I've long thought that among the chief promises of a self-improving AI is not that it will exponentially make new discoveries in Science. But that it will be able to synthesize and combine existing knowledge in a manner too complex for most humans.

The problem with current knowledge is that humans just aren't capable of handling more than a tiny fraction of it. A functional AI (even of the weaker variety) is capable of handling orders of magnitude more information and arriving at novel combinations of existing knowledge.

Part of the reason the impact-per-researcher has fallen so much is that there is simply so much more to know and consider than there was a century ago. And Computers are really, really good at dealing with volumes of information that would overwhelm a person.

And while a self-improving AI may someday slow down it's rate of self improvement, in the short term its rate of self-improvement would appear exponential as it groks scales of knowledge humans can't cope with.