Did some digging. They used 800000 cells, it never got to superhuman levels, human cells are better than mouse cells.

Things to take away, 800k is not equivalent to 800k parameters. 800k cells might equal anywhere between 800m to 8b parameter(1000 - 10000 synapses per neuron). If a bloody 800m parameter model can't learn to play ping pong like a superhuman it's not really that great. However, it probably does converge sooner, given that the way neurons learn in brain isn't like a feed forward network.

The most curious thing here is that apparently neurons work by trying to reduce entropy. Looks like all of life is trying to do the same at many levels.