I guess if I was going to, e.g., use an agent, I would be worried that the more time savings that was promised, the more the work performed would depart from the prompt I made---obviously work needs to be checked, but it seems to me to be a problem that issues get worse as verification gets harder

My strategy has been to increase the number of times I repeat the prompt for smaller batches, but I have learned that GPT is not great at counting slides

If you run a cross-country growth regression with aid dollars as an independent variable and get a negative coefficient is it that:
-countries doing worse will get more aid on average
-countries doing worse today will do worse in the future
OR
-Aid is entirely broken and reduces growth

It's been a long time but Moyo's argument didn't do much for me when I read dead aid. In that genre, I'd rather assign Easterly, who is much more transparent about the fact that our evidence around aid in the aggregate will be, by nature, contaminated by selection bias.

I'd guess going for 2 down 8 happens more often as time winds down, which would probably reduce win rate as well. You could check this as well, just by taking the average time on the clock when the decision was made.

Forgive me for this late night Point Reyes posting

So October 2 in East Lansing was apparently just a grassroots gathering then lol

I followed you!! (Tnkrtlrsldrspy)

Do you have a letterboxd?

Ostentatious stick count at four

If I get accidentally Bronsoned for this, I knew the risks

Packers win, post cheese