magnustuom I've been using the assistant a lot lately for sparring and and general off the rail discussions and I have noticed a worrying thing: the assistant is flattering to the point of sycophancy (quick) (https://en.wikipedia.org/wiki/Sycophancy_(artificial_intelligence)) Are there any steps taken to minimize such behavior? I would prefer the assistant to keep to documented facts and not entertain delusions of grandeur. references (by the assistant): https://www.whytryai.com/p/how-to-reduce-ai-sycophancy https://arxiv.org/html/2602.23971v1 https://github.com/tdeletto/shannon-ai-optimizer this would improve the trustworthiness of the assistant at the very least, maybe?
Dustin Assistant allows you set custom instructions which you can tune how it responds, among other things.
