Writer kept training models because enterprise adoption lags

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 30:33

   sets they want in the company. Tell us a little bit about Palmyra since you mentioned it. You guys made this big decision to kind of train models yourselves. That takes money. It takes expertise. Why did you do that? Well, I think the big decision isn't that we train models. It's that we kept training models, right? Because when we started in 2020, there weren't third-party models to use. I mean, we started Rider to commercialize transformers. I think we continued to train models as we saw that, you know, the labs might be a few weeks ahead, but everybody's following the same research breadcrumbs. And especially as we started using synthetic data quite early in 23, we just never really felt that the capital gap meant a capabilities gap. And sure, they may be ahead, but there are plenty of eras of this chessboard where we were ahead, right? The thing

2. 31:25

   is with the enterprise, it doesn't really matter because they're still trying to catch up to the innovation that you launched 24 months ago. And so we never felt like we needed to go use third party models, right, to be able to deliver for customers and to be able to deliver frontier performance for customers. The ability now for a customer to use any third party model has given them much more commercial optionality, right, that allows them to check a box than like actual usage, right? Because people get in and they're like, oh, Planera is cheaper or faster.

3. 31:55

   There's really no reason for me to switch off the default, right? But they can if they want to. What it's also allowed us to do is to be able to say at a Unilever, at a Mars, you want to create a video, you want to create an image. So, we already do brand and compliance for videos and images. But if you want to create them, well, just hook up a third-party model. At NVIDIA, they use Nemetron because that is a big edict across the firm to, you know, drink their own champagne, et cetera.

4. 32:19

   So, they want to be able to use their own models. At Vanguard, they had fine-tuned a model on voice of the customer and sentiment. They used that for a number of use cases for a long period of time. So it gives folks the optionality, but we're never not going to be a company that builds and trains its own models. But we don't consider ourselves a lab. What makes you not a lab? Research products. product and go-to-market are one team. So it's everything from a single Slack to a single force in front of the customer. When Cigna started using self-evolving models, they're talking to Dan, our head of research. Dan is at the office, right? He's walking the halls. And it really is just this weekend, you know, we've got Palmyra 6 that we are testing. And I came across a use case where I way prefer Palmyra 5, right?

5. 33:09

   And so that resulted in, this is why, you know, I will perpetually be working 90 hours a week, resulted in, right, like a deep amount of data that I was able to provide to the research team. So it's that continuous loop of product and research and then our go-to-market teams who are literally on-site with the customer saying, this works great, this works shit, we need to improve this, right? We need the model to be able to do X, Y, Z.

6. 33:33

   And that is what makes us a product company, not a lab.