It is actually a very interesting approach. The traditional way to fight noise is to have sensor elements as large as possible within the limits of sensor size and resolution.

Basically having larger sensor elements is like having a bunch of smaller elements that cover the same area and averaging them.

But if you can read these smaller sensor elements probably you can do better than just averaging in software, if you have good noise model and image model.

I am really curious to see what algorithms they use for demosaicing/downsampling, and what kind of lowpass filter is in front of the sensor.