Reading: AI Hardware w/ Jim Keller

[embed]

ending: “this is the funniest thing you know you start training with complicated things you end training with complicated things and it’s all incremental changes no human has ever learned that way we start with really simple things but as you learn you make quantitative leaps and the human data rate for training is really low. we do a maximum of 10 million inferences a year. it’s really interesting, so trainings I promise you. you training is going to change it doesn’t make any sense training a computer for three months on the internet with infimal edits to weights no human beings ever learned anything that way. it’s not a thing. (we all now know that everything is a graph) so everything is a graph Hardware guys make graphs go fast yep yay Hardware all right good job got now.”

The following read:

[embed]

References:

[embed]