Comment by jedberg
8 hours ago
Whenever computer chips go into space, they have to be hardened against radiation, because there is no atmosphere to protect them. Otherwise you get random bit flips.
This process takes a while, which is partly why all the computers in space seem out of date. Because they are.
No one is going to want to use chips that are a many years out of date or subject to random bit flips.
(Although now it got me thinking, do random bit flips matter when training a trillion parameter model?)
LLMs specifically are fine with random bits flipped for the results to be 'creative'.
That's not exactly how LLM temperature works. :). Also that's on inference, not training. Presumably these would be used for training, the latency would be too high for inference.
It doesn't work like that, but it can.
Latency would be fine for inference, this is low earth orbit, that is about 25ms optimistically. Well within what we expect from our current crop of non local LLMs.
not a problem for "AI". it's just a bit more spice (temperature) which Grok possibly need! jk!