Moonshine: distilling interactive technical explanations

What I wanna share today is, a skill I've been working on with a friend calling it moonshine. The name is, homage to, Distill, which is a machine learning, journal, I guess, that actually worked on a few articles many many years ago or moons ago I should Staff. And the whole idea was that you would kind of be able to try to understand these more complex thank you. Consulting, algorithms, you know, different aspects of machine learning if you had, like, a richer medium. So not just a PDF, but you could, show, you know, kind of demonstrate these concepts, interactively. And we built a whole lot of different, you know, visualizations, figures, different ways to try to communicate these concepts. People were pretty excited about it. We were excited. We actually you know, some of the stuff I helped here, like, the researchers were researching as we're explaining it. So it wasn't just like, oh, here we we figured this out. Everyone else should learn it. It was like, how do we even think about this stuff? And, you know, but the thing that we heard the most when working on it from other researchers, basically, everybody was like, this is great, but where do we get a team of, like, 3 engineers who are good at UI, you know, the data processing, data and everything outside of the research. Right? And so, you know, fast forward to now, we know where we can get something like that, which is an agent. And and that's, you know, why I'm trying to build this skill. Let's see. So the other thing I wanna point out was that way back in 2017, when they founded Distill, Chris Ola and Sean Carter wrote this research debt article, which, you know, you may be interested in going back and checking out. I think it still holds true, but the basic idea being that there's so much we could advance if we could just understand what's already been put out there. And, you know, publishing and and the papers, the format of it isn't really that friendly for people that actually understand what's been done, how to do it. So trying to find a better way. So I guess I'll just give a couple examples. The the thing that's been really fun with this skill is that it's really, like, sort of easy and addicting to at just ask it to explain something to you. And it has a bunch of stuff built in where, like, you know, the ideas of having the text interact with the diagram. So as you're kind of, you know, playing with it and then you're trying to read and then it's suggesting that you do something with diagram, it actually does it for you. I'll just go through a few of the article series I've made and then my friend Kai has made. And I'll show a couple of things I did at, sort of at my day job where it wasn't as much like this, you know, fun algorithm visualization, but more like software engineering. We're making this, you know, way more complex thing that, like, I can't understand what my coworker is doing. They can understand me. And then yeah. So this first article, this is my in my master's thesis 15 years ago, I worked on, real time fluid simulation, using a method. I used particles, which is a lot like is essentially very similar to the flocking stuff that I'm sure many people have seen where you can, you know, simulate a flock of birds with these simple rules. But it it used it was a lot of partial differential equations, which I was learning at the time and have since forgotten. However, you know, Claude has not and has been able to, like these were all static figures in my thesis, essentially. But now you can sort of interactively play with these concepts of, like, how does how do you calculate density for any given particle? You know, what is the pressure force? How does that change with the parameters viscosity? Back up a couple of steps. Sure. What I'm sure this 1 right here? Actually, no. This is just me saying I wanna explain smooth particle hydrodynamics. Give me, like, a figure for each of the forces. Like, I didn't I didn't even give it my paper. I just yeah. Yeah. So it's so you'll see in a second, my friend has made, like, a 100 of these articles because he I mean, I think he studied math in school. He's got some really trippy ones. But basically, if you say, like, here's, like, a complex thing, I wanna explain it, use the skill. It'll interview you to be like, oh, what's the audience? You know, here's what I think would be some good figures. Do you wanna expand on these? You know, I built that in to try to because a big part of making those System articles, it wasn't just the technical part of making the figures. It was like, what's the story you're trying to tell? What do you you know? And actually answering those questions teaches you quite a bit too. So, yeah, I mean, I don't wanna teach fluid dynamics right now, but it's been really fun. Yeah. So I was saying, like, my friend, he's just got, like, you know, there's, like, 40 articles on different emergence concepts. You probably recognize some of them. You may not recognize other ones. It just I mean and I guess I could click on a few of these, like, you know, just any given 1 has these really interesting, diagrams that you can play with. Right? And so there's obviously way too many to to go through here. And then I wanted to show some of his math ones. Oh, 0, no. And it but it's not just math, algorithms, whatever. He he did a whole series on lithium ion batteries. So, like, what is it and, you know, how to think about it compared to a double a battery or an EV EV pack. Like, the math, you can really get into, like, you know, electricity, how they they fail, what happens to capacity when you change it. So really just like any complex thing. And then I probably talked for too long already. So I'll just show the the work Staff. Like, we were building a system. You can kind of, like, agentically generate these internal workflows or whatever. It's complicated to design a system that designs systems. Right? And so making this article to just break down, like, the the rules that we're setting up, make it interactive so, like, my coworkers and I can talk about these designs in ways that aren't just, like, comments on a Notion LLC, but actually, you know, take each concept and drill down as deep as you need to, but try to have it written up as pros as well. Or even we had, you know, bugs or user misunderstandings of a complex thing. So, like, this is like a incident report essentially, where the user tried to design something in the system, didn't work as they expected. I think the second part, there's like a figure where it kind of plays through what should have happened. We won't really, like, stay with it. Sorry. The very last 1 is, like, I really like I had it figure out how to stick this graph up here and then, like, link it with the text. This is the kind of thing that I would have spent, like, probably 2 months on 10 years ago and I could do in an hour now. How to do 1 from scratch is just you either put this in cloud code, or you can just use whatever agent thing point at the repo and say download the skill.