Yeah the biggest problem with the model right now is that it has fat tails (which yeah okay, still some uncertainty because we're months from election day, guess that makes sense) that aren't correlated with each other at all. That's why you're getting obviously bunk outcomes like R-AZ but D-LA.
I'm not sure why Nate's model is doing this because I've definitely seen him harping on state correlations before. Right now the model seems to just grab tail-outcomes for random states in random directions and it leads to obviously whacky stuff.
With that said I spotted this gem on the back panel a few days ago
The issue is that they're allowing a lot of subgroup variance. If you think about it you can tell how many of the weird maps happen.
It can be, say, "extremely good Biden black voter scenario" (black turnout 80% or something) combined with "extremely good Trump Hispanic voter scenario" (Trump wins Hispanics!) and then suddenly Biden is winning places in the Deep South with all his black voters while Trump wins New Mexico, etc.
One of the specifically weird things is that they allow Corona to be a group thing for states. So when you combine a bunch of different tail end scenarios for different groupings, including some that group states along non-political lines (like Corona) you get these really wild outcomes.