Post

JL
@jaysonlindsley.dev

I'm building a character AI, and, as a result, I'm feeling rather surprised at how this tech is simultaneously saving me so much time while also being completely unable to follow directions and being very lazy. By itself, without chaining agents together, I am unable to get output that follows (1/x)

JL
@jaysonlindsley.dev

follows the rules i set when there are multiple conditions. For instance, i have my base prompt that has all rules. The output happens, then the rule checker fires into an AI evaluator. The evaluator will *STOP* at the first rule it thinks it's in violation of (loosely) unless ...

JL
@jaysonlindsley.dev

I bring the whitelist above it. When the whitelisted rule "Character must be respond with [X] if [Y]" is above "The character must respond", then it's like "Ok, it passes". But if i move it, the lazy ass AI seems to just stop immediately. It simply cannot understand complex rule systems at all.

0 reposts1 like

JL
@jaysonlindsley.dev

So while we simultaneously go "wow this is great!" - it's really just three agents in a trenchcoat trying to make sure the original prompt was actually followed properly given it's entire context. What an incredible waste of compute? resources?