It's worth recognising that a year ago gpt-4 was the cutting edge, and today you can now run a gpt-4 class model on a macbook pro (with lots of ram, sure).
A year ago gpt-4 was the cutting edge, the new cutting edge model takes 11.2 MWh of energy to answer a single question (enough to power a home for 4 years).
Yea but gpt-4 is also only very slightly marginally useful in terms of productivity as it stands today. I highly doubt we're going to get returns on this without huge structural changes to how chatbots work.
Sure, I think that's fair - the tech is only one part of the whole equation. Just pointing out that power demands are really quite a small (and actually quite surmountable) part of the problem.
It's interesting to look at how hardware requirements change alongside acceptable response time. Knowing when you need a reply seems like it might save a lot of money.
The executive order talks about building new geothermal power plants to support AI infra
How about we cut the AI bit and just get a nice C02 decrease, i.e. the actual thing that will support future humans? THat's one more coal plant we can shut down.