Using your scale? ↑7 or ↑8. That seemed to be the sweet spot in capabilities to me, without getting to the point where engagement eats everything else including productivity and future maintenance.

Using semiconductor process nodes? 45-65nm. That was around the point that Moore's Law broke down. At that point, you could do most of the functionality that we depend upon computers for (eg. GUIs, 3D rendering, networking, basic machine-learning, some speech recognition and text synthesis). It also roughly corresponds to ↑7 or ↑8 on your scale, so it's self-consistent.

Conceptually? I'd like to have multiple checkpoints, so that if the ecosystem gets borked you can roll back further.