I'll write that down for reference in case I do ever join an organization like that in the future, thanks.
600% improvement is worth what, 3 days of billable work if it lasts 5 minutes?
Series of such "improvements" could be fame and fortune in your team/group/vertical. In such places, the guy who toots the loudest wins the most.
So THAT's why large organizations want "AI".
In such a place I should be a very loud advocate of LLMs, use them to generate 100% of my output for new tasks...
... and then "improve performance" by simply fixing all the obvious inefficiencies and brag about the 400% speedups.
Hmm. Next step: instruct the "AI" to use bubblesort.