Nice point. This is how in understand why reverse KL is useful here: these methods want to nudge the model’s existing behavior, not overwrite it with the teacher’s behavior. And only reverse KL has that property because it didn’t try to cover the entire teacher’s probability distribution
Not entirely true. NE's drivers are high natural gas/fuel costs + limited pipelines/ports, aging grid upgrades, and weather extremes.
CA also has high wildfire-related costs.
Policy clearly has some impact, but not as much as we might think
This report has more details: https://t.co/cmd1wOtYUf
@mickeyxfriedman I’m partial to a good combobulating, but yes spelunking is my fav too. The absence of a reticulating (“splines” is implied) is clearly the biggest miss by CC tho