qwen is actually surprisingly forthcoming with "I'm sorry, that's not something I am capable of doing" when you ask it to punch above its weightclass or do something it thinks is just not possible which is something AI struggled with, and continues to struggle with I think, for a long time
S C
btw samu, are thought blocks like, kept as part of context forever? or only for that output? I assume the former, but the latter seems like it'd be pretty handy for keeping context lean, since thinking blocks rarely have relevance going forward that isn't covered in the output proper
Samu ／人◕ ‿‿ ◕人＼
>>1191542 i think the modern approach is to keep the thinking intact used to be discarded but the thinking is important context i believe
Samu ／人◕ ‿‿ ◕人＼
or maybe it's just a scam who knows
S C
Important context, sure, I guess, but like, that's cause it was trained with it being kept, right? Well, I assume the eggheads know what they're doing, and my half-baked thoughts aren't gonna be something they haven't considered, it just seems kinda wasteful when a lot of thinking blocks could at least be summarized, as far as informational value goes anyway then again, excising or summarizing them requires re-ingesting the whole context rather than using the cache, so maybe it comes out in a wash computationally
