A while ago we were experimenting with a configurable thing I called a "nudger" when we did we saw that we could make the model run autonomously and be extremely thorough and commit a complete task with minimal human interaction. But, initially we didn't provide a human or model affordance to dial this as appropriate for the situation.
A podcaster I listen to used the term "tenacity" to describe the behavior we were tuning. I like the word to describe how HARD the "nudger" pushes the model and as we learned you need this to be DIFFERENT depending on conversational state, context, and phase.
Step 0: bring back the nudger but put it under "tenacity" and we'll allow a tunable tenacity manually at first then find a way to drive it autonomously.
A while ago we were experimenting with a configurable thing I called a "nudger" when we did we saw that we could make the model run autonomously and be extremely thorough and commit a complete task with minimal human interaction. But, initially we didn't provide a human or model affordance to dial this as appropriate for the situation.
A podcaster I listen to used the term "tenacity" to describe the behavior we were tuning. I like the word to describe how HARD the "nudger" pushes the model and as we learned you need this to be DIFFERENT depending on conversational state, context, and phase.
Step 0: bring back the nudger but put it under "tenacity" and we'll allow a tunable tenacity manually at first then find a way to drive it autonomously.