I am looking at the top500.org Performance Development chart of supercomputers since 1993. What's interesting is how the No.500 spot has improved in remarkably smooth increments since 1993; while the No.1 spot has not - rising in distinct steps. This makes sense I guess, the lumps and bumps of innovation, etc. All that fresh air at the top of the mountain. Naturally I can't help wondering though, at what position would the graph change from one to the other.

But my question really is, is this a generalisable phenomenon? So, say, that in any given rankable performance is there a level at which the rate of improvement is predictable?

Or to put it another way, how good do you have to be to be unpredictable?