3 Comments

You say 1.5 uses Pro levels of compute, but I don't think their announcement says that. It says "less compute" than Ultra. Long context likely uses more compute than their pro model, but how much less than a 32k context Ultra is yet to be known. It seems to take longer for lover tokens at least, so even if the hardware it runs on is the same as Pro(I'm assuming it takes more VRAM), it takes more compute cycles overall. If you're worried about fast takeoff, I think this is a good thing because it's not an improvement in compute requirements.

Expand full comment

Hmm. True, technically they did not say how much less, although I assume the base amount before context window issues is comparable. If you use the whole larger context window, yeah, it's going to get a lot more compute intensive.

Expand full comment

The pace of change reminds me of the Web circa 1994-1995. Nobody can keep up.

Expand full comment