I have no idea what I just read, did they figure out how to carrot and stick ai to self improve?

Basically it appears they've applied found a way to calculate similarity to figure out what skill to apply to downstream tasks. Yes though, I wound up clicking through to the paper.

So they taught it to take the skills from one task to another? Self improvement stuff? Quick get quicker?

Kind of and not really. Basically it's about where in the series of tasks they figure out what skill to use.

So this is more an analysis of how it works? And in knowing that they can focus on improvements?

They're changing the order (and removing a layer of complexity) by using an earlier part of their model to calculate skill use later. It could certainly lead to further improvements in either efficiency or potentially transfer learning down the line.

I think I get it, thank you.