It needs to limit the jerk (rate of change of acceleration). The other day I was following a motorcycle and the neural net kept losing track of it and locking on to the car in front of it. It would floor the throttle and get a few feet closer to the motorcycle, see it and then do full regen (maybe even some braking). Not smooth but I guess it's good for keeping the driver alert! And of course when you're following someone who uses the accelerator as an on-off switch TACC decides that it wants to drive the same way. I do wonder if a convoy of Tesla's were following each other if it would result in an uncontrolled oscillation.