“Spending some time thinking about how do I catch those 5%? How do I get this closer to maybe 99%? How do I make sure that when the 5% goes wrong, it's at least easy for someone to notice? Can we have two different prompts parsing the same chat message? Do these things agree? Do they disagree? There's all sorts of ways we can get creative to get a little more confident at catching those 5% that isn't necessarily relying on a person.”

pitched it and I think the sales traders were a little less than happy. And I think at the same time, it's like, it's hard to figure out like how much is that actually going to cause people to actually, I don't know, bring up that extra scrutiny level. I think.

And you could tell how good they were at it. Yeah. It's tempting. It's

And obviously, you're in a world where the underlying technology is changing really fast. Yeah. It's also maybe not always changing in a way that's good for what we want in the sense that one thing that happens now with models is they spend a lot more time chewing on and thinking about things.…

Yeah. So we've been even thinking about, you know, some of our counterparties, like they just can't move on to LLM parsing because they just want their responses too quickly. And it's very hard to guarantee how fast that response is going to come from versus, you know, other counterparties. Like…

Right. And there's a question of like, do you give the LLM just like the one chat message or like the history of chat messages where it might in fact be better able to build priors if it sees more stuff, but it's of course going to take longer to do its inference in that world. Have you thought…