/moe/

i just mean like you're spending more tokens to try to save tokens
unless this process can be automated
i think it comes down to, like, who is driving
if a human is like "well i want this side question answered" then they can use the tool (someone called it an excursion) which forks the prompt, gets an answer, and then installs it, wasting no tokens on the main branch of the conversation

idk what the state of the art is though
opus 5.5 seems to really know what to do just given a few sticks to rub together
it proactively searches for like existing stuff and uses various bash tools to figure out what's up
i wonder how much of that is the hueg system prompt and harness and how much is its agentic training

well token savings is definitely one aspect, but another is keeping the working context lean
It's kind of just a running issue that at some amount of context, models turn into idiots and stop being able to actually use the information properly
that window can be increased, and is really large, but at some point you do always hit a "we need to compress the whole thing" wall, and those usually lose a LOT of actually valuable information, so it seems like it'd be beneficial to stave that off as long as you can, with low-information thinking processes being a fairly obvious candidate for being excised throughout
>>1191554
I imagine whether it's a system prompt or not, it does come down to the training
otherwise you could replicate it by giving the same prompt to a lesser model

but well-crafted sys prompts can absolutely improve even lesser models

but yea you are probably right that they're just kinda blasting through context to save time
like these frontier models have 1M to play with so it's more about quickly generating a ton of tokens and getting the work done fast than trying to be careful

https://games.mino.mobi/molt/cascade
yo dawg we heard you like snake game