Emad Mostaque
speaker
1,351 appearances
1 recordings
1 series
first heard Dec 2024
last heard Dec 2024
Emad Mostaque’s voice in public audio — every appearance, attributed to the second.
Trend
recordings per month · last 12 monthsNo recordings in the last 12 months.Older appearances are listed below; set an alert to hear about the next one.
Appearances
This is where it becomes again very interesting because
Let's rewind a bit one on stability we did stable diffusion.
And that cost maybe five million dollars of compute in total, right?
To do.
And that's a decent amount of megawatts of electricity, you know, is a million or something
A one hundred hours and A one hundred is what?
Gotcha, can't remember.
Four hundred watts?
So four hundred megawatts of electricity, is that right?
Something like that.
And that's a good amount.
But then to run the model is just a little bit on the edge.
So you think about it, the pre-training and then the post-training, etc., comes out with an artifact, which is this compressed set of weights.
That then means that you don't have to retrain that thing.
Where a lot of them exponential energy usage is coming from is
A lack of optimization, and this is why Meta releases Llama, one of the reasons.
They got three hundred and fifty thousand H one hundreds, they're about to go bigger.
Ten percent are used for training.
They have improved the efficiency of LARM by more than 10% by releasing it.
So the remaining 315,000.
Showing 581–600 of 1,351 · page 30 of 68
← Previous
Next →