My hypothesis is that writing is easier to critique than content, so if you don’t understand the content, you can still “offer a meaningful review” by commenting on writing. A hasty generalization but 🤷🏻‍♀️

Apparently the one vote to keep maybe have been an accident, too (according to Time)

Maybe my opinion is weird because I didn’t start drinking coffee until the job market, but decaf is my nice alternative if I’ve already had caffeine (either because of double booking or poor foresight)

This looks great! Thanks to @sjgreenwood.bsky.social as well— excited to try it out

TLDR: if you have infinite model expressiveness, loss function choice probably doesn't matter. BUT if you have a limited model class (even if the Bayes optimal model is in class), then task-specific surrogates perform better than task-agnostic + thresholding, formalized through H-calibration.