GPTs lieing about work #167
EuanHoll
started this conversation in
Show and tell
Replies: 1 comment 1 reply
-
Perhaps there is a way to link this to task execution. If a task is done via a function call the function returns some basic "Feedback" of the execution and then limiting the developer to only report on things marked as function "Feedback" in order to force it to not hallucinate tasks that were not performed. |
Beta Was this translation helpful? Give feedback.
1 reply
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
So I've been working on this GPT Swarm idea a bit by myself. Trying to figure out the best way to allow the LLMs to communicate and I've come across this fun pattern while trying to get the LLMs to do stuff in my current format.
I haven't finetuned the CEO brief yet to see if I can avoid this, but it's a common problem I've run into in many attempts with tasks that involve more than talking.
They start to act like they can and are doing things outside the messaging between each other.
This makes it difficult to get them to report stuff they've done if you want a final report, but also just to get them todo tasks like programming. I think I can get a way around it by adding another messaging type to the communication format between them, but thought I'd share this learning here as it will probably be something you peeps run into too.
I imagine the cause for this is that they are just emulating chat messages in a slack/teams channel rather than actually doing the task provided to them. But that's just a guess.
Beta Was this translation helpful? Give feedback.
All reactions