[reilly3000]

I suppose one of the ADR’s read something like “…who cares about bitflips, man. Isn’t AI all about probability?”

Knowing the insane level of hardening that goes into putting microcontrollers into space, how to the expect to use some 3nm process chip to stand a chance?

[lovidico]

A trend at the moment is to just hope for the best in cubesats and other small satellites in LEO. If you’re below the radiation belt it’s apparently tenable. I worked somewhere designing satellite hardware for LEO and we simply opted to use consumer ARM hardware with a special OS with core level redundancy / consensus to manage bit flips. Obviously some problems will present for AI there… but there are arguably bigger problems with AI data centres like the fact that they offer almost no benefit with respects to the costs of putting and maintaining stuff in space!