The Exchanges

Q Next up, we have Koki Moshida coming into the studio. Uh, our last teal fellow of the day, finishing it out strong. Welcome to the stream. Koki, how are you doing? Sorry for keeping you waiting. We were running long, but you're here. Please introduce yourself and the company that you're building.

A Hey guys. Uh, thank you for having me on. Uh, thanks for the time. Um, I'm Koki originally from Japan. Um, you know, I spent a lot of time growing up thinking about natural disasters. I think, you know, in SF you don't, you know, it's like nice weather, everything's, you know, kind of nice and, and you get to work on really, really interesting things. But, you know, I was in the 2000 loving earthquake where, you know, 18,000 people died in a couple of days. And a lot of my friends, you know, homes were destroyed and you have these like, you know, typhoons, right. Um, that like blow away my grandparents' roofs every like three years or so. And I keep asking them like, like why, like, you can just like move Because, you know, you're in Typhoon Alley, like, what's the whole point of, you know, having to stay there? But it's like something so deeply ingrained, right? It's something, you know, that you accept, that you, you know, kind of just accept as, as a feat. And I was talking to the environment minister and the, and the infrastructure minister of Japan recently, and they were saying that they, they spent billions of dollars on, you know, disaster infrastructure. Like, they would build these, like, You know, columns under Tokyo, right? So like, when it floods, the water can go in. Like, they were like, oh shoot, we haven't actually solved the problem, which is the intensificati…

AI assessment note: “I'm Koki originally from Japan. Um, you know, I spent a lot of time”

Q Yeah. Uh, you, you mentioned diffusion, uh, is there a distinction between weather forecasting using diffusion models versus transformer based models or token prediction models? Like, is, is that a meaningful distinction or are you kind of using the term broadly?

A It gets really complicated, uh, in a sense that I think, you know, both of them Are good at, you know, kind of milking, um, as much, you know, information, you know, looking at, okay, what, what are the lane parameters? How can you, you know, figure out, you know, somewhat unnoticeable, um, characteristics within, within other models. But I think the interpretability, um, of weather models also become very important, right? Like if you can like characteristically like define, okay, which parameter within a high dimensional vector Right, it's creating a lot of bias. How can you, you know, tweak that so that you can, you know, get back on track? And it's like, weather is so chaotic that, like, even one small perturbation, um, in an initial condition can really mess up the forecast, and that's why it's really difficult to forecast, like, a week out. And so, being able to, like, assimilate data as you go, um, and that's why, like, we're focusing on, okay, how can you, you know, collect data that no one really has, um, So that you can, you know, make sure you're not, like, de-biased too much.

AI assessment note: “both of them Are good at, you know, kind of milking”