🔴 LIVE | Research Stream | NeRFs incoming!

Let's go ahead and go live. All right, let's see. I should be live on YouTube in just a second. Yes, am I live now? I think I'm live now. Hello. Hey, how's everyone doing? Uncle Bill, what's up? Anyone else who might be lurking in chat or waiting for the stream to start. Hey, thanks for being here. Um hello, wonderful people of the internet. Welcome back to another episode of Alejo's Labs. Um yeah, so uh we were just having a brief Twitch discussion about what's going on the Twitch side. So, uh for people on YouTube, you may not have known I tried to stream yesterday and actually we got a lot done yesterday in principle, but in practice I just felt kind of sick at the end of the stream yesterday, so I closed it early. Um and because I was feeling sick, I just kind of unlisted the video as well. So, you guys aren't really up to date as to what's been going on. So, let me just catch you guys up. So, at the start of this year, January of you know, 2026, um I got a heart condition known as myopericarditis. So, myocarditis is the inflammation of the inner region of the heart, pericarditis is the inflammation of the outer region of the heart. If you put it together, you get myopericarditis, which is what I got. It sucks. It takes months to recovery. Uh and unfortunately, there are days where I just feel sick. That's just what it is. I feel really sick. I can't really walk around very well with my without my heart rate like spiking to 135, 140 or something like this. And um the real problem is it not only does it go up really high, but it also goes down really quickly. So, like my blood pressure just constantly goes in and out and I just feel dizzy, nauseous, you know, I just don't feel really good. And it's impossible for me to have any sort of mental clarity when that type of stuff happens. So, it's just Unfortunately, that happens every now and again and it's been getting better. Obviously, it's been getting better, but I had to take a break on Wednesday. Um and yesterday I came back, but the stream was a little bit just mixed. I wasn't able to keep a train of thought very well and it was too I don't know how to put it, just too uh goopy. — I don't know what word I'm looking for. It just wasn't good enough for me to keep uh live on YouTube. I It's fine either way. This is another one of those streams that might not be like remain on YouTube for too long. It might be that I'm live and then when I'm not live, I leave. And there's a reason for that. — I um I've got a little bit of a problem. So, here are my Twitch statistics. Um yesterday, if you were not here for the stream, you would not have known, but I was for some reason view botted yesterday. I don't know why. Um where my like Here, you can see it in the uh stream summary. You go here, go down to not chat messages. I want average viewers? No, I want Wait, stream summary. It should be Where Oh, no, no, it's Friday. I want Thursday. Uh here we go. We are going to go to uh live viewers? No, we're average viewers. Um and we see a huge spike in here. So, basically, it's like 32 average up and now it's 637 average. And it kind of goes up and down and then drops down to 50 average. I guess there's probably still a few bots here. I'm not really sure what's going on there. Um but the interesting thing about the view botting and I didn't know this beforehand, um my live views did not increase significantly. Um but my average viewership did. Is that not weird? Like I don't understand. How do your live views not go up? Uh what is live viewership? Maybe I just don't understand what the metric is. — Um but yeah, if you look at chatters, you'll see that there wasn't really I mean, I kind of a vague increase in chatters and stuff like that as well. Anyway, I'm not going to go into all the details here. So, that's it. Uh hey Al, what's up? Uh right, just wanted to point that out. Weird stuff was happening yesterday. So, uh what are we doing today? So, today, I'm going to be honest, we might not get too much done today. I'm going to do what I can, but I'm a little bit tired. We're going to see what we can do. Maybe they weren't watching live in an older segment. Could be. Maybe. Yeah, could be. I don't know. I don't know what they were up to. Um it doesn't really matter too much for the purposes of the stream or anything like that. But um as far as Twitch is concerned, there's only one other thing I want to mention about Twitch, uh which is just for people who are viewing right now. I do realize I don't have transcoding options. It's because I'm not affiliated with Twitch, but um I don't know. I don't know what to do. I think I like Twitch. I like it as a platform. I like streaming on Twitch. Uh I don't want to go YouTube only, uh but YouTube only is the only way I can guarantee that people don't get ads on my streams. Uh and so, like I'm in this weird like middle ground where if I want users to have transcoding options and I want them to watch my stream, um then and I don't want them to have ads, you know, then my only choice is YouTube, right? Because if I get transcoding options, uh that means I have to affiliate myself with Twitch and unfortunately, affiliating myself with Twitch necessarily means um that uh you guys will have to see ads on Twitch. And I don't mind it as much because if I'm always dual streaming onto YouTube, it's fine. You guys can just go to YouTube. That's fine. I don't care about it that much. Okay, let's kind of get into the topic for today. We we did all of the updates. Um there's one other thing I guess I want to show you guys, uh which is this. Let me just go to my channel, I guess. And uh let me go to uh YouTube Studio over here. — [clears throat] — So, uh let me uh get some water. You want to see something fun? I always love looking at this. Every time I start making content again, I lose subscribers. A significant number of subscribers. I've lost 127 subscribers this month. And it looks like I'm just

going to keep losing subscribers. It looks like this is going to keep going for a while. Um when I started this, I had, you know, 97167 or something uh YouTube subscribers and now I have 97 I might even drop into 96, you know, uh territory, which is not great. Um 96 is not quite as nice of a number as 69. Uh but, you know, it is what it is. Um So, yeah, I just want to point out that this happens every time that, you know, you start uploading after a long hiatus and especially because the content that I've been putting out is basically like live streams and shorts, which are not content that I'm known for. So, for people watching the YouTube vod um and are maybe just tuning in to see how I'm doing, um I I will say that my passion for making math explainers has dropped a significant amount. Um I want to I do want to do math explainers — in a sense, but I want to make sure that these math explainers are things that I've actually done and not things that other people have done. You know, I don't want to do an explainer on Gaussian elimination or something like this anymore. I want to do it on like new methods or algorithms that we've personally created in the community. Um or like, you know, uh something like this, you know, something that like it's like, "Hey, here's a research paper that we did and here's an explanation of that research paper. " Um or something like this. I want it to be that I've done something meaningful and I'm presenting that meaningful thing to you guys. Not like, "Oh, here is something else that somebody else created that's meaningful. " Whatever that happens to be and let's present that. Like I don't want this to be like my channel is not geared to be an educational channel. It's more geared to be like a me channel and I'm more or less a researcher, you know? And so, I just want to point that out for people who are here and they're looking for content — uh that's like 3Blue1Brown-esque cuz I know there are a lot of people like that. Um that's not the type of content that I think I've ever made on this channel and it's not the type of content that I'm intending to make on this channel. I I'd rather do uh like research-related streams and stuff like that. So, that's what I want to do. Hey Foster, by the way, for people in the US, it's quite early your time, right? Like it's like 5:00 a. m. What time is it in the East Coast even? Um time in Boston. It's um yeah, 7:00 a. m. for people in like Boston, Florida on the East Coast. And for people in uh in California, like I don't know how you're here. I'm so sorry. It's way too early. Um but yeah, okay. So, lots of things I want to talk about today. So, uh as of the last stream, woke up 30 seconds ago and immediately go on to a Alejo's Lab stream. That is uh commitment right there. That's what that is. Um I do recognize one more thing. If I were to stream at a much later time zone, let's say like midnight my time, like midnight to 2:00 a. m. Um or even like 10:00 p. m. to midnight or something like that my time, you know, something that's not ridiculously late, I would be able to hit basically all of the US and the viewership would be much higher because like, you know, there's just more people around who are willing to watch an English stream during that time. Um it's not like I don't want to stream during like this time zone. It's just that this time zone's a bit rough to find people um What is this? I don't think I've ever seen this one. Uh it's a bit rough to find people um who are like live and active and willing to watch because it's just a weird time zone. European people are working. Um American people are sleeping, you know? Um so, it's kind of like a quite bad time zone. Um so, anyway, I'm toying with the idea. idea of doing another stream at like midnight my time uh just to get work done. And the other thing I want to say is I'm also toying with some ideas to uh like kind of get to work a little bit more efficiently. You know what I mean? Like basically, what is the difference between a pencil and a pen? Here's the pencil. And here, see, is the pen. I don't know the difference. I can't tell. Anyway, like maybe like if I don't get work done this stream, we have some sort of punishment or something like that. I don't really know what the stream would be, but it's like, you know, uh something in order to motivate me just a little bit to get like meaningful work done instead of just chatting. Don't get me wrong. I do enjoy the just chats. They're great. But uh I also have some meaningful work I want to get done, you know? Um I don't want this to be just a chatting stream. I want it to be like, "Hey, we're doing something here. " Um — Good morning. What's up? Uh 10:00 midnight your time would be the next day in Tokyo. Uh that's Yeah, that's right. Georgia, I'm never up this early unless I'm still up, which is why I'm rarely here. I get it. Totally I totally get it. 100%. I'm sorry for people in the US. I know it's a tough time zone. Uh why is my chat on screen? It's on not uh YouTube. Right, probably. No, so so so I have two chats. Oh my goodness, what did I do? I Why do I suck so much? This is the button I meant to press and then I just ruined my life by clicking something I didn't mean to. Okay, look. Um so, I have YouTube chat hidden behind my face and then I have Twitch chat up over here. I just because I have two chats and I don't know how to show them both. And it's usually true that the Twitch chat's a little bit more active than the um — uh Twitch chat. Anybody who wants to know how much time I've been live, I you can look, I'm sure the browser will tell you. I don't know. Do do do. Interesting The difference is the second one is a laser pointer, I think, and it will not persist. It won't persist? What? You guys were noticing this and no one told me. — I'm just an idiot, aren't I? It's a laser pointer. Why does it say

pen? It says pen. — What time is it there? It's currently 1:30 in the afternoon, 1:00 1:20, something like that. — Well, that was easy enough to understand and I I'm just too stupid. Um I mean there's I don't know how they were supposed to notate it, but I feel like this notation is not good. Um Okay. Last thing I want to say is let's talk about things that we're talking about yesterday. So, I've been working on this paper. I I've been consistently working on this paper on stream, although the work has been slow and that's okay. It is expectedly slow when we start. Uh so, let's see the projects. I'm already there. So, let me just go and open up ocular main. pdf. So, for those who weren't here for the last stream, which again I is not live on YouTube, so I totally understand why you weren't here. Um the we were talking about a bunch of different things that we could do. And one thing I said that we could do is we could use a signed distance field. Um this is a way of rendering pure mathematics. It's a very well-known method uh that people use for like Oh, it's like shader toy and stuff like this. And we could just render that to a square, similar to how I render these objects here. And then we could just do spatial manipulations on that square. And again, because of the way that the method works, you can get like any anything you want. You could, you know, poke holes in it, you can duplicate, you can do whatever you want, right? So, it's um you know, it gives you a little bit more power than what you're used to with signed distance fields, which typically require you to kind of add a bunch of Euclidean objects together and smooth them together in a way that is just quite unintuitive, at least for me, right? Um so, um with that in mind, um I thought basically putting a signed distance field onto a square and then showing these transformations that you can do more easily with the method that I'm proposing here would be a good uh it would be good figure. But after a lot of discussion, we kind of realized I mean — I'm already showing that you can just create uh like a fragment shader and then put that onto a square and then manipulate that. So, going further and showing that you could do um a signed distance field isn't particularly worthwhile. And there's other things that are probably more worthwhile. And so, it brings me to other things that I would rather show. And one of the things I mentioned is that we could Okay, that's my stupid I Okay, that will turn into a figure. It's a sketch of a figure. But one thing that I mentioned was um it would be nice to show something that's difficult to do with other methods. So, one thing could be caustics. Um caustics aren't particularly hard to do. I mean they're easy enough. But it's nice to show. It's easier to do with this method than something else. — Um the other thing would be moving a hand around, showing some form of rigging. Um and then the Let's see, there's one other thing I was going to mention. Oh, lip-sync as well for some form of meta-programming would be great. And stuff like this. So, I think at the end of the day I settled on um uh rigging point clouds. I still want this mirror frame. I as much as, you know, maybe we talked it down about it the other day. I think it's important just to show that look, we can improve other methods by incorporating it here. Like one of the things I'm proposing here is to use the method that I'm proposing as a middle layer. Like the idea is you render to this and then you get these additional transformations for free. — Um in a much more straightforward sense, right? Um so, I want to show rigging point clouds and meta-programming difficult animation workflow. So, these are the two I want to show. So, this brought up some discussions about how do you rig a Gaussian splatted image? Um and that is a good question. Let me go ahead and catch up on chat. I just woke up. What time is it? Did you get Twitch fixed from yesterday or is it just Twitch? I don't know what's going on with Twitch. Test test. I thought you were Spider-Man for a second. What is going on with the Spider-Man comments? I used to get those like in 2016 or something. I've not gotten this comments in a very long time. Um I mean I guess I appreciate it cuz Spider-Man is like an actor and actors are cool. Um but like I this is it's an old comment that somehow is coming back. Maybe I am um uh I'm like missing something. You look like Peter Parker. Maybe. I guess I still I guess that's a compliment. I'll take it as a compliment. Uh 10 minute your time would be next day. Interesting difference is the second one is laser pointer. Okay, yeah, you got it. I was catching up on YouTube chat, but I already caught up on YouTube chat. Okay, good. Um — Is it cuz I shaved? Is that why? It's probably cuz I shaved, right? So, I you can't see it. I'll I'll explain why I shaved. Give me a second. So, there's two reasons why I shaved. One of which is more embarrassing than the other. The first which is I just forgot what my face looked like. Okay, that's number one. Um so, I guess there's three reasons why I shaved. Um but uh I have scars right here. I don't know if you can see it on camera. Like there's three scars right here from when they they stuck um uh a line into my jugular for like a week. It was just sitting there. Like I had wires sticking out of my neck for a week when I was in the hospital. And I noticed that I had like dots there and I wanted to see how bad the scarring was. So, I just wanted to make sure I shaved clean just to see what it looked like. Um but also I said that was number one. Uh so, number one is like, you know, I just wanted to see what my face looked like and that. Uh but number two, I really need a haircut. Like desperately need a haircut. I Like here, let me just show you really quickly. I and I'm going to have to tie it back up cuz it's insane right now. Um so, keep in mind I have not done anything in uh How do I put this? Uh I have not been able to get a haircut this whole year because I've been sick. And now my hair is insane.

— It's never been this long before in my life and it's quite rough. So, I was like, how do I cut my hair? I don't know. I've no idea. I don't know what to do with it. Um but I really genuinely Cuz I actually do like the hair is the problem. Uh but I can't I genuinely I can't What? I mean, like I could cut it short like I used to have it always. But then I lose this and it's all shiny and I like shiny hair. So, I don't know what to do. I genuinely Uh Just cut 15. Yeah, I could cut it here. That's what I was thinking, cut it here. — No, the issue was it was at a length that I was happy with. Like here-ish uh as of October of last year. And then what happened was I was sick for 6 months and I have not been able to get it cut. So, uh it is uh it's a problem. It is an active problem that I need to figure out. Uh but I don't know what to do about it. So, we are keeping it for today. — Uh has it always been that long? It's been long for a while. — Turn some I could actually. Okay, I need to not do that. Now I need to tie it back up. Okay. Um also, it needs to dry is the thing. I it like the problem with long hair is it takes forever to dry. So, I do have to have it down. As an electrical I have an electric hair cutter every few months just cutting all to 3 cm myself. Yeah, I also was thinking about just buzzing it off. Like just going bald. But I don't necessarily want to do that cuz going bald is probably not the best look on me. Uh Why did I do this in the first place? What am I doing with my life, guys? Like what genuinely? life? Why did I just That was a waste of time. Now, people on YouTube like I've lost all my viewer attention on YouTube. That's what I've done here. That's like the whole thing. That's all I've done. I've lost all my viewer attention on YouTube. Um that's the only thing that this segment has done. I wanted to create a 3D solar system movement simulation, like you know, helical kind of thing instead of 2D projection, but I couldn't figure out the formulation really. 2D is easy, but 3D is a bit over my head. What do you suggest for me to handle the situation? I saw a planet image. It reminded me of this old struggle. Yeah, I've got a question for you. Um why aren't you just following like some sort of like I know I can't do it. It's too long. I was going to try the hat trick. Um — Why is anyone I need to get off of YouTube right now. Guys on YouTube, you guys are going to be gone, I think. — Can I make that look good? No, I can't. I was trying to do the bang trick. I can't do that anymore, guys. Uh let me tie up my hair and then we'll figure this out. Give me just 2 seconds, guys. Jesus. Jesus. I was at the hospital for a condition known as myopericarditis. It's um the inflammation of both the inner and outer region of your heart. Um and I was hospitalized for 10 days and then it's been about 6 months of well, it's been 3 months of recovery, but it's 6 months until I will feel like normal again. So, that's where we're at. As far as the 3D simulations, I don't really understand why you're not just using like good old-fashioned F is equal to GMM over R squared. You know what I mean? Like why not just a normal force law? Is there something wrong with that? Um Yeah. Um because in that case I don't really understand how 3D is significantly different than 2D. You know? That's my question. Um Because really it's the same thing, right? The only difference is that for 3D you have to do, you know, your R is now going to be XYZ, whereas for 2D it's going to be XY. Um I hope you're fully recovered. I'm not fully recovered. That's why I had to take a day off Wednesday and half a day off yesterday and actually some of today off as well. Long hair looked good, but right needs more time to dry and generally more work. Yeah, I do think I like long hair. That's my problem. I genuinely like having long hair. So, I don't know what to do about it. I'll wear a hat though. Maybe that will solve my issue. Um good. So, for your planet simulation, which again I'm just doing something stupid, you know, you have F is equal to G um MM over R squared, right? This is your typical force law. The only difference is that your R in this case is going to be equal to uh square root of X squared times X plus uh Y times Why I can't I talk right now? Y plus Z times Z, right? Like this really is the only difference between 2D and 3D. So, I'm curious as to why you're having Thank you for letting me know that. There's a button. Here it is. So, you have F is equal to GMM over R squared. That's your typical force law for gravity. And then you have R is square root of X squared, Y squared, and Z squared. Um that's the only difference between of Ah I think I need to remove my YouTube uh VOD at this point. It's

That was pretty embarrassing. — Uh I have this problem where I don't like having my VOD up on YouTube unless it's like absolutely perfect or like I feel like it's like compelling good content that you know, the whole way through. So, if I feel like I lost energy, I just don't like um don't like having it up on YouTube anymore. This type of thing. So, I uh I tend to — um I tend to remove my VODs on YouTube way too often. That's it's a me thing. Um okay, so, uh what am I doing What was I going to do? Yes, yes, yes. This is what I was going to do. I messed up. I messed up big time. There was a Peter Parker comment which got me on like why I shave, which I think makes me look like Peter Parker. And then there was the whole um Yes, so many comments. So many What am I doing with my life? Okay, look, here's the deal. Right. I was looking up how to rig a point cloud. That was what I was looking up That's uh that's before stream. How to rig a point cloud. So, here's the deal, right? Let's say we have points in space. Okay, why space, right? Here we have a whole bunch of points. Point, point, point, point, point, right? Now, typically, if you want to rig these, you might create like I don't know, something like this, right? And then you can move this up and down, right? And this will allow you to move the points around, right? Now, this is what we typically do for a mesh. And we just kind of chain all of these to the um Sorry, we points and these would be vertices in a mesh to the um uh to the skeleton and then we move the skeleton, right? It's not particularly complicated to do. Um but there was one other thing that I was going to say. Oh, right, right. Uh the reason this works for a mesh is because in a mesh we have triangles that are all connected, right? So, connected, connected, connected, connected. So, if you move this point over here, you know, this way, that's okay because the connections are still connected. And so, you can connect All of a sudden the mesh warps, right? Um and so, because This allows for you to create a skeleton that um dynamically moves. Why can I not do it? Why do I suck so much at everything? Geez. Um create a skeleton that can then move the mesh and you can basically make sure that all of the pixels are colored in the right way. Right? Uh the problem with doing the same with a point cloud and specifically Gaussian splatting, which is just rendering the point cloud, is like, you know, if we move these points around, right? There is no connection between the points. So, if we move this arm over here, right? Uh we move the skeleton, we have to make sure that the points are um still going to render this gap that you've just created, right? Do Does that make sense? So, with a mesh, the reason we can move the skeleton around and then move the points around is because the mesh is connected. But here in the point cloud, the points are not connected. Therefore, if you move the mesh too much, you're going to have to like overpopulate a particular section to make sure that like it's covered when it goes onto screen, right? Um the advantage of Gaussian splatting is that like, you know, every dot will turn into a Gaussian of some shape and size. So, there are a couple solutions to this. Um the first solution is, okay, if we see a significant warping of um Sorry, let me explain this a little bit better. So, what happens is if you have a single point, when it's splat onto screen, right? With Gaussian splatting, it splats like a Gaussian, right? So, it looks something like this, right? Um and another point would splat over here as a Gaussian, right? But let's say you have to move the points uh like stretch the points a little bit like this. Well, one thing that people do is they say, okay, well, how about if we know we have to move in the X direction, what we're going to do is we're going to splat like this, you know, we're going to make this splat a little bit more elongated along X, right? So that we deal with the difference in the uh like the gap that appears in the point cloud by stretching the Gaussian in the right way. Um the other way is to overpopulate, right? To say that, okay, well, we know there's going to be a bend here, so let's overpopulate this area so that if we do the bend, it's not going to like create a gap in the final resolution uh in the final image. This is why when it comes to like Gaussian splatting, um it is really good for mirroring an environment. It's not so good for uh like dynamically creating a scene that you can then render for like some 3D environment uh for like Pixar films or something like that. Like we This is one of the challenges. It's not even a huge challenge. There are solutions to it. It's just figuring out how to like populate the right area depending on whether there's movement here or not. This type of stuff is kind of complicated. The solution that I'm proposing in the method that I'm proposing here is to say, look, we know what the functions are that you're going to move stuff around with. So, let's overpopulate very specific areas and then create a distribution that allows for overpopulation so that when you do that bend, it's going to make sure that it uh renders to screen. It's okay if like two points happen to collide. There's no problem with that. Um Hey, I just joined. Do you plan uh NeRF with auto rigging? No, I have questions about I So, the thing is, I understand splatting. Splatting is very easy. There's no problems with splatting. I don't really understand NeRFs as much as I would like to. So, for today, I kind of just go through the NeRF paper, not the splatting paper, the NeRF paper, um because that's I don't really understand it as deeply

as I wish I did. Um and then go back to the splatting paper and just kind of look around. I I want to see what people are doing in this space right now because one of the demos I wanted to show for my paper is kind of some sort of rigging thing and I just wanted to showcase that. Um cool. Let's see. Is this Newton's laws at least if you don't want it super realistic with relativistic effects? I did full galaxy simulation some time ago. Uh thanks. Yeah, it was interesting to see emergent behavior. This part was not implemented uh explicitly. I'm going to take a look at this. All right, let me uh let me copy link. Let me see. Um paste. Uh what do we have? Do you have Oh, the YouTube. Uh one second. Wait, how many particles was this again? I We talked about this I think a while ago. In fact, maybe you I see that's what I'm doing here. Maybe you um you were doing it when I was doing my own like 4D galaxy simulations. I don't remember. Uh but like Yeah, good. Cool. One second. I got to get my board away. Uh ba ba. I bet you're fun at parties. That was not sarcastic. You're funny. I was not intending to be funny. Uh I am intending to be very serious. So, I apologize if I'm coming off as funny. — What was that? What am I talking about? Um only a thousand particles. That's actually quite good. I was expecting that each particle would represent like a meta particle or something. But yeah, a thousand, I mean, that's pretty cool. Um I asked you about specializing in GPU programming maybe a year ago. Soon I'll be doing a PhD focusing on ML for chemical reaction networks in Julia. Oh, cool. These networks are faithfully represented by direct hypergraphs and are a generalization of direct graphs. I will be focusing on visualizations first. By the way, despite doing ML, we have very critical approach to LLMs together with a professor. The professor is doing AI ethics work and I'm interested in personal LLM hazards. Look, here's the thing. I've been very outspoken about my um not about my feelings on AI and I realize that this has come off as very negative to a lot of people. But I don't really see AI in general as a negative thing. Um it's just specifically LLMs I think are very difficult to integrate into many workflows and you have to be very careful about how you do that. That's been my take. When it comes to the stuff that you're going to be using for chemical reactions, I mean, we did this before in like Molly. jl. Um Molly. jl is the package that I've been working on or I've worked on before. Uh and it's a molecular dynamics simulator in Julia and doing like machine learning potentials, MLIPs, and stuff like this is um you know, it's uh that type of AI is really good because it allows for you to get like a heuristic approach before doing like actual physical simulations on top of it. Like integrating AI into normal scientific workflows, obviously I've no problem with that because for the most part, you're not going to be integrating an LLM into like a physical simulation or chemical simulation. Um so, like it's just a different thing, you know? And obviously, you know, if your goal is to create a heuristic, you know, to kind of like guess as to what an initial state is something like it doesn't really matter. I mean, you're just creating an initial state and you're finding a good starting point, right? Um ba ba. It is a bit unrelated, but what is your recording setup? How do you overcome overheating? Uh the computer is not in this box. So, I'm in a box um and the computer is not in this box. It's outside of the box. So, there's no heating there. Um I just overheat though. That That's just what I do. I just overheat sometimes, unfortunately. Um my frame rate is at 25 frames per second because if it's at 50 frames per second, which is what I used to do, um then the camera would just overheat. Um that's unfortunately what happened. So, it goes to an Elgato something. I don't know what it is. Elgato stream capture HD something like that. Um and uh I just have to limit the frame rate to 25, otherwise it will have issues. One of the goals real time couldn't do much more because of the naive Oh, right, right. You weren't using uh Box-Muller or anything like that. Box-Muller, no, no. Barnes-Hut. Why did you say Box-Muller? It's a completely different thing. Why are you in a box? Are you being shipped somewhere? No, I'm not being shipped somewhere. The box is kind of for sound protection um because like I said before, I have plans to stream at kind of like midnight and I'm in a city. And so, I have to make sure that I'm not like uh disrupting my neighbors. It's not very good for sound protection though. Uh it's mainly just a recording setup and to make sure you guys don't see ba ba da ba ba. Oh, you can't see it. Anyway, it's a mess. — It is an incredible mess back there. Um not overheating uh our room. Yeah, yeah. Okay, uh 5D points XYZ and transform index. The index of points uh index The index points to a transform matrix. Wait, why do you need to index it to a transform matrix? Am I being stupid here? Um Good. Okay. So, this has been a very rambly start. This has been probably one of the most I mean, it's difficult starts to stream I've had in a long time. I mean, I was asking how you overcome camera overheating. Yes. So, let me just be very clear about

this. To overcome camera overheating, I limit the frame rate of the camera. That's what I do. I'm using a Sony 6100 Alpha camera. I can do this with it because Hello. — How are you guys? I hope you're having a great day. All right, I'm going to go back now. — Um so, it's a nice camera. I can do some stuff with it. Um but uh yeah, I go that I put that through Elgato. The camera does have overheating issues every now and again. Um and then I uh I just limit the frame rate and that's how you do it. Sometimes I'll put the Sony Alpha 6100, I believe is what it is. Um it it just it's one that allows for me to just send it through Linux without me thinking about it. So, I don't have to worry too much about the details. — Um I I definitely looked up what camera to buy that works with Linux and can be captured properly and all this kind of stuff. I should have gotten a different one looking at it now. I hadn't realized that overheating would have been that big of an issue, but now that I know overheating is an issue, I should have been more careful about it. Finally, Cosmo Druid. Um just a mention, I'm very happy that you're doing this. I don't know what package you're going to be working on, but I would recommend that if you are working on a package in the Julia ecosystem for chemical simulations, I don't know what you're doing, but I would really recommend working with um uh Julia MolSim community, right? They're great. Uh honestly, very great people. I had I really enjoyed working with uh Joe and um uh Michael as well. Um And a lot of people in this group. Like I really they were a great guy a great group of people and I think they can probably help you, right? Um so, I would reach out to them and I'd communicate with them is what I would if I were you. Um Science in a Box. Thank you. Okay, this too rambly, guys. Uh this VOD is definitely going to have to be removed on YouTube, guys. It YouTube. Um so, what was I doing? I wanted to show you guys a couple papers that I was looking at before stream. Where are those papers? Here they are. So, here's the paper Drivable 3D Gaussian Avatars. And then there was another one. This is a NeRF paper that I want to take a look at. Uh and then here's a survey. Um and if you go through the survey, they only got like one mention of rigging, which is here. And I did want to kind of take a look at these guys. Well, I just want to see what people are doing in the in the space here. Uh my understanding is like you have um — like when we were doing the archive stream the other day, we actually found a paper that was already doing this type of rigging thing. I just want to see what they're doing. I don't want to like propose a solution if the solution's bad, you know. Um I know for example, just using um SDFs and rendering that to like a uh a point cloud in the way that I had uh said before allows for you to do some smear frame effects and stuff like that you can do before. So, there's a reason to do this. Um and I can provide like meaningful advantages. Uh but for rigging, I I just hadn't thought about it very much until we talked about it on stream the other day. So, I have to dig deeper into that. I definitely need Okay, I don't know if Did you guys hear that? You might have heard that. Okay, anyway. Um So, let's go here. Ahem. Drivable [clears throat] 3D Gaussian Avatars. We're going to start here. Uh I'm probably going to look at the Gauss uh the NeRF paper. And actually probably let me just go get the Gaussian Splatting paper as well. Uh scholar. Oh, of course, I'm not I'm on the wrong browser. Bop bop. Uh one second. Uh recent advances in 3D Gaussian Splatting, Survey on 3D Gaussian Splatting as a new era, a survey Gaussian Splatting SLAM. This None of these are what I am looking for uh at all. — Maybe I should just go here to the survey and I should go to the top and they have to have it cited. Um the objective of image-based 3D scene reconstruction is to convert a collection of views capturing a scene into a digital 3D model that can be computationally processed and analyzed and manipulated. Uh the hard and long-standing problem is fundamental for machines to comprehend real-world environments facilitating a wide array of applications such as 3D modeling and animation, robot navigation, historical preservation, blah blah blah. Um Where do we have it? Um okay, this is starting way too early, man. I just want the paper. I want the paper that I'm talking about. Like everybody knows the paper. Give me the paper. Um a Julia MolSim can be used for our project. Well, it's not necessarily could it be used. Uh I'm talking about maybe proposing a collaboration with that group, right? Like I do believe you're probably going to have to write your own software, but I also believe that it might be possible to get their tools to help you with that process um and to integrate with that group. Like it's always my opinion that especially with researchers, it's best to collaborate instead of compete. And specifically in the Julia ecosystem, you have all the tools that you need to allow for that uh collaboration to happen. And so, there's no reason for you to like How do I put this? There's no reason for you not to work with people if they're doing something similar, right? Uh it's just I've spent time uh some time. It's like the paper the one that like people weren't thinking about Gaussian Splatting and

then sometime in 2024, the Gauss there's a paper that came out on Gaussian Splatting that showed it in real time. There's like a specific method like a specific paper that like I'm sure if I just look up Gaussian Splatting, I'm going to find the one that I'm thinking about. Um Gaussian Splatting bop bop Uh let's see. Now, we're talking about with NeRFs. This is the problem, right? Like obviously, it's one of the earliest rendering techniques, but we're talking about specifically having that real-time example in a browser. Where is it? Um Oof. I should have brought it up before stream, but I didn't really think about it. Uh 3D Gaussian Splatting. Maybe it's this one that's the thing I'm thinking about. Uh 3D Gaussian Splatting paper explained. Uh Mm. Maybe it's this 3D Gaussian Splatting paper I'm talking about. 3 3D? This is I'll find it. I'll find I think I found it. This one. This is what I'm talking about. I'm sure It's like the big paper in the field. This is the one. No, that's not the one. It's a slide deck. You telling me a slide deck has 8,000 citations? 9,000 citations? Oh my goodness. — Why do I suck so much? Okay, they must have the actual paper like not in slide deck form. Why is this a thing in computer science where like for some reason the slide deck is cited more than like the paper discussing the thing, you know? Is it because everything is like um uh like based on conferences and stuff? Um Reading papers is one of the is one of his favorite activities. Uh let's see. What do we have here? Uh original reference implementation. Give me the paper. Here's the paper. Thank you so much. That's what I'm looking for. I think this is Let me go down. I'm looking for a particular image. Is this the one? Uh Is it? — Somewhat Oh, yeah, it is the one. Yeah, that's the scene I was remembering. Okay, perfect. That's perfect. Perfect. Okay, there was a question here. Um isn't rigging Gaussian Splat easy? Uh you transform the points and its Gaussian since the transformation matrix is affine, it should be trivial. I agree. It should be trivial, but the thing is you will have gaps in the mesh, right? Or the the point cloud, right? It's the question how do you like either overpopulate? What do you do um in those cases, right? Because, you know, it's a point cloud, not a mesh. The points are not connected, right? That's it. Like that's the thing. Like it's not a hard one um to uh thing to understand. I just want to see what people are saying about this cuz I also thought it was a relatively trivial thing to do. But then I realized when we were reading the archive the other day that there was a paper doing this rigging of um uh a point clouds in a you know, a 3D Gaussian Splatting environment. I was like, "Oh, okay. I didn't realize this is a hard problem until I saw it on the archive. " I was like, "Oh, okay. This is a hard problem. " Um do you work on meshing? I don't work on meshing. I do computational science. Uh in particular, I've been working in the Julia lab for the past 5 years. So, I do kind of like GPGPU engineering, research software engineering. The stuff I brought up with Molly was something that I'd worked on before, which is a molecular dynamics engine. I work with the Julia GPU ecosystem. So, I've directly contributed to like GPU arrays. jl and um kernel abstractions. jl and like these packages for easy metaprogramming and stuff like that in uh in GPGPU stuff. So, this is actually my first paper in graphics proper. Everything else I do is more computational science, right? Um so, for that reason, I understand there's a lot that I have to learn and I'm not sitting here pretending that I'm an expert. Um but what I am saying is that like I'm happy to push and learn new stuff in this area. Um that is something I haven't made particularly clear on YouTube. Uh and I think people on Twitch automatically understand it and probably based on my persona, you could probably get it just based on like what I'm saying on stream. But what I but it's very clear I'm not like a like key expert in the area. Like I have a paper that I want to publish because I think it's cool. I learned a lot while doing it. Um and it seems like it has a place. It might not be like the greatest paper ever, but like it seems cool, right? Um and I have additional things that I want to do with it um like formulations for quantum uh tech and stuff like this that are cool and I want to do that. Um and so, to do that, I have to get the paper out. That's kind of where I'm at right now. And to do that, I want to have like another like good application. And one application I thought was the rigging of these point clouds. Um here we are. Perfect. So, let me go ahead and take a look at this, right? We present Drivable 3D Gaussian Avatars, a multi-layer 3D controllable model for human bodies that utilizes 3D Gaussian primitives embedded into tetrahe- tetrahedral cages. Can I talk about science communication really quick? So, I had I was that guy, you know, um who would always use the largest word you could possibly come up with at any point in time just because that's what my brain would naturally like conjure up. Um And then I went to OIST, the Okinawa Institute of Science and Technology, an English-speaking institution but with most people there probably being non-native English speakers. And I realized there's certain words you can just cut out of your vocabulary entirely and it's going to make your life way better. One of them is utilize. There's

basically no situation where you can utilize the word utilize where you could not utilize the word use instead. Like for example, you could say "3D controllable model for human bodies that uses 3D Gaussian primitives embedded into tetrahedral cages, right? " Like basically, you every use of this word utilize can be replaced with use and it will be more understandable by more people. It will just sound less technical. It'll be easier for people to grok like everything, right? There's just certain words that if you hear it's like "This guy's This guy utilized utilized, yeah? " Um but it's posh and sounds experty. Right, right. That's exactly right. If you hear someone using the word utilize, you kind of know the person. You know, they're trying to sound smart even though like they might not have anything smart to say, yeah? I don't mean that necessarily because there's a lot of non-native speakers who just use the word all the time. Um And that's totally fine and valid. I'm not saying there's anything wrong with it. But I'm just saying that like, you know, it's certain words you can cut out. Like that's one thing that OIST really taught me is that science communication is not the art of being the smartest person in the room. It's the art of like I mean, science communication was never the art of being the smartest person in the room. But the ability to communicate with other scientists, you know, it's okay to use If you don't remember the word but you remember the concept, you can say wibbly wobbly timey wimey. It's okay. People It's fine. You don't have to use like the correct term all the time. Just get the point across, yeah? And as long as you get the point across, it's fine. I mean, most scientists are non-native English speakers, you know, it's their second language anyway. If you say wobbly thing, everybody will know what you mean by wobbly thing, you know? Like if you forget the word for a particular algorithm, totally fine. It doesn't matter. Say like the algorithm that does this thing and they'll be like "Oh yeah, I know that thing. " Yeah? Um and that's kind of how it works, right? You don't have to sound smart all the time, you know? It's perfectly okay to say things like "I don't know. I think it's probably this. Who knows? " Like you know, it's fine. Just just talk normally, yeah? Um so, the advantage of using cages compared to Did I miss something? I feel like I missed something. We provide drivable 3D Gaussian avatars, a multi-layered 3D controllable model for human bodies that utilizes 3D Gaussian primitives embedded into tetrahedral cages. There we go. The advantage of using cages compared to commonly employed linear blend skinning. That's what I thought. Yeah, that makes sense. Is that pri- primitives like 3D Gaussians are naturally reoriented and their kernels are stretched uh via the deformation gradients uh of the encapsulating tetrahedron. This is interesting. I want to look at linear blend skinning. That's what I want to look at, right? So, this is already has given me information that I didn't have before. I should have known this before, but I don't, right? This is me entering a new field. Um what does this mean? Uh linear blend skinning uh is the idea of transforming vertices inside a single mesh. Okay. Stupid question. Am I reading a dumb paper? Is this not what I thought it was? Cuz I looked down at this figure and I really thought it was what I thought it was, but this might be a different paper. I want to see them basically move points around as if it were a mesh. That's what I want to see. I'm going to see if that's actually what this paper is if or if we have to find another paper. Um What do we have here? Trump voice, I've utilized Gaussian splat rigging eight times. Nobody has ever utilized it even once. Maybe that's what I'm learning though. Maybe I'm learning that this is actually a little bit more difficult than I thought it was. Um which is fine. It's totally fine to have a problem more difficult than I thought it was. That makes it a valuable problem to look at. Um but like I Like what Hannah said earlier, it sounds trivial. So, I do I am a little concerned if there is some difficulty here that I didn't think about before. I was watching PBS Spacetime video about Earth's core and they use the term squigy uh for a type of matter with certain properties. Yeah, but it's a squigy matter, right? Like That's a good word for it. Um — [clears throat] — I don't know. I think the less scientists take themselves seriously, the more easy it is for people to see them as people and the more easy it is to see scientists as people, the easier it is to engage with what they're doing, yeah? Like that's what I believe. Um Hence why I'm streaming. Really and truly, that is why I'm streaming. Like I started streaming a long time ago as a PhD student because I wanted to do research on stream. Um turned out I did pretty well at the time, you know, considering all things considered. Man, this chair needs to go higher. Okay. And [clears throat] now we're here years later and I am now a failed YouTuber. — That's my career path, a failed YouTuber. Um Is it primitives like 3D Gaussians are naturally reoriented? Uh additional offsets are modeled for tetrahedron vertices, effectively decoupling the low-dimensional driving poses from the extensive set of primitives to be rendered. This separation uh is achieved through the localized influence of the tetrahedron on 3D Gaussians, uh resulting in improved optimization. Using the cage-based deformation model, we introduce a compositional pipeline uh that decomposes an avatar into layers such as garments, hands, or faces, improving the modeling of phenomena like garment sliding. I feel like this is a weird approach. Am I wrong? Am I stupid here? Like basically what I'm understanding is they encapsulate these points in the point cloud into like boxes and then they move the boxes around, which is fine. There's nothing wrong with it, but

I'm wondering is this actually the best way to do it? I don't know. I feel like it's not necessarily the best way to do it. Uh these parts can be conditioned on different approach uh different driving signals such as key points for facial expressions, uh joint angle vectors for garments and the body. Our experiments on uh two multi-view data sets with varied body shapes, clothes, and motions show higher quality results. They surpass PSNR and SSIM metrics um of other SOTA methods. What's SOTA? Did they Stop using acronyms without defining what the acronym is. State of the art. That's what it is. I should have known, but I didn't know. Um using the same data while offering greater flexibility. If we're doing point cloud motion interpolation, we could also look into parallel transport or optimal transport. Now I know I um What do you mean by this? What do we mean? Uh mines and factories? Is this what we're talking about? The Hitchcock problem? What? What are we talking about? What is the parallel transport optimal transport problem? Um In mathematics and economics, transportation theory or transport theory is a name given to the study of optimal transportation and allocation of resources. The problem was formulated by the French mathematician Gaspard Monge in um — 19 1781. In the 1920s, A. N. Tolstoy uh was the one who to first study transport problem mathematically. Is this the one that we're talking about, Bolster? If it's something else, I'm happy to look into it. Um Okay, right, right. As um which paper? This paper. There we go. Uh developing drivable photorealistic human avatars is crucial for better long-distance telecommunication that provides an immersive experience to the users. Is it though? Like is it actually crucial? Like that's my I might Is any of this crucial? Uh Demonic Draconic, thank you so much for following. Really appreciate it. Like crucial would indicate that like human lives are at stake if you don't get this working. This is Hmm. I I'm My question is I see this as incredibly useful for other effects like creating appropriate starting points for molecular dynamics simulations and stuff like this. I can see it as quite valuable. Um But if it Yeah, anyway, let's keep going. I'm digging at stupid sentence. I shouldn't be digging at this. Um the motion and deformations across various segments of a complex avatar's body are influenced by distance signals such as facial expressions and body movements. This complexity poses challenges for accurate modeling using a single layer, right? Multithreaded avatars uh become essential to represent these different regions, ensuring proper motion and visual fidelity. Uh similarly, garments present challenges such as sliding, necessitating uh separating uh models of each clothing piece. Yeah, how do they deal with like You know, like the big problem with clothing is that like you have clipping, right? How do you deal with clipping in this case? I don't really know. Um the general idea of area of study is how to map one point cloud to another while maintaining certain properties or minimizing distance. Um No, that depends on the context. Okay, there are two statements there and so, first of all, the general area of study for transport, is that actually the area for transportation theory? Um Have you been keeping up with your alchemy studies? I have not, unfortunately. If there's something crazy in the field of alchemy, feel free to send me a paper. I'm very interested in what's happening in alchemy in 2026. Um One second. Clipping is solved with better cloth dynamics. Yeah, that's the thing though, right? I feel like hmm I feel like when it comes we spend too much time dealing with I I'm not going to talk about this. Basically, we spend way too much time dealing with cloth simulations. Like we spend too much energy in cloth simulations. Um as proven by deep by VTubers, you only need an avatar to represent yourself. There's one How do I put this? There's one dream of mine and that is to actually create VTubing software. I want to write like a very simple piece of VTubing software so that like I can do The idea is to do a like a digital mirror of my immediate environment um and then to find some way to learn what my specific point cloud looks like, right? And then to replace those points uh with like the points dedicated to uh some avatar or something and move the points of the avatar around instead of my own points, right? Like I really want to be able to do that cuz I feel like using lidar and stuff like this in order to create the appropriate like this is something that people are kind of doing. They're doing digital twinning. They're doing all this kind of stuff. I feel like you could do this, you know, for like really good VTubing software, I think. Like there's a way to make this work. Uh new Grim War just dropped. Hasn't? We must solve it. Albert Einstein turned pitbull into gold. However, it was only the bones. Oh, no. I feel bad for the pitbull. You want to VTube, too? Sign me up as a programmer. I mean, like who doesn't want to VTube? If I'm being honest, like genuinely speaking, VTubing is so cool. I don't know. I know some people find it weird, but VTubing is so cool, right? Like

Like Am I dumb? Like way back in the day, I actually created an avatar in like in my very old animation I think it was Chi the L at the time. I don't remember which one. The Cairo visualization library. With the sole purpose of VTubing and I never did it. Just cuz I didn't I you know, that was way before VTubing even existed before it popularized or any of this kind of stuff like Holo Live and Kizuna AI or whatever. Um Kizuna Ai I think is actually her name not AI. I think I have to be very careful about that now. But yeah, it's I don't know. It's I think it's really cool. We must solve it. Albert Einstein VTubers remind me of me characters very scary. VTubing is really cool. I don't have an application for it, but still very cool. Yeah, I mean it's one of those things. I like The thing about it is I would definitely be a VTuber if I could get away with being a VTuber. Like I definitely would, but I can't get away with being a VTuber because I want to do research and what am I supposed to do? Like publish under an anonymous identity? I can't do that. Let me know if this could help. Let me see. Copy link. Paste it over here. Um There we go. Optimal transport in a nutshell. Optimal transport is mathematical framework that finds the most efficient way to transform one probability distribution into another. Oh, it's not just point clouds. It's probability distributions. That's actually useful. No, no. I've heard about Yeah, yeah, yeah, yeah. Yeah, I know exactly what you're talking about. And this I need this for another thing that I'm working on right now. Actually, I do need this. Let me So, we're going to keep this, but not necessarily for today. Maybe for today. But probably not for today. But I will need this. Thank you. Um If Satoshi Nakamoto is here today, he would be a VTuber. I don't know who Satoshi Nakamoto is. Sign me up as a programmer. Anyway, I don't know who that is. I maybe I should. I'm not like, you know, I lived in Japan for 5 years, but I'm not like that into like I mean, I watched anime. Like it's not like I don't watch anime, but it's just like, you know, I'm not so deep into it that I know everybody. You know what I mean? Like I'm not Um like I'm definitely a nerd. I know anime. You know, my favorite movie is Koe no Katachi, which is, you know, an animated movie off of my favorite, you know, written work, which is Koe no Katachi the manga. But you know, I'm not like I I'm not into the shonens as much as other people are. Okay. So, why were these guys clumped together? Is that what happened? How did I do that? What what does this mean? Okay, I don't know what's going on with my browser right now. Um Sounds like an additional input source that could use one. So, like MediaPipe for the general face hand position and then lighter data to morph and blend shape. Otherwise, you'd have to figure out how to map the point cloud to the bones of a VRM model, which would probably require training your entire own regressor. No, no. My idea is much more simple than that. My idea is tool Well, let me do journal new page before. Here's the stupid idea, right? If everything is a point cloud, then all you got to do is the following. You first create a digital twin of the environment, right? So, that's step one. And you hope the environment isn't moving too much, right? That's one key aspect of this that I suppose maybe I should be more clear about. So, you have the environment here and you assume it doesn't move too much, right? And then if you have a point cloud that represents the person, right? Here's your Um Then this is an awful drawing, but I'm really hoping you're following. Let me try this one more time. I need to erase this down here and I need to do this in here, right? Then it is much easier to identify what the person is because you have um Uh you have a digital twin already of the environment, right? And so you can basically exclude points and you can more easily identify the person, right? Now that you have identified the person, all you have to do is find some form of transformation. Um Uh Uh one second. Uh from this point cloud to the point cloud of the anime avatar that you want, right? I know this sounds really stupid, but if it's a 3D model like it's not actually so difficult to create like a simple transformation. Uh like you should be able to learn this, right? It should be a you should be able to learn the way in which these points kind of go to the person and then you have a map, right? You might have to have the person like stand in a very particular pose or something like that in order to get to sync to begin with. But then you can have a map of the points. But the problem is because Gaussian splatting works by creating new points. I haven't gotten that far. far, but you should be able to create some sort of field that allows for you to map the movement of each point. Like, you know, for example, if here's your hand, right? And you move it here to here, right? Well, you know, you'll see a movement of some mass of the hand moving from like point A to point B. And like I don't know how this is going to work. I haven't gotten that far into the details. But the point is if you can find this movement vector, right? And you can find that movement, then you should be able to create some sort of like You should be able to keep track of the point cloud's movements in such a way that you can also

like continually attach that back to the avatar. And then you should be able to move the avatar with the movement of the hand, right? So, you got a point cloud for the person and you will always transform the point cloud of the person into avatar. So, you have to learn that transformation to begin with. But it's simple algebraic manipulations. Like all you have to do is like, you know, whatever XY position it was before XYZ position is before. You just have to move this into like, you know, X * 0. 5 to shrink it in a little bit and Y + 7 to move it up. You know what I mean? Like it should be simple algebraic manipulations. Nothing too complicated. And so if you find if you can find that movement from one point cloud to the next, then all you got to do is find a way to track the points, which is more difficult than it sounds because the points are regenerated basically every frame. But it should be possible, I think. This is like one of those like I it's such a vague notion. I don't really have it in my head exactly how this is going to work. But to get this to work, I actually have to do two things first. One, I have to get my engine out. So, the current paper done. So, paper one. But then the other thing I have to do, which is actually crucial here, is I have to find um a way to lower the run time. Right now, I'm in a situation where if I want to draw a circle, I overpopulate the circle. And the reason I overpopulate the circle is because even though I'm generating the circle as a uniform distribution, it's a uniform random distribution, right? Which means it's totally possible there's a big gap like right here, you know? And so I have to overpopulate the circle. And specifically if I'm doing transformations on the circle, I have to do like overpopulation in X, for example, if I'm going to stretch it in X or overpopulation Y this way. And so this makes it so it's quite inefficient if I am spending too much time like um if I'm in the transformation scope. Like if I'm doing too much with the circle. So, on the one hand, it doesn't cost too much to do extra computation. On the other hand, if I want to limit [clears throat] that, I'm going to have to find some way to make it so if I do X is equal to X * 2 um that this initial distribution was kind of bigger in the X direction, if that makes sense. But good news is people are reading in these transformations, right? These are read in and there's some form of a D-like code transformation I can do on this to create the appropriate initial distribution here. And I don't have to make that initial distribution random, either. So, paper two would be actually in how do you create the perfect initial distribution such that you guarantee at the end of the day when it's splat to screen that you don't have gaps, right? That's the question. And that one's a lot harder to do. Um but if I can get this done, right? Which again is a hard thing to do, but I kind of have vague notions, then doing these real-time transformations and stuff like that should be a viable target. Right now, it's not a viable target just because it it's a little bit slow with the run time. — It's not that slow, but just a little bit slow. So, it's like kind of like it's just a little bit annoying right now. And so I have to kind of work on that. But the nice thing is that finding this transformation is I should be doable. I'm not going to say it's easy, but it should be doable with the method that I have. So, this kind of paper one, paper two, paper three to becoming a VTuber. — That was kind of what I was looking at. Um And yes, I I'll probably need an additional input source. I'm looking at something like lidar. Uh or something. I'm not sure what you want to do here. Like I mean, there's a lot here that I don't know and I have to learn about it. But luckily, someone that in the comment section even or in the chat, somebody somewhere somehow sent me messages and they actually work in a lab that focuses on digital twinning, right? I was like, this is perfect cuz if I want to get this to work, digital twinning is exactly what I need, right? Um and so I really do want to see if I can get that to work. Like that's one of those like paths of development in addition to some other paths of development. So, those are kind of three kind of papers that I wanted in that direction. Then I've got another paper, which I've told you before I want to get this a quantum formulation for this, which I think is really cool. So, that's four papers that I'm looking at right now that I really want to get out but and I want to work on as well as whatever we can do together in as a community, right? You can create tab groups in Firefox in a couple weeks. The input is a bit finicky for me still. Should be dragging tab above another tab for 2 seconds. I think I just didn't know that. Uh so, you need to normalize the difference of locations to see an average of vector. Right, I don't know exactly. That's the truth. It's all very vague. Like it's just in my head. It's like, maybe there's a way to do this. I don't know the actual details. Um target an EG PNG VTuber instead of 3D model, if I understand correctly. Unfortunately, I think I have to target a 3D model. I think doing it 2D would be a little bit too difficult at this point in time because you yourself are 3D. Well, maybe that's not true. You get two cameras. You could probably target You could target a 2D. But the whole point is to be to have the model more expressive. Like that's the whole point. You should be able to do hand like motion and stuff like that much better. Cosmo Drew, thank you so much for following. Really appreciate it. Gaussian splatting looks like works like expectation maximum. Fits and Gaussians then based on Gaussian size it splits the large ones, prints small ones, then fits Gaussians again. Uh what is this? Follower goal met. I met a follower goal. Hey, great. Mishivy, thank you so much. I didn't know I had a follower goal. — Thank you.

Um Uh fits and Gaussians then based on Gaussian size it splits large ones. Well, I mean the splatting is literally just you take a ball, you splat it on the screen, and then that splat is a Gaussian. But I think you're talking about like how do you create this distribution here, which is a thing I don't know either. Like that was kind of the point of today's stream to learn that. Um so, maybe this is like already solved and I just wasn't aware of it and that's totally fine, too. Um this is me just not uh understanding the details, right? — Um so, that's kind of also the point of today's stream, right? I just need to learn some stuff. Um There's also something I want to point out. I know this is going to sound really stupid or arrogant. I'm not sure what word I'm looking for here, but yes, I'm aware that there is a Corridor Crew video on Gaussian splatting that I am aware I should probably watch. I have a problem in that I can't return retain any information that is given to me through video. I don't know why. Um it's I just can't do that, right? Uh is the Gaussian splatting's papers algorithm? Is that the one? Oh oh, the 3D Gaussian splatting paper. Or the one from like 1930. I'm sure you mean the the 3D Gaussian splatting one. Yeah, that's the one I haven't fully read through. Um in particular, I want to learn more about NeRFs cuz they mention NeRFs a lot in this paper and I didn't know about it. So, I had trouble reading it. Yeah, the paper. Yeah, yeah, yeah. Yeah, okay, then the paper, right? Um I need to I need to go through this one in — a lot more detail cuz I don't know it as well as I should. I feel that some input methods just don't work for learning for me, either. Yeah, specifically video content, I can't learn from video content. I I Here I am making video content, but I just can't. Um Azrael and Verad Rath, thank you so much for following. Really appreciate it. Everyone treats AI like a world-ending threat um when it's mid at best. It's not the threat of AI itself, it's the fact that uh sub-mid people will use a mid AI like it's AGI for decisions which could be Yeah, exactly. Yeah, I agree with that statement. Um I will be honest, AI is a useful tool. I mean, it think anyone who says it's not a useful tool I think is lying to themselves. But on the other hand, like when we're talking about specifically generative AI, um it's not as useful I it's not like it's not as crazy as people are making it out to be, I think, in my opinion. Ahem. [clears throat] Okay, so, let's keep looking at this cuz this is the one I want to look at. Um Yeah, what at is we design a method for the use of minimal set inputs. Man, my throat is just Give me a second. — [clears throat] — Um I don't know if he's asking if there's a ray marching like algorithm for point cloud generation. Um yeah, I mean, there are a whole bunch of them, right? Any of the functional methods that I've talked about before on stream. There was one that came out uh a while ago. Yeah, the point is yeah. Yeah, there are a whole bunch of them. It it's not uh like iterated function systems, for example, is that thing. Um then there are um There was a paper that does exactly what you're talking about doing like a recursive thing. It's like Oh, what is it? It's like it's it was a paper that published in SIGGRAPH a while ago. There's a nice blog post about it. Uh I can't remember the name of this paper. I have it cited in my paper, but there are a bunch of these. Um basically, any method that allows you to draw um based on purely functional approaches is kind of doing the same thing, right? It's creating a point cloud that it's is rendering onto. Um I'm just trying to figure out how to get my brain to focus on written information. The motivation is there, but my ADHD just gets bored because slow uh because slow and loses interest. I'm like directly the opposite. And I know it's like totally maybe it's impossible to explain, but like for me, any type of video content goes in one ear, out the other, right? Like I know everyone loves 3Blue1Brown. I worked directly with Grant for like years in a sense. Um I worked on the Summer of Math Exposition, which is mainly video content. And yet at the same time, I have to be honest and say that I if you give me an explainer in a video form, I'm not going to understand it. I just won't. I I cannot explain why my brain does not process this information at all. Um I wish it would I wish it did. I wish it did, but I it just doesn't, right? And so like a lot of times I have the problem where like, you know, again, I have no issues specifically with Grant. I like Grant. He's a great guy. He makes wonderful content. Um but like sometimes [clears throat] I realize like when I'm watching the video that I've kind of lost track of the plot entirely, you know? And I have no idea what he's talking about um anymore. And all I'm doing is like following the animation. It's like, wait, this doesn't make sense to me anymore, right? And I feel like it's in part because he has a tendency to focus on what makes a good visual and what clicks for him as opposed to like what clicks for everybody. And it gives people the impression that they're learning even if they are also kind of lost to the plot, you know? Um I'm again, I'm not saying there's anything wrong with that cuz I do think it does click for a lot of people. But I think the people who doesn't click for, they don't realize it doesn't click for them until it's significantly after when they have to utilize the information. And there I go utilizing the word utilize. Wow, what a person. Um No, I don't want to swap because it reading papers is fundamentally one of the greatest joys I have in life and I would never give that up for anything. Um like I know that sounds stupid, but the ability to read a paper and just sit down and enjoy it with a hot cup of hot cocoa next to you, genuinely the best experience, period. There's nothing that I I'd rather do than like just go back to OIST

in the middle of a typhoon day, have the you know, the winds and rain, you know, going right against your window, you know, nice white noise going on there. Um your power maybe potentially flickering. You don't know because it's a typhoon. And just having a paper that you printed out the day before cuz you prepared for it and just sitting there reading it. It it's genuinely the best experience I I've ever had. Uh you look like Toby You were like the third person today to say I look like Toby Maguire. I appreciate that. I guess because Toby Maguire was an actor, therefore I I appreciate the comment. Um but I'm not Toby Maguire. I am significantly less talented. Um — [clears throat] — Okay, so. Related work. I think this is where I want to start. Uh D3GA reconstructs controllable digital full-body avatars using multi-view video uh and joint angle motion by combining 3D Gaussian splattings. Um this is probably the paper that we were looking at before. Uh with cage-based deformations, right? Okay, so these cage-based deformations I think are important. Current methods for controllable avatars rely on dynamic neural radiance fields, uh point-based uh These are the ones that I want. The point-based models I think are what I want. Or hybrid representations, right? So, let me take a look at 3575 and 8135 is going to be uh the power of points for modeling human humans in clothing. Okay. Uh what else do I want? I wanted 75, which is going to be uh where are we? Um point NeRF, point-based neural radiance fields. Okay, I think actually that might be worth reading. And then finally, the last one is 81, which is going to be uh deformable point-based head avatars from videos. So, what then how is this I Mhm. Maybe I'm I need to learn more about this. what's happening here. Webs, point clouds, uh close enough? I think there's a fundamental difference between webs and point clouds. I'm not like an expert here, but I feel like there is a fundamental difference, right? Um Uh you are talented if you do tech. No, I think there's a lot of people who think that tech takes talent. It doesn't take talent. It just takes having a broken brain. Like I I mean, maybe I'm wrong about this, but like, you know, like I'm not Michael Phelps, you know what I mean? Like it's significantly harder to be Michael Phelps than it is to be uh a professor at a university, you know what I mean? Um Or it's significantly harder. I mean, being an actor is maybe different because like it takes talent, but it's also a skill that you build. Like if you just see like learning or you see research as a skill that you can build in a similar way as you see athletics as something that you can build up or acting skill as something you can attain, then it's the same thing, in my opinion. I believe visualizations are absolutely critical for math, um but common problem of math videos for me is a lack of dry residue. Uh I need precise and formal definitions, not hand-waving anal- analogies. I think there is a problem with YouTube. Look, if you're making YouTube video, you have to have people watch the YouTube video. Otherwise, what are you doing, right? Like if you get 5,000 views on your YouTube video that you spent 5 days on, right? Or worse, a month on, which has happened to me multiple times. What's the point, right? Because at the end of the day, you know, who is learning from this? You got 5,000 views. That's great. That's a huge number of views. Just for the record, 5,000 is great, but it's like $10, right? And if you spent I don't know, let's say, you know, 48 hours editing, which is uh common for my videos back in the day. It took me 2 days of just purely editing to get it to work. Um and then, you know, a couple days of visualization as well. So, it was not uncommon to take, you know, 5 days in terms of full hours. So, this would be, you know, something like We'll just put it at 100 hours of work, right? Uh and then again, maybe more or less depending on the video, but 100 hours of work was not right for my videos. Um and then you get 5,000 views from that. Well, I mean, $10 for 100 hours of work is nothing, right? So, you do have to get money back. And the way in which you get money back is by, you know, audience retention. And the way you get audience retention is by making your videos clicky so people can watch. And so there's a balance here, right? And a lot of times that balance is at the detriment, I guess you could say, to the people watching the video. They will not How do I put this? They will not be able to retain the information as deeply because you focus on things that are more clicky, if that makes sense. Like you focus on visualizations, which are again, essential, but will not necessarily help you retain that information in the long term, right? That's my opinion. Uh unless your life is to revolve around something. Uh notes to self I'm taking from this. Try to coax my brain with hot cocoa and close everything on my PC but the paper. This is something I have a question about. I feel like I don't know what ADHD is. I think maybe that's what I don't understand. Like what is it that makes it so a video is more engaging than a paper? Cuz for me, I thought I had ADHD because I can't focus on video content. But it seems to me most people with ADHD can focus on videos, but can't focus on papers, right? It feels the opposite to me. But the thing about a paper, which is so amazing, is that every single paper it I I say amazing. Like I'm really hyping this up. The thing I like about papers is that every single sentence is itself like a burst of new information that you wouldn't get before. It's like, you know, it it's like it has to be cited. well understood. What I'm trying to say is that like every time you're reading a paper, like every single sentence it's like poetry. It's like it was composed in such a way as to present a very specific point. — Um and it's up to you to understand what

that point is. And so there's like a little game you're playing as you're reading the paper because like every single sentence is just new information, right? And so if your brain is after new information, which is what I thought ADHD people were, um and what people scrolling on Tik Tac Tac Tik Tok want, and what people on YouTube want, like they want this new information, then papers are the source of new information. Like you just read and every single sentence will give you something that you had not you may not have considered before, right? Which is significantly more. Uh the information density is significantly higher than what you would get on YouTube. And I think that's why I can go through papers, but I struggle with YouTube because the information density on YouTube and lectures, for example, are is a very low, right? And I need that high information density to go through. Um can you explain the process in detail of how you learn through a paper? I guess most people just don't know the options when they get stuck and encounter something they find hard to understand. Hey, let me talk about this. It's good to talk about. You guys are great. These are great questions. Let me talk about this. I found using a screen reader to read with me helps. Um I think it's simply information bandwidth visual auditory textual. Maybe you're right. Maybe I'm like misunderstanding the way in which most people's brains work and mine just doesn't work this way. Um here's the thing about a paper, right? I keep talking about reading a paper because I genuinely enjoy it, but I also have to be very clear about this. There is no such thing as a single paper in isolation, right? When you're reading a paper, you're not just reading one paper, right? The whole point is to build and craft your own narrative. So let me just I'm going to say some stuff that's going to sound very arrogant because I don't have any other way to say it. I How do I put this? This is going to sound really stupid and I don't I can't say this in a way that doesn't sound stupid. So I'm just going to say it with coming to my brain and I'm hoping that the information gets across. So I actually can't read books, okay? I Um and the reason I can't read books is because when I read chapter one of the book, right? What happens? Well, I read chapter one, they set up all of this great story, right? And then in my brain, after reading chapter one, right? I now have all of these different directions that I want the story to go in. I want it to go here. I want it to go here, right? And then I in my brain, I'm already all the way over here. Like I'm like, "Oh, imagine if they did this path. Imagine if these two paths actually collided here. Imagine if this goes over here. " And this is what happens after chapter one, right? And so now, chapter four, chapter five, chapter six go by and I realize they're significantly diverging from this path, aren't they? They're doing something else. They went here, but then they did something that's quite frankly not interesting and boring because this path to me was the path I wanted to see, right? And so my problem with books is that I actually can't focus on them because after going through chapter one, two, three, four, I'm now immediately bored cuz I see where the story's going and it's not the direction I had in my head for the perfect story for me. You know what I'm saying? The same is true for videos and in movies and stuff like this. If I watch a movie, specifically I remember The Dark Knight Rises being like very disappointing for me cuz I watched scene one of The Dark Knight Rises and I immediately had in my mind this crazy story of Batman um you know, fighting the Joker, but the Joker isn't actually in this reality. He doesn't exist. He's more of a metaphysical concept of the Joker, kind of like they did with the Joker movie a few years afterwards, right? And I was so disappointed when I saw the Joker on screen. I mean, look, the guy did a good job acting, but it's just I was hoping that like the Joker's goons were the embodiment of the Joker. And and seeing like, you know, an actual physical representation of it really took me out of the moment and I was very disappointed. You see what I'm saying? So like um this is why I have trouble going through narratives because every time people set something up, I want to see more. And then unfortunately, I'm always disappointed in the end results. So when I read books, and I know this is stupid, I read chapter one and then I read the end and then I will actually read the middle bits that I find interesting. I do this with um TV shows as well. I'll watch the first bit, I'll watch the last bit, and then I watch like a Monte Carlo like kind of random assortment in the middle. That's what I do because otherwise I can't focus on it. It's not engaging to me, you know? So how does this go back to papers? Why is this related to papers? Well, when you read a paper, right? Here's my representation of a paper, right? You're not going to understand that paper. You're just not. No matter what, you're not going to be able to read this paper. Um it's going to be filled with information you do not understand. But here's the thing, every paper has a bibliography section attached to it, right? And every time a term comes up that you don't know, you can Google it. And this will bring not only one paper, but two papers and three papers and four papers. Um and then this will lead to more papers here, right? Here and here. And every time you read any of these papers, you're not going to understand it, right? You're not going to understand any of these. You're not going to understand this one. one, this one, this one, or this one. But then eventually, you get to a paper you can understand. It's way over here in the line. Right? And then you have a very satisfying thing that only happens when you're reading papers. You now understand Let me go ahead and remove these X's. You'll now understand this box over here. And you're like, "Ah! " Then you go back to this paper. "Oh, I understand that paper now. " And then you go here. You go down down. Check. And now you understand this paper. — And now you do the same down here. Check. Check. And then suddenly go back here and you understand the whole thing. And this process of like not understanding going through just not knowing, not understanding, being completely out of your depth, and then suddenly having something click

and then everything falling right into place is so satisfying, right? It Because you had no like it it's I don't know. I don't know how else to describe it. But the reason I brought up reading books and watching movies and all that is because this is exactly what I want out of a narrative. I want to be I want to read a story where like I you know, I can build my own path and kind of go through with my own interests. And then finally, once one path is has finished, I want to go back and I want to understand the other bits. Like that's why I enjoy reading papers because again, it's not me reading just one paper. It's me fundamentally understanding the whole wealth of research in the area and figuring out how the little bits and pieces go together. And once you understand this paper, well, all of a sudden these other papers that are also on the boundary of human knowledge that you were looking at that you didn't understand before, they're not immediately able to understand and you can't immediately understand them. But what you can do is you can already like understand 50% of them. Like now all of a sudden this paper over here that you were struggling to understand, well, half of it you already understand and now you only have this half that you have to figure out, right? And it just becomes more and more like eventually you start to understand more and more, right? And this is just this branching structure of building your own narrative to try to figure out what you know and what you don't know um is really enjoyable. It's even more enjoyable because you know that these are like every single paper is itself like a world first, you know? Like no one has ever done this before. That's why it's being published, right? And so because it's being published and because it's new, like there's a level of just excitement that you get when you're reading papers because even if it's on something stupid like I don't know, some synthesizing of some uh a genome that allows for better viral packaging. I don't know. I don't know what we're talking about here. Whatever it is, even if it sounds very boring, the fact is it was published. And someone spent months on this work doing something that has never been done before and that alone is interesting, right? And so each one of these check marks I put down is again something that's never been done before um at least at the time of publication. And so put that together with the fact that you're building your own narrative and you're able to craft your own adventure. It's really fun. I really enjoy it in a way that I can't enjoy anything else to that extent. Um let me go ahead and catch up in chat. Um Uh so when whenever people talk to me and they say, "How do you read a paper? " I always say don't read a paper. Like don't start with just this box. Don't try to don't hold this piece of paper and say, "I have to understand this in a week. " Look at it. Figure out what you don't know and go through all the un all known unknowns and just figure those out. And eventually come back to the paper because a paper in and of itself should not take more than 30 minutes to understand at least at a cursory level. If it's taking you longer than that, then it means that you don't have the prerequisites and then it's a matter of figuring out the prerequisites. Sometimes these paths are very long, by the way. Sometimes it takes, you know, weeks um to understand all everything you need to know before you can understand a paper. But it that's kind of the [snorts] process and that's why it's fun. Um Uh that's weird. It is weird. I guess I'm doing the low-fi version of that. Yeah, Wikipedia is a great start. Sounds to me more like my OCD uh not the must lock the door 20 times. I do actually have that, though. Every now and again I'll go to the um I'll go out of my door and like I'll fear that it is not locked. So what I do is I lock the door and then I will specifically try to open up the door and make like a pattern in my head so I can remember that pattern so that later if I have a fear that my door is unlocked, I'll be like, "Ah, but I remember the pattern, therefore it is locked. " If that makes sense. So you spoil all the shows uh at the end to yourself. Yeah, exactly because if I don't spoil them, I get very disappointed and then I can't watch the rest. Like it's either I watch the end at the start or I don't all. Like that's how it is. I've turned down so many books. I have um like stopped so many movies or so many yeah movies um and also TV shows halfway through just cuz I you know, they're great and I hear the ending is great, but like man, I can't sit through that. You know, I don't want to sit through I I you know, I'm so disappointed by the middle of the middle stuff. Uh we could be doom scrolling, more like doom branching. I think this is normal when you are researching something um not just reading paper to paper, but necessary parts and connecting the dots, right? Papers are nice for this because as long as a branch is consistent, it's valuable enough to be made rather than a single linear narrative. Um yep, which is classic inattentive ADHD. You would have issues with hopping around references literature because while you were doing it, there's no immediate dopamine reward. Wait, wait. Let me So you're saying that like I'm inattentive ADHD or I'm OCD? I I've got I don't know. I don't know the details here. Tell me more. I want to learn more about myself. The ability to poke at things I find interesting or I didn't understand is something that also makes traditional journalism difficult for me. Every piece of news prompts the question of what created the circumstance for that event. Hm, this is true and I think journalism and research in the way that I'm describing it are very similar. Uh I'm here rarely locking my door in the first place. Wait, what is that's actually efficient if you have ADHD, then it's ADHD plus something else. Wait, tell me more. What is the definition of ADHD? What's the definition of OCD? Yeah. Um Uh same but some Wait, wait. What are we talking about here? I found using a screen reader uh That's why they say show don't tell. A picture is worth a thousand words. And why translating books into a movie requires massive change. I'll give a fanfic writer brand. You read one plot, imagine the entire That is actually what I do. That is correct. I I'm 100% a fan fiction author, I guess.

The same but sometimes I actually enjoy the different direction they took. That's true, too. I actually agree. Sometimes the direction is better than what I came up with. Um like books like this actually are amazing to me. When they come up, I love these books. For example, and I know this is such a classical case and I hate that I'm like I feel like such a normie for saying this, but one of my favorite books of all time is The Sound and the Fury written by um Faulkner. William Faulkner wrote the best book, I think, of all time in Sound and the Fury. The way in which he moved like specifically Quentin's story Like the thing about The Sound and the Fury is it's so incre like in incredibly interesting to read just like word after word. It just it I don't know how to describe it but like the reason I can enjoy poetry but I don't enjoy traditional literature too much is because poetry like allows me to spend time and just be in the moment and enjoy reading the book, right? Reading the poem. Um whereas traditional literature requires me to think deeply about what paths we're taking and like where it's going and stuff like that. I can't just enjoy a book cuz I'm always wondering what where this plot line is going. What's going to happen here? All this kind of stuff. But with poetry, I can just enjoy reading it. And that was true for Sound and the Fury. Not only like the story wasn't great, but the reading of the book was fun. Um And any book where the reading of the book is fun I greatly like. Um me with Wikipedia. What is LHP BV? You said something down here in the YouTube chat that I don't understand. Anyway, that was my stupid ramble about how to read a paper. So, if you actually want the details of how to read a paper uh beyond that. Like so what I always say is a good paper is defined by its bibliography section. That is this bit down here. Like what has led to this paper being useful in the first place, right? Um and to me that's the most important part of a paper or one of the most important parts of a paper, right? Outside of that, the next most important part is probably the introduction way up here, right? So, developing drivable photorealistic human avatars like understanding this is also very important, right? Um and it should be basically read a paper, you go through the abstract, and if the abstract sounds good, you read the paper, right? The abstract is your AI summary. Like there's basically no reason to do an AI summary if the abstract is written well. Like that's just how it is. Um I I don't really understand why people do AI summaries of papers like single papers. I understand why you use it to search like a great amount of literature at a single time, but like to do a like a single summary of a paper I just I don't fully understand that but maybe that's just me being a boomer. Um — [clears throat] — but uh the details here are pretty straightforward. So, for me if I just want a cursory understanding of a paper and its context within the literature, which is what I do a lot, I'll read the abstract and introduction and then I'll pause. I'll just pause and I'll look at everything else at that point, right? Sometimes I'll go into related work as well, uh but I don't usually go as far as methods unless I feel like this is something that I need to implement for my own research, right? And then I go into methods. But usually for me the paper mostly ends at related work um and then uh and then I try to understand the other details more deeply if I need to use it, right? But for me I just want to know if this exists, this is something that people are doing, this is how it's it fits into the greater literature. Let's move on. Um and then if I need it, I'll look into the details and actually implement that. That's why it takes like, you know, maybe 30 minutes for me to get a good understanding what's happening. Um and then it uh um I can go from there. So, reading a specific paper, you go through abstract, you go through the introduction, you go through related work, you go through conclusion. Everything else you leave if you're interested enough. Um reading the whole research literature again, you do exactly what I did just a second ago. I went to related work and I said, "Okay, well, specifically I'm interested in these point-based methods. So, I'm going to go 35, 75, and 81 in the um uh in the bibliography and I'm going to read those papers. I'm going to understand where they come from and stuff like that. And so that's how I go through papers anyway. I go to the bibliography and I read paper after paper from there. Um Anyway, I was ranting. Uh I'm not going to lie I'll start to use your method. That's smart. Wait, what method is smart? So, you basically like consuming content that you know where the plot goes to. Yes, I think that's correct but Jesus, I have a drinking problem. Did anybody see that? That was very embarrassing. Oh my goodness. — Oh, I'm trying to act cool and I just spilled water all over myself. What an embarrassing person I am. Um So, uh in case you know more about inattentive ADHD um can you recommend a book for inattentive ADHD? Most books just target the hyperactive type and I still haven't found any good resource for it. I want to learn more about ADHD and OCD. Like what what is wrong with me? — So, you basically like consuming content that you know where the plot goes. I like consuming content where I can create my own plot, I think is the actual answer. Um So, like papers allow me to create my own plot because, you know, it's not just one paper, it's multiple like papers. Um video games sometimes allow me to do this but I don't really enjoy gaming too much cuz it it's just I don't know like basically every time I think about playing a video game, I think to myself, "Why am I not reading a paper? This is more fun. " And the same is true for reading books or anything else. Like every single time I try to get into something, I'm like, "Why am I not reading a paper right now? Or why am I not writing a paper? thinking about things to do for research? " Like just my brain it's so incredibly focused on research. I can't And it could just be because I have this heart condition and this heart condition is for forcing my brain to break. It it's one of those things. Um ask AI to I ask AI to ask me Socratic questions

from the paper up front. That gives me some idea of what we will get to answer as we progress through the paper. Many papers on IEEE are very dense and the abstract is not very clear. I find IEEE to be not so bad. Um I find for me and I know I'm going to be very annoyed or I'm not going to be the right person to talk about this. But when you're like I'm going through like iterative function systems all the time, right? This is a very deep mathematical underbelly that I am scared of. That's what it is. I'm scared of certain mathematics. That is how it is. — Like I will be reading a paper. I love the paper, it's great, the concept is great. Then all of a sudden they throw at me some very terse mathematical equation that I should understand but I don't. You know, there's just a lot of this that I really need a better deep understanding of that I don't have, you know? Um and it's because I'm like a systems engineer. I'm not that. I'm a like I'm a software guy. I'm not super great with the math. And so what ends up happening is that they'll throw like even mathematical symbols at me that I don't know. You know, it's like, "Man, I wish I knew that but I don't. " Um okay. Uh what I do find very useful let the uh better AI models compare a handful of related papers. I think it's up to you how you want to integrate AI into your workflow. For me like I don't want AI to do something that I want to do, right? And I want to read papers. Therefore, I don't want AI to do that for me, right? Like that's just kind of what it is. I want to write software so I don't want AI to do that for me. Like that's because part of my life is like finding a way to do things that I enjoy. I like writing code to some extent so I will try my best to use AI minimally in that. You know, I like reading papers so I'll try my best to make AI not read papers for me, you know? Like I you know — [clears throat] — I understand that maybe I could save like 30 minutes here or there if I were to use this tool to do that but like who cares about saving 30 minutes if I'm miserable afterwards, you know? Like keeping motivation up is your own way of ensuring that you're working productively, yeah? Um so, I think that's one thing that a lot of like middle managers fail to understand. Like when it comes to it, you can't just force people to use AI for specific workflows. You could but like it changes a workflow. Like people were programming because they enjoyed the ability to creatively express themselves in code, right? That's why a lot of people were programming. And so if you force people to use an AI tool that they weren't ordinarily going to use, maybe their productivity does go up in the short term but their like overall enjoyment of the job and it goes down. Therefore, productivity might will probably go down long term, right? Because people are people, not like not machines. Anyway, I'm talking too much. Um Same, I want to learn without stressing my brain. If there's no learning, uh it's like exercise you got to push to failure. Yeah, I also like pushing myself to failure in exercise so maybe this is kind of where I'm at. Um ML math and implementation sometimes gets so decoupled that it becomes too hard to infer the implementation from math. For machine learning the math is the clothing for the implementation and not the starting point. Um I think I understand what you're saying and I think I agree with it but I'm not sure if I fully understand. But I do agree with you that like the math behind machine learning is kind of not super useful uh for most people. Like it's really like it's quite simple usually. Um but uh we don't have good models for understanding like, you know, the deep nuances to it. It's mostly like I don't know. It you know, ML is always had the problem that it's like a giant black box, you know? I like programming because as a kid I was obsessed with Legos. I mean, that might be true. I was not obsessed with Legos as a kid. I actually found Legos to be relatively they weren't my thing. Um I have friends who are very obsessed with Legos but not for me. Um I was obsessed with writing. I love writing. That was my thing. I genuinely love writing. I have like 700,000 I probably have like at least a thousand pages of writing that I've never published before because why would you publish, you know? Okay, so we are now rambling and I've been rambling for too long. Um but I did find some valuable information here. The valuable information is that there are point-based approaches for uh doing the um uh what is this? A drivable avatar. So, my assumption is that when they say something is a drivable avatar, uh my assumption is that means similar to a model in some form of a mesh-based operation, right? That's what I assume is what they're talking about here. Um so, I want to read through their uh I think related work. And because I think they talk about it here, the point-based rendering. And I just want to see what they say. And then I want to go through it. So, related work uh ba ba ba Uh current methods for controllable avatars rely on dynamic neural radiance fields, point-based or hybrid representations, which are either slow to render or fail to correctly disentangle garments from the body, leading to poor generalization uh to new poses. Okay, so I didn't realize the garment thing was like a core thing to this paper. Like they care a lot about clothing, which is fine. You know, lots of people care a lot about clothing. Um I don't think it's super relevant to the stuff that I'm like I would be interested in, but yeah, it's fair enough. Recently incorporating 3D Gaussian splatting into dynamic scenarios has opened new research avenues. For a thorough overview, we refer readers to state-of-the-art reports on digital avatars and neural rendering. — Um 63 64, I'm going to take a look at this really quick. Uh 63 uh is going to be Where is it? What is the title? Advances in neural I have

that one up, I think. Don't I? No, I have a survey on 3D Uh it's fine. It doesn't matter. I'm sure they're all the same. Okay. So, I want to read these. So, dynamic neural radiance fields. Uh NeRF is a popular parent's model for human avatars representing scenes volumetrically with density and color information using an MLP. MLP being what? Machine learned potential? No, it can't be a machine learned potential. MLP is what? Why am I having so much trouble here? MLP. What is an MLP? Uh nope, that's I need using an MLP. They haven't defined what an MLP is. I So, this is something I should know, but I don't know. Um Yeah, I don't know. Uh so, machine learned potential is my guess. Uh machine learned field? Machine learned I don't know what it is. One second, I got to cough. Sorry. Okay. Um we'll keep going. I'm not super interested in this, so it's okay if I don't really know what an MLP is, but I should. Perceptron. Perceptron, of course. Um machine learned perceptron. That must be it. Um although I don't fully understand what a perceptron is either, so I need to learn about that. Images are rendered via ray casting and volumetric integration of sample points. Um sure, easy enough. Uh I see the difference here. Okay, I'm already learning something about NeRFs. See, that's not a splatting method then, because splatting is the opposite of that. If you're doing any type of ray casting, then that's Wait, is that what they're calling splatting? — Cuz that's not splatting by the fundamental sense. Multilayer perceptron. That makes a lot more sense. Thank you, guys. You guys are really good at this. I'm an idiot. Um Yeah, why Thanks, guys. I appreciate I'm just an idiot. I'm an idiot. That's what I am. That's what I've learned. It's okay. I can be stupid. Many methods have successfully applied NeRFs to dynamic scenes, achieving high-quality results. However, most methods treat avatars as single layers, which complicates modeling phenomena like sliding or loose garments. Yeah, I don't know how you would actually do that type of stuff within uh a NeRF uh or with like the traditional models that people are using for this. Um models like six and seven address this by using hybrid representation combining explicit geometric geometry from the um multilayer the S multilayer perceptron SMPL. Uh MLP. SMPL. Okay, whatever. Um I'll get through it. The implicit dynamic NeRF. Despite um impressive garment reconstruction, these methods struggle with novel pose prediction. Sure. Uh TeCHOS extends SCARF to a network of framework enabling prompt-based generation of NeRF NeRF-based accessories or hairstyles. I would like people when you when people write a related work section, I really like it when they specify like, "Hey, this is related and here is how ours contrasts it. " I know they kind of probably already did this in the introduction, but I do prefer it when people tell me why their method is better than this or why their method how their method is related to it, you know? Um Maybe that's just me from reading too many physics papers and not enough computational science papers. Uh so, before 3D Gaussian splatting, many methods use point-based rendering or sphere splatting. Yeah. With optimal positions and sizes. — NPC by Su et al. defines a point-based body model for avatar representation, but requires lengthy nearest neighbor searches during training. Yeah, that's fair. Uh it takes 30 minutes for training on their method. Whoa, that's something I didn't realize. That is actually something we could talk about. Um making impractical for dense and multi-view data sets. Uh Ma et al. represents uh garments as a pose-dependent function mapping. Why does it take so long to learn? Am I misunderstanding what they're doing here? I thought they were literally just like moving an arm. Why does it take 12 hours to learn? Under what situation could you sit for 12 hours to learn? Also, 30 minutes is insane, too. Am I stupid about this? 30 minutes sounds like an insane amount of time. When all you have to do is oversample the points. Am I feel like I'm misunderstanding something. I must be misunderstanding something, cuz otherwise I'm not fully There's something I'm that's not clicking with me. Um they should do that in the conclusion section, how it improves and contrasts with other methods. I agree. I think it's good to put in the related work section as well, because like I think it's good to motivate the reader to understand why they're doing what they're doing. I understand like giving motivation for this is where my paper or my idea fits into the greater research literature is the point of related work section. And if I don't have that, then it's um not exactly clear to me why I'm reading this, right? Uh let's use some collab experience around the paper. What exactly do you want to do? K-nearest neighbors are quadratic, I think. Multilayer perceptron. Uh at least that's what Jen and NatSci papers do. Yeah, it depends on your area. I mean, it also depends. Like some papers don't even have related work sections, they put it all into the introduction. And if I'm being very honest, I actually prefer that as opposed to um Like I prefer everything to be one big introduction as opposed to having introduction and then related work. That's my preferred method. But if you are to have a related work, I want to know exactly how it is related, right? Um like what you're doing that differentiates it. Yeah, 30 minutes, right? It sounds It's 30 minutes for training, then the assumption is that

you can reuse the model for other things. So, I I think there's something about that I'm not understanding why it takes so long. Um but 12 hours for learning is woof, that's tough. Um Uh making impractical Ma et al. So, like for example here, making impractical for dense and multi-view data sets. I actually like the fact that it's a 12 hours versus 30 minutes, cuz now I understand a lot of context I didn't have before. Um This improve Wait. Ma et al. represent garments as pose-dependent function mapping. Uh SMPL points to uh the clothing space. This is improved in 49 with a neural deformation field, but both models only address geometry, not appearance. Um uh Jiang et al. represents the upper part uh an avatar as a point cloud um grown during optimization and rasterized using a differentiable renderer. Uh while achieving photorealistic local results, the avatars suffer from artificial light holes. Yeah, see that's what I was afraid of with this type of thing. Any type of point-based method, you're going to miss up with these point holes. Um But, I need to now understand something. I'm missing something deep about this. The hugo42, thanks for following. What I'm missing is So, obviously we're using a bunch of we're mixing and matching a bunch of different rendering methods. I'm not too worried about that. What I'm missing is why it takes 30 minutes to do training. Why it takes 12 hours to do training when at the end of the day, like if you're just doing you're doing K-nearest neighbor searches in order to figure out what like points Basically, my understanding here is that they're they've got a point cloud that represents the person, right? They're going to throw these on a screen and apparently they they sphere splat it. So, it's not exactly Gaussian, it's like a sphere. Um cool. [clears throat] Not a big deal. No matter how you splat a screen, it doesn't matter. Um But, what does matter is why you know, if they have a point cloud, you know, and this point cloud represents the object, I don't understand why they're spending so much time training as opposed to just, you know, uh watching the possible movements and saying there's a joint here, you know? Like and if then we know we have to overpopulate this area, because we're doing a point-based rendering anyway. So, why not just overpopulate the arm? You know, this is what I don't understand and I need to learn more about. No state of computational science it could just uh be a be a sound but very efficient inefficient algorithm. It could also just be that again, it is inefficient, that this is like a first-time thing that people were intending to use later. Like but 12 hours is one of those things where like, you know, if it took 60 seconds to do something, I'd be like, "Okay, fine. I can optimize this down to 1 second. If I'm really good, I can optimize it to sub 1 second. " But saying 12 hours is like, "Man, I can optimize this down to at best probably 10 15 minutes, right? " Um and it means that the there is a significant learning phase no matter how you cut it, right? And that's what I'm like you know, that's fine, but it's something to keep in mind. What you say is image yes, search for it. I should search for it. Um Let's do some collab experience around the paper. Digging deeper into some paper might help learn how to verify and extend ideas. Reading is good, but having some explorations might be chef's kiss. I do agree, but we're doing that right now and writing our own paper. I thought about this a lot, like what type of meaningful content you can do as a researcher that you couldn't do as a content creator, right? Like for example, Ludwig is a fantastic content creator, but I can't just go to Ludwig and say, "Hey, write a paper with me. " I would love to be able to do that and I had considered Actually, he had a charity last year where if you paid enough, you could actually buy a bro versus bro from him and I considered trying to do the bro v bro so that I could, you know, no life it win at the games that we selected and then I could force him to do a research project with me, which I thought would be kind of funny. Um but I thought about it, you know, it's not going to work out, right? Like no matter what you do, it's not going to work out. Um but what I am saying is that like uh Xander brought up this idea of reading a paper and digging into it deeper and extending that, right? But the problem is that like — um you know, every single time I think about that, I always realize it's just better to collaborate. Just to just to, you know, work with your community. Work with another researcher who happens to be going live, you know? Like there are other people who have PhDs who do streaming. I should work with these people and I should collaborate with those people and we should do a research project together with our own mutual experience, right? Not just take a paper and extend it, but like, you know, do something like significantly meaningful that we both find valuable, right? So, I agree completely with what you're saying. I want to do research on stream, but the question is what how to do this meaningfully. I can't just like, you know, and I haven't quite solved that problem yet. Uh does the paper give data size or other hints as to why it took so long? Uh I don't know either. We're going to have to go through it to figure that out. Um you should look at Matt Parker's code. Someone did optimize the code 1 million times faster from minutes to it on Python to sub-seconds in C. So, there is a difference between a mathematician and a computational scientist. My assumption is that most people here are computational scientists uh or computer scientists, right? And because of that, my assumption is that their code is not going to be a million times slower than what it could be theoretically. My assumption is that there it's probably within a factor of 100 at of the final runtime, right? Um unless there's some algorithmic changes that you can make. Uh sometimes the right method can yield

massive optimization. Right. Let me That's exactly what I said about choosing the right algorithm. Uh but you asked me to look up Image GS. Let me take a look at this. Oh, wait. I'm on Firefox, aren't I? Image GS. Uh image -gs. Oh, come on. This is not what I want. Um Uh Image GS content-adaptive image uh representation via 2D Take a look at the archive and we PDF. Um content-adaptive image representation via 2D Gaussians. Sure. Yeah, this is I think what I was thinking about. Uh neural image representations have emerged as a promising approach Is that their whole abstract? No, no. The abstract ends here. Geez, I couldn't grok that for a second. Uh neural image representations have emerged as a promising approach for encoding blah blah. I want to see I just these images are weird. All right. I'll take a look at this later, but it does sound like this is kind of what I was expecting to do. And in fact, this might be a paper that I want to cite because that's what I was planning to do is just overpopulate um and allow for the mesh warping. Like, you know, I want to do something like rigging with my paper. Um but uh I don't want it to be super complicated. Like, I'm not doing a rig as in like, you know, you have like a specific — um like bone that you're moving, right? I'm using it as a mathematical field. So, it will move adjacent points, right? Um as well. And so, there's like limitations to what Basically, I'm seeing it as a field uh operator instead of as a skeleton that you're like connecting points to. So, there is no necessary necessity to like learn the skeleton or which points are connected to where or any of that kind of stuff. Um the difficulty comes in uh creating a field that hits the right points in the right spot, — um which is not particularly difficult if you can just move the points around as as you want and overpopulate, right? Um curvature might be the word I was looking for, yeah. Um okay. Let me just keep going here. I want to at least go through the uh related work section and see what they have here. Um because it sounds to me like this is actually solving a different problem than I thought they were solving. And they're really focused on clothing here, which is weird to me. I don't know why they're so focused on clothing, but I guess I guess, you know, they want the dream of being able to walk up to a mirror in Victoria's Secret and see and have that mirror like, you know, tell you what you look like in this lingerie or whatever. I guess that's what they want. Um I mean, everybody wants that. Everybody wants to, you know, see what they look like without having to put on the clothes cuz, you know, everybody puts on the clothes. That's gross. Anyway, um cage-based deformations. So, cages are commonly used for geometry modeling and animation, serving as sparse proxies to control all interior points, enabling efficient deformation by manipulating only cage nodes. — Um E von All et al. Oh my goodness. Introduced neuro cages for detailing pres- uh detail-preserving shape deformation, where a neural network rigs the source object uh to the target via a proxy. Uh Garbin et al. extends dynamic nerves with tetra- hedron catron tetrahedron cages to unposed ray samples based on tetrahedron intersections. So, this method is real-time, highly high-quality, and controllable, but limited to objects with local deformations like heads, not suitable for highly articulate objects like full-body avatars. But they're using a tetrahedron method. So, they must that the big thing that they're doing here must be to break through that. By the way, I don't know if I said this, but the Hugo42, thanks for following a few minutes ago. I think I missed that. Um uh Peng et al. used a cage to deform a radiance field in Cage-NeRF. Uh while their low-resolution cages can be applied to full-body avatars, they fail to model uh detail features like faces or complex deformations. Um time-conditioned methods. Playback methods represent a scene as a time-conditioned function that can allow for arbitrary control uh controlled Wait. In Wait, wait. Represent a scene as time-controlled functions that cannot be arbitrarily controlled. There we go. Allowing for a novel viewpoint synthesis while traversing time axis. Um it This isn't controllable, right? That's the issue. Yeah. Why are they even mentioning this? This isn't drivable at all. I'm missing something. Um Young et al. gang Young Yang extends the representation to 3DGS into 4DGS, which I assume is just time dynamics. Um effectively incorporating time into the primitive representation. Wu et al. combines Gaussians with 4D neural voxels inspired by HexPlane, which achieves real-time rendering and novel view synthesis. However, these solutions fell fall into different class of algorithms. Yeah, that's what I thought. I don't know why they're mentioning this. This is quite different than what they're doing. Like, they're focusing on the drivable. But maybe the reason they're putting this up is because some people wouldn't focus on the drivable. And so, anything that could do a 3D avatar Gaussian avatar is what they're focusing on. Could be true. That's fine. Um and they're putting that in related work. It's fine. It's fine for related work. Um uh dynamic Gaussian splatting. D3GA is based on 3D Gaussian splatting, um a recent alternative to NeRF for model modeling neural scenes. Uh due to its real-time capabilities and high-quality results, 3DGS has inspired numerous follow-up papers, of course, um — in areas such as physics simulation, hair modeling, head avatars, and fluid

dynamics. Several works recently uh introduced convolutional neural networks to regress uh Gaussian maps. Despite achieving high-quality results, fixed convolutional architectures do not allow for local conditioning or adjusting the number of Gaussians during training. Yeah. Uh similarly, I I don't imagine them being able to do the deformation that they want in this method. These methods also allow also use up to 23 times more parameters. Really? 23 times, causing the model size to reach almost 1 GB. In contrast, our pipeline remains lightweight and flexible, offering garment decomposition and localized conditioning. Finally, using CNNs — can slow down the pipeline to run 10 FPS, whereas our model remains real-time. But it takes 30 minutes to learn, which is fine. I guess it's fine. 30 minutes to learn is still fine. Uh what hardware are they using? Are they using just like, you know, a standard GPU or are they like using a uh Let me see. GPU. Where are we? Uh V100 GPU. So, they're using Well, V100 is relatively old now. I mean, V100 came out in 20 20, right? Am I wrong? 2019? 2020? So, I mean, this is a the 6-year-old GPU. Um it is a server GPU, so it's kind of faster than what you get on your desktop, but um not by much, I think, compared to like the A100s, which I think are the most recent generation. Uh you know, it's fine. The fact that they can get 30 minutes training means that with like the most recent uh GPUs, you could probably cut that down by a factor of two, maybe. I don't know. I don't know what the performance difference is. I don't think factor of two um is what you get. Probably 25% performance. Um and if you find a way to distribute it, you could even get that down more. So, if you use a um But what I was really wishing they would say Sorry. is that like they were doing this on Apple Silicon or something. You know, so like a desktop computer that you have physically available on your system. Right. That's what I was hoping, you know. Um but yeah, obviously not, you know. I guess I was being too naive. Um you know, or if they were using like a AMD Radeon 6700 XT or something like that like I've got. Just something that's like oh, this is somewhat old um consumer hardware. — Um the these are always my favorite papers cuz it means that like the algorithm's actually useful um as opposed to stuff that's running on like V and A100s, you know. Okay. So, I usually learn a lot here. Um and I do want to do some follow-ups here with the point-based methods cuz I think that's the closest to the method that I'm using and I think it's the most relevant to what I'm doing here. Um this paper and also this paper, I think both of these are worth citing if I start talking about meshes. Should I Can I How do I combine these into a single tab? What How do you do that? This? It would Come on, please. Come on. There we go. I don't know what I'm doing. So, these are good to look at later. Um and the point-based methods. So, uh where are we? Point-based rendering, sphere splatting. Uh I don't know if that's actually relevant, either. But um let me go back to the citations that they had in the related work. — So, it was 35, 75, and 81. 35 is on the power of points for modeling human clothing. Um let's see if they just have more general-purpose uh visualization here. 75 uh is going to be Point-NeRF point-based neural radiance fields. I think this is the one I want to look at. I think 81 was the other one. No, 85 Uh no, I don't remember what the other one was. But let me take a look at this as well. Uh so, if we go back over here to this browser, where I can easily get to DuckDuckGo, scholar, and paste this in here. And let's grab this paper. Um 936. I figured this would be a the CVAS. com. Uh yes, I think this is exactly what I'm looking for. Perfect. I'm going to take a look at this as well. Uh Point-NeRF point-based neural radiance fields is exactly what I'm looking for. Perfect. Okay. So, um this I need for something else, which I'm going to put over there. And so, we're going to 3D Gaussian splatting. So, I think I'm getting to where I want to go. What is this figure? Can I just be honest? Like, we're running We have a trend in research right now where for some reason I it's fine. It's just people are being like quite artistic. Like, you know, like the crayon background here is just a little bit I I'm picking at straws. I shouldn't care about this. It's fine. It's just every now and again you'll see stuff and it's like, "Why did they use that font? Or why did they do that? " It makes the paper kind of seem less trustworthy, you know? But maybe I'm just being biased. Okay. Uh Nerf, I just want to read the abstract and go through this cuz I don't understand neural radiance fields. I don't I should, but I don't. Um I really don't know. Uh so this is something I have to look into. Uh this I'm not interested in, but the 3D Gaussian splatting I should be reading through this one in more detail. So let me also see if I can I combine these into a Yes, I can combine these into a tab and I can move the tab over here and then I want this to stay there. Perfect. All right, let me take a look at this. PowerPoint template energy. Yeah, exactly. Exactly what I'm saying. Like I it's fine. I don't mind it too much. You can do whatever the heck you want. It's fine. But it is a bit

weird that we're seeing all these like really pretty diagrams and stuff like that. Um I just looked at the viewership. I don't normally look at the viewership, but we got 51 people here on Twitch. That's a pretty big number. And then we have seven people on YouTube. We have like 10 times the viewership on Twitch as we do on YouTube right now. Not a problem. It's just interesting. Oh, Google research, yeah? Um the thing about computer science that's different than like physics or um probably chemistry. Biology might be the same as computer science is that like when I see a university in uh a paper, like if I see like Waterloo or if I see Berkeley or San Diego like in these cases, I'm like, "Yeah, these are good universities. " But I don't like my mind doesn't immediately say that like you know, I know people from that university who are doing good research, right? But if I see a company like Google or um uh Nvidia or something like this, I'm like, "Oh, I probably some good research goes into this. " You know what I mean? Like it somehow I have a bias towards companies in computer science, whereas I have bias against companies in every other field. You know what I mean? Um I I it's just how it is. I don't know why. Nerf is MLP trained to uh spit out coordinate and color. I need to learn because I don't know the problem is I don't know what it again, an MLP is. Right? I think that's my problem. Uh I thought it was machine learned potential before. Um it is multi-layer perceptron. And I don't know what a perceptron is. So I just need to go through and like learn. All right. Uh I no clue why you got recommended, but I'll take the adventure. Yeah, thanks for being here. I appreciate it. Um Libertario Lonco, thank you so much for following. Really appreciate it. — Okay, we present a method that achieves state-of-the-art results for synthesizing novel views of complex scenes by optimizing an underlying continuous volumetric scene function using a sparse set of input views. Sparse set? Is this sparse? You're telling me that's what you consider to be sparse? I I don't know if I'd buy that necessarily being sparse. I mean, I guess it's not dense. But that's a lot of views, isn't it? Right? There's a lot of views. I I don't know. Am I dumb? Yeah, I mean, the answer is yes. I shouldn't even ask that question anymore. Okay. Uh our algorithm represents a scene fully I I'm sorry. using a fully connected non-convolutional deep network uh whose input is a single continuous 5D coordinate um spatial location uh and viewing direction. Sure, that makes sense. You need XYZ and then kind of the viewing direction camera angle, right? That's fine. And whose output is the volume density and view-dependent uh in view-dependent emitted radiance at the spatial location. So I think the thing I don't understand is emitted radiance. Um I I'm guessing that's just like the amount of light that you have on the scene at that particular time, but I need to I need to figure out what we're talking about here. And this is how deep you know, I don't understand Nerfs at all, right? Splatting is literally you take a ball you throw it at screen. It's not hard. But like, — you know, whenever we talk about Gaussian splatting, they talk about in context to the 3D Gaussian splatting paper that came out recently. Um which again, you could always do 3D Gaussian splatting. Um it wasn't like this was difficult to do before this. It's just now you can do it in like real time and render scenes directly, right? Um so that was the big thing that this paper put out, right? Um the real-time radiance field rendering. Like that's the thing. The 3D Gaussian splatting, which is what everyone talks about, like we could always do 3D Gaussian splatting. That's not difficult. The difficult thing was combining it with real-time radiance field rendering was the thing that made it so this is a new thing, right? Um these splatting is like one of the oldest rendering techniques in the book, you know? Like it's before meshes, there was splatting. It's just throw a ball at a camera and you know, look at the results. That's it. Um anyway, [clears throat] uh that's my rant of many today. We synthesize the views by querying 5D coordinates uh along camera rays and use classic volume rendering techniques to project the alpha colors uh and densities onto an image. They use volumetric rendering. They use volume rendering techniques. So that So for the record, volume rendering techniques have always been a little bit bad. I don't have any other way to put it. They're not great, but I do understand why they're doing this because it makes the most sense um that they have basically they've created a volume and they have to find some way to um put that volume onto your camera. Um so yeah, I totally get why they're doing it this way. Um In fact, now that I'm reading that they're doing it this way, I have a much better idea of what they're doing. And I now have a much better understanding of the 3D Gaussian splatting paper. My idea right now, and I again, I might be wrong, they take their cameras um and again, their cameras have an XYZ position and then a an angle uh by which they're shooting the rays in, right? They got a bunch of rays and then they just got to learn based on this ray going in here, cameras, uh they got to learn what the scene looks like, right? Um so that's step number one. So it's going to be some sort of field. Um and then from that field, right? Which is now going to be some sort of 3D uh like kind of density plot of what it looks like on the inside, they have to put that on screen. So how are they going to do that? Well, they use traditional volume rendering approaches in this paper. But the 3D Gaussian splatting paper um turns this into a point cloud um again by the method that I think stripe one

was saying before, uh but I I don't exactly know exactly what they're doing to do that, but they have some volume and then what they're doing is they're turning this into something that could then be Gaussian splatted on the screen, right? So the thing here is saying, "Okay, we can represent a scene via neural radiance field, which is going to be, you know, some sort of 3D density. " And this is saying, "And then what we can do with that 3D density is turn this into essentially point cloud that's splatted screen. " And how do we do that, right? Um is what I'm understanding is happening right now. Uh because volume rendering is naturally differentiable, the only input required it to optimize our representation is a set of images with known camera poses. Uh we describe how to effectively optimize neural radiance fields to render photorealistic novel views of scenes uh with complicated geometry and appearance um and demonstrate results that outperform prior work on neural rendering and view synthesis. View synthesis can be uh results are best viewed as videos, so we urge readers to view our supplementary video for convincing comparisons. Anytime your abstract is like, "Oh, please look at the videos associated with the paper. " I kind of at least for me, it kind of makes me not think the paper is serious in a way. Um I mean, I know it is serious. It's a good paper, but like it kind of makes me feel a little bit, you know, a little weird about it. Um So what do we have here? Related work, ba-bum. So what I want to learn about is perceptrons. That's about. So let's go here. What is a perceptron? Uh step Okay. So multi-layer perceptron. Um What What is this? This is what I don't understand as deeply. Like a biology paper without with no confidence intervals. Wait, what? Did I Was I Yes, exactly. If you see a biology paper without a P value, I agree with you there. 100% Like you see it without um uh like any type of uh metric to show like what the variance in the measurement is. Yeah, 100% agreed. I think that's the same thing I had a problem with like these papers. When you see people doing like stuff that's like just a little bit off the beaten path, like you're kind of like, "Is this a serious paper or am I reading a scam? " You're very rarely reading a scam. And like I said before, papers are not about the individual paper that you're reading. It's about the set of papers in context, right? And so even if you are reading a scam paper, like what matters is the context around that scam paper. Um and like why are they interested in this to begin with? So even if you're reading a bad paper, it's fine because like you understand more about the field as a whole in that somebody was trying to um uh like do this thing to begin with, right? Parker Sham, wow. Hey, thanks for following. Really appreciate it. Um it does feel a bit weird, but then again, I often wish physical papers had hyperlinks or animations available as content. I agree. I mean, like I'm making Twitch streaming content right now, right? So Y is equal to ReLU(x) + bias is one MLP paper. One MLP layer. Okay, so I don't know what ReLU is. Uh so that that's how far gone I am. That's how little I know about this and how much I need to learn. Um these should be tied together, but I don't think they are. All right. Um let me just figure this out. Um for resulting in perceptually distracting Oh, okay. Uh they only mention perceptron once. Um so let me just Google MLP. Multi uh layer perceptron. Multi-layer perceptron. In deep learning, multi-layer perceptron is a kind of modern feedforward neural network uh consisting of fully connected neurons with non-linear activation functions. Okay. Um organizing layers. I'm so sorry. Uh notable for being able to distinguish data as linearly separable separable. Is that what this is? Or is it Wait. Notable for being able to distinguish data as not linearly separable. Okay. So I need some more data here. Non-linear function ReLU is x if x is greater than zero, else zero. So literally Huh. This seems straightforward or am I misunderstanding something? Um like I don't have to understand how they work. I just have to understand what they do, you know? It's okay. Um This So I don't know anything about deep learning. Like I've implemented my own learn like not neural networks before, but they've always been like, you know, I it's not like not deep, you know? I haven't done deep networks. Um I love early AI research. You can feel the optimism in the future of AI until it's crushed by the reality of the past hardware. Yeah, I mean to be fair like again, these AI papers have been coming out for decades, you know, and that's just how they are. But but Okay. But nothing especially deep about deep networks. Right. But let me I just I want to learn more about like Let me go to Wikipedia because I'm sure on Wikipedia they give me a citation that give me more information. Um So what are we doing here? Mathematical foundations here. Here we go. Activation function. If a multi-layer perceptron has a linear activation function in all neurons, that is a linear function that maps the weighted inputs to the output of each neuron. Sure, so that's just input output and then we have some weight in between them. That's So like Journal new page before. This means that we have a box. Here's a box. And then we have input you know, A B

C I should D I'm committed now, but I feel like I shouldn't be committed cuz I should just do this on my tablet. G H I J All right, fine. There there's our inputs and outputs. Then in the middle here like somewhere here, you just have some weight that represents the influence of A on G and H I and the influence of A on J, right? And then you have weights for all these guys as well. So basically that that's where I'm This is all I'm thinking. I don't think it's complicated. Then linear algebra shows that any number of layers can be reduced to a two-layer input output model. I'll believe that. I'll take that. In MLPs, some neurons use a non-linear activation function that was developed to model the frequency of action potentials or firing in biological neurons. Fine. The two historically common activation functions are sigmoids, which are described by this. Sure. The first hyperbolic tangent. Why are we so closely modeling neurons here? Why do we care? Like you know, we don't have to follow the biological Okay. It's fine. Um But is that what the perceptron is? No, this is this can't be what the perceptron is. A feedforward neural network consisting of fully connected neurons with non-linear activation. Is that it? Is that all it is? Organized in layers not really being able to sing. That's it. It's just it's literally just a neural network. That's it. It's just a network with weighted with weights — with non-linear activation functions that allow for them to more easily differentiate between like This is easy then. It's literally just a neural network that just has the activation functions the um The activation functions being non-linear, I think is the trick here. Um So I think this is me not understanding the like feedforward neural networks as deeply as I think I should. Um So modern neural networks are trained via backpropagation and are colloquially referred to as vanilla networks. MLPs grew out of an effort to improve on single-layer perceptrons, which could only be applied to linearly separable data. A perceptron traditionally used a Heaviside step function as a non-linear activation function. However, the backpropagation algorithm requires that modern MLPs use continuous activation functions such as sigmoid or ReLU. Okay, that's why you said ReLU. Okay, sure. If you would have said sigmoid, I would have been like sure sure, I get you. Multi-layer perceptrons form the basis of deep learning and are applicable across the asset of diverse domain. Works for nature, so the approach might just feel like promising starting point. Yeah, I do agree with you there. I mean, we you know, one of the things that I did back in always like one of my first research projects was creating a biologically inspired analog electronic neural network, which was fun, but I've forgotten everything about it. Like genuinely everything. So when I say I've implemented neural network before, I'm like yeah, of course. I mean, who hasn't? But also like mine was an analog electronic thing that was probably not like it's probably not what people think of when they think about neural networks. So I have a lot to learn here. You spend an afternoon going over the ML and DL terms, you'll get it. I think a lot of the ambiguity comes from that area inventing odd terms for common terms, features versus variables. No, I get it now. As long as a perceptron is like a form of neural network, I'm fine. I get it. I don't have to worry about it too much. Um And then a multi-layer perceptron allows for Like I don't have to implement a nerf. to. I just have to kind of vaguely know what it does because I'm not using it for my work right now. I just have to know what it is. And then I can go on from there. And I do realize I kind of clickbaited everyone. Maybe that's why people are watching right now because I put nerfs in the title. Is that what happened? And I'm not even going into nerfs in that much detail. But anyway, we'll see. I'm really going back into my AI intro curriculum. I So the thing is there was no intro AI course when I was going through undergraduate and there was no AI anything when I was growing up. So I just kind of was like reading you know, I was reading papers on this and I've forgotten all the papers cuz it happened a long time ago. Is it normal dense NL NN stuff? It's typically used to create a sort of smart function. That's fine. I get it now. Like I totally get it now, but I didn't realize that's what a perceptron was cuz if they would have said a like what you just said, dense neural network layer with non-linear activation functions, I would have been like oh sure, easy peasy, no problem. But like they called it a perceptron, which is like okay, well that sounds like something different, you know? And that I just didn't know. That's me not knowing and I'm learning something new, you know? And like Stipe one said, they have odd terms for common terms, right? That's just you know, that's how it is in every field and I just didn't know. This is me just kind of avoiding this for the past 10 years and coming back to it after not caring about I still not caring that much about the AI tech. — But like you know, I recognize that it is useful for some things and I have to use it for some things. Yeah. And also it's different than the stuff that we use for science, right? Like machine learning potentials are fundamentally different than this, you know, and they work differently. Yeah. So it's Also it's not common for people who do ML to know deeply how ML works. I think is a cool but outdated term. Maybe. I didn't know about it. I this paper came out somewhat recently, 2020. I mean like I

haven't read an AI paper in the past 5 years, you know? Like not since I finished my PhD. So that's been 6 years. So I I understand that I'm like I kind of actively have been avoiding AI papers cuz I don't find them particularly compelling. So like when I see this kind of stuff, I just leave. That's been my go-to. I know that makes me sound really lame and that I should be like — like learning this like in more detail, but like but probably perceptron's old and probably nerfs are new. I just always looked at the term and I'm like okay, this is some neural network thing and I just didn't really look into it too deeply. But like when you say you say like when I guess it was some neural network thing. It ended up being exactly a neural network thing. There was like nothing crazy about it, you know? So yeah. But we don't Yeah. This is me like I never did deep learning network stuff. I we never did like — basically anything beyond like a simple network we never had to do during the PhD, you know? We only did like deep learning. What we Deep learning was just then coming onto the scene when I was leaving my PhD. And since then I've been done basically no AI outside of like what we need for some scientific simulations. So that's where I'm That's why I've missed this and I know it's a blind spot in my understanding, especially when I'm doing a graphics paper that is using a splatting method. I should at least understand this to a some extent. Yeah. Um Anyway, where were we? We were just looking at multi-layer perceptron. When did this come This came out 1989. So this is quite old. Approximation by superpositions of a sigmoidal function. Yeah. So you're right in saying that this is very old and I should have known it before, but I didn't. I forgot. There's a whole bunch of things that I've forgotten. I'm old, you know? I'm quite old at this point. Which is a sad unfortunate reality where I'm in of what I of the state that I'm in. Okay, so survey on I think this is enough for me to go from here. I don't think I have to go much deeper into this. So What do I want to do? What I want to do is I want to I've got three things I think I want to do, right? I got do. So let me go over here and let's write these things out. Journal new page before. So number one, right? Number one is I want to do some form of rigging is what I'm calling it, but it's not it's not going to be exactly rigging. Just moving fingers around and stuff like that. And then two, what am I going to do? I want to do I really still want to do this mirror frame on an SDF, right? The reason being that I just I know there were people in chat who said that mirror frames on SDFs are quite easy and trivial and maybe you're right. And I am also showing elsewhere in the paper even here in the in figure one, which is really the only figure I have right now, that you can very easily do these types of texture warping without too much trouble, right? So I — I don't necessarily need to show this, but the fact is that SDFs are at least in my opinion from what I can see limited in their ability to do like certain complex manipulations because ultimately you're combining Euclidean shapes together or shapes together in Euclidean space, — which makes it hard to draw some objects in it, right? Like just going to Inigo Quilez. Where are we? Inigo Quilez. We go to his website and he just has an intro to SDF somewhere in here. 3D SDFs. I guess probably this is the one I'm looking for. If you go down here to the very bottom, he explains the problem in more detail. Where are we? Come on. Keep going. Keep going. It's really far down, isn't it? Here we are. So deformations and distortions right? So he's basically this is what I'm going off of, right? Deformations and distortions allow allow people to enhance the shape of primitives or even use different primitives together. The operations usually distort the distance field and make it non-Euclidean anymore. So they must be used carefully. So one must be careful when ray marching them. You'll probably need to decrease your step size. Um if you're using a ray marcher to sample this, right? So, what Basically, the point here is that these additional deformations are relatively annoying to do. Um and uh I They're not inherently difficult to do if you were to put them into like my space, you know, uh the space that I have for the method that I'm looking at and then just do the transformations on top of that, you know. So, what I'm trying to say is I still want to show that you're able to do this type of stuff relatively that easily with an SDF. I think it's I just because it it's not like it's impossible to do with an SDF. You obviously they're showing you how to do it. Um but I think you have to change the method so it's a time-based ray marcher instead of a spatial-based one, I think is my understanding of that. But someone who knows more about SDFs will tell me more. Like this is what I don't know. Um but my understanding was you have to uh do like very small time steps, kind of like what we did with the invisible uh simulation like 10 years ago now. Um so, however long ago. Um so, that's my understanding. So, I still think it's useful to do a some sort of smear frame on an SDF. And then three, the other thing I want to do What was that last thing? Um Oh, there was something big that I

wanted to do in the paper. What was that last thing? Let me see if I can find it. Um do do. It was Oh, right right, just lip sync. This is it. Lip sync. It's not as particularly complicated, but the reason I want to do is just to show you can do a type of meta programming where you basically take the CSV, you compile that into my set of Quibble software that I have and then that compiles into a usable kernel for what you want. Um and so, just showing that compilation workflow for something that's not It's trivial, but like, you know, it it's you know, historically it is useful to do. And it's also useful for the animation I want to make um is worth doing. And then finally, I want to go through this space example here, uh clean it up a bit. Where are we? Here we are. The space example down here at the bottom. Uh I think I will use the space image, but uh I'm going to clean it up um so that it's not it just so it's better. Uh and figure out how to actually work this uh figure so it's actually nice. And the other thing is I want to count down the specific number of bytes that I'm using here cuz I think the for a thread is like 10. Um but I'd like to keep it under 10 if I can. Um that's my goal to keep it under 10 bytes per thread. Um that'd be the best case scenario. But I don't know if that's possible. I think because I'm failing something. — Um anyway, see you folks later. Thanks so you guys so much for coming in, by the way. I really appreciate it. Okay, so I think these are the big things. So, the last thing is count number of bytes. So, the goal is to do this kind of like throughout the next week uh because today's Friday. So, I think I'll slot this one in for Monday. Um we'll do lip sync. This is going to take a couple days, so three days is probably going to take for that. Um this one, honestly, probably just one day, uh but sometime next week. And then some form of rigging, again, this is probably going to take another three days. And then I think that's it. I you know, I think we could do it all in that time if I'm really focusing on it. Um so, this gives us seven, eight, possibly two weeks, and that's exactly what I'm looking for cuz I was hoping to have everything ready by basically May 1st. My goal is to have the paper done by May 1st so I can go ahead and put that on the archive. Um that's my kind of goal for the future in what we're doing. Um I think, guys, that we're even streaming for 2 hours and 15 minutes, and I've been quite tired today. Um I might just chat for a bit. If you guys have any comments, questions, concerns, ideas, anything like that, feel free to let me know. Um I've got a couple things I just want to talk about, but um I I'm going to chat, and then we'll um uh we'll kind of call it for the day. So, if you guys have somebody that you want to raid as well, feel free to let me know if you're on Twitch. Um Uh do you uh You're asking have I used a Gantt chart? The answer is yes, I have used a Gantt chart, but uh I don't use them very effectively um because I kind of the So, for people who don't know what a Gantt chart is, it's basically like you're doing this activity for this time, this activity for that time. Um and like you have these different ideas that you're working on simultaneously. It's good for that. But um I don't I usually I just kind of keep it in my head. It makes it more dynamic. I feel more like uh I don't know. It just Like I was the type of person that never really had to keep uh like write stuff down like when I'm doing what. I don't really need a calendar cuz I just kind of usually I'm able to keep it all in my head. Um and that's what I do. That's not always true, but um it's been true more recently than not. Um Yeah, I don't have so much more to say. So, if you guys have anything, feel free to let me know, and then I'll go ahead and close it. Uh mind be loathed. Definitely mind be loathed. Um but I will say I have a couple like kind of bookkeeping things to talk about with Twitch and YouTube. So, first thing with Twitch is um yes, I actively know I've said this before. I actively know that I am losing subscribers on YouTube. That's what we're doing. We're actively losing subscribers on YouTube, and it's fine. Um I I we just have to be honest about ourselves. We're going to be losing subscribers on YouTube for the remainder of this month, and hopefully we'll start getting subscribers back in May. Like that's the goal. Um I understand that we're doing different style of content that people are normally used to on YouTube. People don't usually watch live streams, and I've been putting out shorts, which is weird for me. But my hope is that near the end of May, I can get the first video out. And the video I really do genuinely think is going to be a banger. Um it's actually a an animated music video that is related to the algorithm that I'm putting out. It's got a really nice story. Uh we'll be doing the animations on stream as well as the other research projects. So, my goal is to kind of finish out this paper by May 1st, and then kind of slot in these other projects that people have been talking to me about um in the next month, right? Um one like see what we can do about that. Uh you gained me as a subscriber. Oh, hey, well, thanks for being here. I appreciate that. I see you've got the Clippy um uh the Clippy profile picture, which means you're a Louis Rossmann fan, which I'm also a fan of for the most part. Um but yeah, anyway, long story short, that's kind of the goal. I I'm going to be getting some stuff done this month, and then next month we'll start working on other people's stuff. Are there recommendable starting points for someone without an academic background to dig into the topic? Uh which topic do you want to dig into? Um I would argue that the paper that I'm specifically working on right now, um the stuff we went on stream, probably it's not a bad start. I mean, but when you're reading the research papers, like we talked about this before. Um right. So, here's your list of the papers. You start with paper one, then you go to paper two, paper three, paper four, paper five, all this kind of stuff. Um there are eventually papers

that you hit that are review articles. And these review articles are papers that give you a really deep understanding of the literature and where it's at right now, and they have a huge bibliography to jump off of. Uh we hit a review article today. Uh the review article we hit was here, um a survey on um 3D Gaussian Splatting. And it's good. It's good to hit these review articles and to kind of go through and see where the literature's at and what people are doing. One nice thing about computer graphics as opposed to like, you know, heavy mathematics or physics is that there really isn't that much mathematics involved, so you can usually get through it without um understanding it too deeply, understanding the mathematics too deeply. Um and so, really just if you just Google um the stuff that we're doing today was mainly just Splatting stuff. So, if you just Google Gaussian Splatting, you'll probably find what you're looking for. Just start there at like the Wikipedia article and kind of go through the 3D Gaussian Splatting paper, similar to what we were doing today. Um it would probably be the best starting point. Uh the other thing I want to say is that uh yes, what was I going to say? Where am I? I want to I don't know what I want to keep on screen while I'm talking. Maybe I should show my face. Maybe Yeah, maybe I'll we'll put my face forward. Uh here, there. Hello. I'm big now. Um the other thing I want to say is that um uh you know, I'm trying to encourage people to like do research out in the open. So, if you're watching this and you happen to be a researcher, and you think it's kind of maybe a good idea to stream that research, like just kind of talk to me about it. I'm I'm willing to help. Like I want to help. I want to grow the scene of more people doing research on stream. I've been thinking about ways to like engage uh a larger audience with like doing research and stuff like that. And every time I think about like gimmicks that I could do, you know, like, you know, Ludwig's vote bro v bro or like, you know, playing games with other people or the equivalent in the research space, I always just think of things like, you know, getting together and like doing a research proposal together. And whichever proposal is better, that's the one we go with or something like that. Like actually getting a meaningful like research results out, I think is the most important thing. Um and you know, my hope is with, you know, subsequent videos, like basically every video comes with it a paper or at least something interesting uh for people to look at so that we're doing more than just like making videos, but we're also like doing interesting stuff that has not been done before, right? Cuz I think there's novelty in just saying like, "Hey, this hasn't been done before. " Like you can every one of your videos can be like, "Hey, we did something that's never been done before. " Like that's the cool thing about research that you can't say in any other video. You can start the video out by saying, "This has never been done before. " And I think there's a lot of novelty there um that has not been properly explored. Um and so, if you're a researcher and you're interested in uh like working with me, um or you know, whether you want to be a content creator or not, if you're just interested in working with me, um just reach out. I'm more than willing to work with you. I'm probably Um Paper exchanges could be fun even between disciplines. Uh I see the disparity in language understanding. Yeah, so that was one thing about OIST, the Okinawa Institute of Science and Technology where I came from. It was truly interdisciplinary, multi-disciplinary, whatever you want to call it. So, like I would very regularly watch uh or listen to lectures from people talking about crocodiles and stuff like that. It was fun. Like I truly do believe that your ability to communicate across disciplines is um it it's a very valuable skill, and I want to see more of that. Like I would love to do some biology research. Like I I have some interesting biology research I'd like to do. Um by the way, Intergalactic is super. Uh you you surper. How do you say that word? Anyway, thank you so much for following. Really appreciate it. Um but yeah, the I have some biology papers that I kind of want to do with um uh viral creation. There's like a lot of people who are like synthesizing viruses based on like genomic data, and I find that really cool, and I thought it'd be cool to do on stream um and like to do with people. So, if you're biologist, like, you know, I know some bioinformatics. I've taught a course on bioinformatics, technically, although I only taught the programming parts. So, like, you know, I'm there. I'm game, you know. Uh though I'm just beginner, I but will be following you through the journey. Much love and wish you good health. Yeah, again, my health is kind of a big problem right now. Um I don't really have too much more to say, I don't think. I think we're pretty good to go. Um so, I think that's it for today. Um we actually did get a little bit done. I think the only other thing I want to say is that I'm a little bit I need to move faster. I've been moving quite slow. Part of that's because of the health, but part of that's just because of me. And so, I'm thinking about ways in order to move faster. Um, I'm not sure what I could do. I could, for example, like I thought about a stupid idea, which was like um I could like do a stream that's like um uh I'm a VTuber until I do something meaningful or something like that. And so, I just I I'm like an anime uh VTuber character until like something meaningful happens, like I make meaningful progress on the paper, we do a meaningful simulation. And like at the end of the stream I ask like is this meaningful or not? And if you guys say it's is meaningful, then I don't have to be a VTuber anymore. But if you say it's not meaningful, then I have to keep VTubing. Something like this, you know? Um As a key reason I have my current job, capacity to explain computer topics at various levels of comprehension. Yeah, I agree. The ability to talk to people as if they're like do the ability to communicate what you're doing to uh another technical professional is very important and very valuable. Okay. Anyway, I've got ideas. If you guys have any comments, just let me know and I will address them as soon as I see them. Um my suggestion right off, do a novel project 1 month, present your results. I do have backlog of small projects uh that I want to do. Yeah, this is kind of my goal as well. You can use FujiTech uh

for an easy PNGTuber for that idea. I can I have a whole bunch of different VTubing. I I actually do VTubing. Uh I have done VTubing on the side. I find it really fun. Um but I would not um I don't know. You guys don't understand. That sounds like encouraging chat to not let you get stuff done. Maybe that's what it's doing, but it also would force me to get work done, um I think, because I think I really do genuinely think that if I were to do that, chat would be very upset with me in like 30 minutes. You don't understand. I I can make the stream very annoying. I can definitely Um I but it would be very funny, I think, for people who who've seen the the channel before. Uh we could just make it so every stream you add on an additional like if I don't do anything meaningful and the answer is no, then the next time we have to add on like an additional constraint or something until eventually the stream becomes basically unwatchable. You know, like uh first stream is like I have to VTuber until I do something meaningful. The next stream is like I have to VTuber and like a pixel on my screen will turn dark every now and again. Um or else I have to You know, like we can do like different things like that. There's a way to make it work. Anyway, I'm going to close for you. Thank you guys so much for coming in. Have a wonderful day, night, afternoon, morning, evening, whatever time it happens to be your time. I'll see you guys around next time for the next episode of Leo's Labs. Hopefully it'll be Monday. We'll see how my health is, but I think Monday is good. Um so, I'll see you guys then when we will hopefully um go through that space example that I showed before and regenerate that figure, um rework that section of the paper, figure out exactly how many bytes we're allocating. Um And I'm hoping it's under 10. If it's not under 10, I might rework it so it stays under 10. Um and that's the goal. Yep. Uh fail failure to meet goal equals half resolution. Yeah, maybe something like that. Anyway, thank you guys so much. I really do appreciate like all the chats, all the comments, all that kind of stuff. I'll see you guys next time. Bye, guys. Uh where am I? Twitch.