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
as it were, which is maybe a story for later on other day.
But dedicated infrastructure build out to create an intelligent capital stock that then gives capability.
Just like building factories gave capability, but we don't need factories anymore, right?
You actually need pure hardware running as standardized models as possible.
And those standardized models need to be based on standardized data sets.
And just have regular releases and be predictable in the way they operate.
Yeah, so 8B model is runnable on a Mac Neural engine.
So a Mac Neural engine will use about five watts of electricity, as will the Intel MPU.
But if you want more, like I've got my M two Max and I'm waiting for my M four max, that will use up to a hundred watts of electricity.
Right.
If you look at the H one hundred, the H one hundred uses a thousand watts of electricity.
So you can use more energy and it'll go faster.
But
Ten tokens a second is human reading speed.
So you can you know, you see these models go like that.
If it's faster then you can read, what's the point?
In some ways, right?
And there's a question of how many tokens does a human need to flourish?
And so we can do some calculations around that and we're like, well, if you don't need everything synchronous and you can let it go, then five watts of electricity is probably all anyone needs.
And again, you're seeing that because you can now run models on a smartphone, and smartphones tend to top out at twenty five watts of electricity.
Showing 561–580 of 1,351 · page 29 of 68
← Previous
Next →