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 months
No recordings in the last 12 months.Older appearances are listed below; set an alert to hear about the next one.

Appearances

newest first · ▶ plays the moment
You know how to do it.
Like will it be a Lama type model versus or transform model versus something hybrid, Jamba style?
Who knows, right?
But just make your pick.
It's pretty straightforward to pre-train.
Host training and optimization, distributed stuff.
becomes very interesting there.
The verifiable inference allows anyone to contribute their compute.
And then it's about having secure computation for running these things in regulated industries.
And that's some of the TEE work that we've seen.
But most of it's coming together now and especially like I said against the context of two other things, which were large context windows and function calling, like making it more deterministic on the outputs.
And the final thing we just need is maybe continuous training.
For the individualized stuff, but again, we've seen big advances in that.
So continuous lady.
Yeah, like I said, I think distributed training on millions of GPUs doesn't make sense.
But sending a packet to you that contains
Indonesian classical architectural law and checking that for consistency and rewriting some of the data there does make sense.
Putting it through a Lama's seventy B or eight B model.
So I think distributed data augmentation makes sense, distributed fine tuning makes sense, and then comparing those things.
And then the mining operations as well.
Showing 981–1000 of 1,351 · page 50 of 68 ← Previous Next →