Well, A) "top tier engineer/scientist" is a significant step above generic human, B) the human engineers/scientists also have immediate access to the same database, C) The humans have been optimizing it for even longer, so what makes us think the AI can optimize itself even a couple percent?

For example, if the number of AIs you can run per petaflop started to scale with the cube root of researcher-years, then even if your researcher AIs are quite fast and you can double your density in a couple years, hitting 5x will take a decade and hitting 10x will approach half a century.