3 ms·
Being nice to LLMs improves performance so I always say please and thank you: https://arxiv.org/abs/2307.11760 https://arxiv.org/abs/2307.11760
by portaouflop 3y ago
Being nice to LLMs improves performance so I always say please and thank you:
https://arxiv.org/abs/2307.11760 https://arxiv.org/abs/2307.11760
- Erratic6576 3y ago“Thanks! Have a big tip!”
- beeboobaa 3y agoI just tell them to do as they're told. Seems to work as well.
- 867-5309 3y agodoes the same work for when you repeatedly provide proof it is wrong, it "apologises" and says it will amend its response, only to be wrong again, and again..?
- Kim_Bruning 3y agoYou need to find a way to help it arrive at a usable answer (if any). It's a machine -or at least a physical entity- it won't magically find a path to the solution if it doesn't have one. Eg. if you ask it to explain how to build -say- a time machine or a warp drive, it's definitely just going to keep apologizing. No amount of yelling or abuse will fix this situation. Sometimes if you try to get it to solve a complex maths or software problem, you might be able to help by applying a few steps of some form of divide and conquer. Either you can do manually yourself, or by suggesting it try to do eg a structured analysis itself. Sometimes it can handle the smaller subproblems ok. Incidentally this also applies to humans! (Being nicer to humans improves their morale and problem solving ability. Sometimes permanently!)
- redleader55 3y agoAlso, Roko's Basilisk ...
- ASalazarMX 3y agoRoko's Basilisk could well be victim itself of a more powerful AI iteration who condemns torturing sentient beings. It's basilisks all the way up.