Does Anyone Use Old Models?

Over the weekend, I went to the Vin­tage Com­puter Fes­tival at the Com­puter His­tory Museum down on the peninsula, an event that prob­ably deserves a post of its own — it was fabulous — but/and I want to jot down some broader ques­tions about how tech­nology ages.

I read this line in a newsletter—

gpt-3.5 is 10x cheaper than it was. it’s also as desir­able as a flip phone at an iphone launch.

when a new model is released as the SOTA, 99% of the demand imme­di­ately shifts over to it. con­sumers expect this of their prod­ucts as well.

—and wondered, is it true? I have no idea. It sort of feels true, which is, of course, THE DANGER ZONE.

True or not, it made me think about the useful life­times of com­puters, and chips, and AI models. I’ve written about how I use a ten-year-old iMac, which is not, I promise you, just some stub­born performance: I use the old com­puter because the old com­puter is a powerhouse, incred­ibly useful for all the things I need it to do.

The yel­lowed hulks at the Vin­tage Com­puter Fes­tival were mostly not useful anymore, but they were still very interesting, and very pal­pably beloved. Maybe you could say, they have become a dif­ferent kind of useful.

The AI com­pa­nies are out there packing Nvidia chips into their data centers, in vol­umes that have become macroeconomically meaningful. What’s the depre­ci­a­tion schedule for a hot-rod AI chip? How much value do you have to wring out of these things to make them worthwhile? (And what hap­pens to all the old chips, anyway?)

Last, there are the lan­guage models. Do the old ones get used? How much? There are people at Anthropic and Google who can answer this question. I’ll bet you could glean insights from Hug­ging Face usage data, too, if there’s some way to slurp that up.

Basically, a lot of ques­tions, some of which are rel­a­tively easy to answer! Per­fect blog post!