Keri Briske

Because we need to know how the models can form, what it was trained on, why it was trained on that.

You can't put this into critical applications unless you can truly trust what it's been built on and kind of open up and see inside.

Yeah, I think it really shines in our efficiency.

I'm really proud of the hardware architecture that we've done.

Again, of course, we do.

We build it for ourselves.

We want to make sure we build the most efficient systems, and it's this virtuous cycle.

And I think it really shines on the reasoning, right?

We have really great reasoning data that we've put out.

We've even helped translate it to some other languages.

And so I'm really proud of the team doing that.

Room to grow, I think...

If I had to be really critical on my own product, I think that right now we're working on two things.

One is long, multi-task, complex steps.

So the more tools and steps that you have to take, you can kind of get lost along the way.

And so now we're doing more data collection for those types of scenarios.

So really complex tool calling is something that we want to do better on.

And then maybe this is a little bit visionary, but, you know, we have a really great speech team, world-class speech team.

We actually have, talking about open models, we've got really great speech models at top leaderboards and are open to, they're called Riva and Canary and Parakeet models.

We'll pull them under the Nemotron umbrella and maybe start some more Omni models.