I wonder how much of that is caused by the fact that the models are so slow they're forced to stream their output to the end user? What if the they could produ

I wonder how much of that is caused by the fact that the models are so slow they're forced to stream their output to the end user? What if the they could produce the output and feed it back to another session that gets continuously asked to analyze where the conversation is going and whether it's likely to break policies?