David Cox
speaker
634 appearances
1 recordings
1 series
first heard Jan 2026
last heard 13 Jan
David Cox’s voice in public audio — every appearance, attributed to the second.
Trend
recordings per month · last 12 monthsRecordings per month over the last 12 months — 1 in all, peaking in Jan 2026 with 1.
Appearances
But really what happened under the hood is that it attached an adapter, put in a structured template or input, and got back out an output and parsed it and brought it back to you.
And the LLM ends up being like a little computational engine inside a larger computing framework.
So I think this goes deep.
And it's very different from where I think the frontier model folks are going.
I think there's going to be a lot of opportunity for us to do lots of really interesting things.
So one of the scourges of, I mean, LLM is all this stuff.
This technology is moving so fast.
So a lot of the stuff we have today is kind of like, is it the best thing?
Like, I don't know.
We're moving too fast.
Like, stop asking questions.
We're moving this way.
So a lot of things, we've just kind of dragged them along.
And they're super memory intensive.
You need to have lots of GPU memory.
not just because the parameters like you know people talk about like a small model or big model they're talking about the number of parameters there but also because these things have this context that they're keeping which you know people talk about context length you know long or short but that creates something called a kv cache okay key value cache
And those things are getting huge.
It's like every there's like all these blocks of this transformer and they're all just spewing out tokens worth of KV cash.
And pretty soon that's way more significant than the amount of parameters.
So if you only have a certain amount of memory, then like how much context you can have, how big a model you can have, it all trades off against each other.
Showing 161–180 of 634 · page 9 of 32
← Previous
Next →