This is a special post for quick takes by jelly. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
You can imagine this problem as trying to have an extension of you that thinks exactly like you but does it 10x times slower (so other things come 10x faster to them).
This helps transfer your short-timescale thinking strategies to long-horizon problems
The slower you discounts future things less, forgets long-timescale-relevant things less, doesn't update too much on random noise like e.g. recent news, subjectively experiences richer feedback in low-feedback problems, intuitively feels how much long-timescale tasks or history have progressed overall, ...
This could reframe how we think about rationality! ("rationality is partly about how we extend human thinking to long timescales")
But another part of rationality is to maximize something to the extreme, so that's something too
Another possible counterpoint is that short timescale problems are systematically easier than long timescale ones, bc we spend more compute on longer timescales, and so the bar is naturally higher
... which is evidence against attempting to generalize short-horizon strats to long-horizon being successful
or, maybe because they are so slow, the slow agent could only be used for bookkeeping and aggregation, with the original agent doing most of the cognitive work