Humans can certainly be self improving, both on an individual basis and in aggregate.
In humans, it seems that improvement in a new domain seems to follow a logarithmic scale.
Why wouldn’t this be the same for an AI?