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
But I don't think distributed training makes sense, or pre training, shall we say.
So post training and data augmentation.
And for those, you really want to have verifiable, you know what the models are.
Otherwise you might get junkiness, as it were.
So I think that unlocks something.
It's not required, it just helps.
Like right now what we're looking at is literal H one hundred clusters in various countries and how we can bootstrap that.
Because they're inevitable.
So you might as well put them there and use them for healthcare later, right?
And then that gives you enough
To then tune models in every country and spin up teams and national champions, and then they can decide what they want as part of this network, right?
And H one hundreds are a lot easier to track than forty nineties or M four maxes, etcetera.
But if you can access that then
Like I said, you'll be able to build the better models even better 'cause you'll build better data sets even better and faster.
I think again, this is not synthetic data, it's augmented, filtered, cleared data utilizing LMs.
You put a massive amount of compute in putting a fine tune over all of it, checking it for consistency, adapting it as appropriate.
Like again, you can afford to go over the top because then you have a gold standard base.
It's again like you have a good curriculum for your kid.
You feel comfortable about that or comfortable about the mechanism and the methodology.
So I think again this is a different type of
Showing 1001–1020 of 1,351 · page 51 of 68
← Previous
Next →