Tim Davis

When you train a model, whether you're using PyTorch or other frameworks,

you actually need to do a lot then to serve that at scale in production.

It's not like it just instantly happens.

And so we realized, well, hang on a second, if we could build a platform that not only helps abstract away some of this compute, but began to work towards this vision of what you train is what you serve, then maybe a very significant contribution to the planet and to humanity is helping to realize this sort of vision of super intelligence.

Have you heard that?

Okay, yeah.

I was just going to say, and it's an interesting observation, right?

Because I think most programming languages today in the world, they're very CPU-specific.

It's not like Python was never truly designed to go and run on a large scale, highly parallelizable, accelerated machine.

That actually wasn't why it was created.

Most programming languages today were very specific on CPU related programming because most of the world's compute was CPU.

As now everything's sort of turned over and we're going through this massive upheaval of infrastructure that's spinning towards parallelizable, accelerated compute.

You know, what is the compute model that supports that, right?

And I think that's where what we realized at Google was you really need not only a programming model that jumps on a CPU and can do everything that you want on a CPU, but that can then also carry that over to an accelerated device.

And the most famous of this today is CUDA, but that is very locked to NVIDIA's systems, you know, and they've made a bunch of incredible, you know, progress and design decisions as to why they did that.

But then...

It doesn't help the developer who's out there going, but hang on a second.

I don't want to just necessarily use NVIDIA.

I also don't want to use other types of silicon.

There's all this new hardware being invented.