Tim Davis

There's all of this hardware that exists in a phone.

Yeah, but like...

It's funny, Corey, even in a mobile phone, you think about it, one, NVIDIA doesn't actually get you there, but two, in a mobile phone, you have a CPU, a GPU, a digital signal processor, a DSP, and then an NPU, so your neural processing unit.

You have four pieces of silicon.

Um, and so again, you like, you know, to really, uh, unleash the power of that, to really enable people to say, well, what are these amazing applications that we could be building on mobile phones and, and other devices, wearables that are out there in the world?

What is the compute model?

What is the platform that enables us to do that?

And there isn't really a good one today.

Um, until, you know, here's a plug until we came around, I would claim, uh, and that's what we're, you know, that's what we're building.

So stepping back, what is intelligence?

I would argue today the path we're on, and I would argue I don't necessarily know this is the right path, but the path we're on today is let's keep training...

according to the scaling laws let's keep training bigger and bigger models right so i think grok was trained on 5e to the 26 flops um you know now there's discussion about let's go and train models on you know 10 to the 20 28 29 flops which which ends up being gigawatt facilities running model training cycles for months and months and months like it is an enormous scale um

But fundamentally, what are these models, right?

Like, what are these LLMs?

They're autoregressive models.

They're taking, you know, basically history of text and then trying to predict the next token in many ways from, you know, left to right generation process, right?

And so is that actually intelligence?

Like, are we actually, you know, creating, you know, intelligent machines when we do this?

I mean, the perception...

I think to humans is, well, clearly they're intelligent, they're spinning out incredible solutions to problems, but all of that has been trained on human data, this incredible massive corpus of text that now has existed on the internet.