anything other than that, if you do interact with it "iteratively"/back and forth
but this is also a fantastic oportunity. we now kinda know the shape of the beast, it's strengths and weaknesses.
we can now focus on properly integrating these new tools into our engineering practice. there are a lot of low hanging fruit of drudgery we can delegate trustfully to the models, if we also design the needed guard rails around them.
yay, engineering!
@guelldemarko Dica: pede sempre links e o modelo as vezes tem a tendência de só concordar com o que vc fala, então se vc falar alguma coisa com "mt confiança" às vezes o modelo alucina alguma informação pra concordar com vc. Infelizmente os modelos são pick me kkkkkkk