it's just an experimental optimization. Implementation detail... Irrelevant to "safety". I mean its going to have to go in that direction anyway... eventually the models will just be constantly thinking, refining their internal thoughts / weights... External input and output will be rare, just as it is for most humans.