My latest work in progress is:
My idea is that there will be 2 personas powering the system:
- Persona performing triage - the one defined already today (triage bot)
- Persona that interacts with moderators / high trust users (mod bot)
By chatting with @mod_bot moderators (or very high trust users) will be able to guide @triage_bot on how to behave.
For example:
@mod_bot, be sure to let @sam know if anyone talks about ai
This will trigger mod_bot to amend the system prompt on triage bot. Which means being in this specific chat room will be enough to allow any community to train the robot to behave the way they want it to.
It’s an interesting twist on implementing memory. Not sure how well it will do in practice, but it is a very worthy experiment.