Better Assistant Panel layout
Rajveer2009 opened this issue · comments
I respectfully disagree. It is useful to have Assistant blocks editable.
LLMs get the whole conversation sent on every completion, not just the last message. This means that every message, both my and assistant's are sent on every turn. Every token costs money. So it's useful to remove unhelpful parts of the previous messages. Things like hallucinations, extra verbose explanations, ideas that turned out to lead nowhere can be removed. Useful parts can be rephrased more succinctly. This can lead to better results, while saving some money/time.
I respectfully disagree. It is useful to have Assistant blocks editable.
LLMs get the whole conversation sent on every completion, not just the last message. This means that every message, both my and assistant's are sent on every turn. Every token costs money. So it's useful to remove unhelpful parts of the previous messages. Things like hallucinations, extra verbose explanations, ideas that turned out to lead nowhere can be removed. Useful parts can be rephrased more succinctly. This can lead to better results, while saving some money/time.
This has sort of been our goal with the assistant since day one - give the users control over all the text, nothing is inaccessible - so you can correct and guide the model as the conversation develops.
I think for now, I will close this issue out, as our current design is very intentional and we are currently building out more features that lean into it.