Keri Briske

It's more than a data set.

So that's why we call it a gym, because you're kind of working it out.

And if you really think about it, it's like a little agentic system on the side that you're kind of running the model through and you're exposing it to tools and saying, here, here's these things in front of you, go figure it out.

I've asked you a question, you've put tools in front of you, go figure it out.

And so you're right, you're absolutely right.

People are spinning up ISV or software vendor

interfaces to be able to say, hey, here's where this button is.

Here's what this function is.

How would you do this invoice?

How would you write?

And so it's very interesting.

I think the space is going to get even more interesting.

No, I think my answer, and you're not going to like it, is it depends because it really does depend on the use case, your compute that you have, your latency requirements.

So it's all about throughput, latency, and accuracy.

And so when you triangulate between those, you kind of find the right model size for you and your task.

And so we have found that the sweet spot has been the super because, again, it fits in a single data, at least for enterprises, it fits in a single data center GPU because enterprises want all the accuracy they can get in an itty bitty compute footprint.

Yeah.

Right.

And save on cost, right?

So they want the most accuracy that they can and the lowest latency on the smallest compute footprint.