With human drivers: traffic light turns green. The first car starts driving. The 2nd car waits 2 seconds and then starts driving. The third car waits another 2 seconds (4 seconds total) and then starts driving. The fourth car waits another 2 seconds (6 seconds total) and then starts driving. etc.
With computers driving: traffic light turns green. All cars simultaneously start driving. It'd be like a train but without the efficiency.
Similarly, with human drivers: some jackasses drive into the box and the light turns red. Now perpendicular traffic is either fully blocked or must proceeed slower to maneuver around the jackasses. With computer drivers, they shouldn't intentionally break the law and they should have plenty of sensors to figure out that they cannot make it through the box.
Safety margins still will require some level of delay between cars that aren't mechanically linked. Even with perfect reaction times, the physics of driving (maximum acceleration rates, possible loss of traction) dictate this, it's a non-trivial control theory problem. Besides, it doesn't seem to be a goal of Waymo; I've seen lines of their vehicles before and they all behave the same way as in mixed traffic.
As a sorta informed outsider, conceptually this makes intuitive sense. But in practice, how does this work? It seems a lot of the intuition breaks down if we don't assume it's network (aka 1 vendor). Fundamentally it's a bunch of external actors where we cannot verify trust and in order to solve for the needs of the individual, suboptimal choices must be made. To put it another way, even if computers can drive cars, what _else_ needs to be in place for this vision?