🔴LIVE | What am I doing?

— Am I live on YouTube? Hello. Hey, hello. I should be live, I think. Hello wonderful people. Hey, hey, hello. I just never know. I never know. Hello wonderful people of the internet. Welcome back to another episode of Leo's Lab. So, today we're going to catch up on some stuff that I was working on yesterday yesterday. So, basically, while I was streaming yesterday, there was like you know, a really good amount of conversation happening about the research project that I'm doing, all this kind of stuff. And I just kind of lost steam halfway through. I know it was only a 2-hour stream, but you know, I my heart was getting all palpitati. Um not quite that bad, but you know, like I was starting to feel like I was going to kind of crash a little bit, so I left. And I did feel a bit bad in the evening and in the morning. So, today is going to be kind of just a chill, you know, we'll do what we do type of stream. I'm not it might just be a short stream. This might be one that's unlisted on YouTube afterwards. I'm not really sure what's going to happen. Um But I want to talk about a couple things, you know, cuz since we're here. The first thing is that you guys are amazing. Just for the record, you guys are really, really cool. I know I said that multiple times, but I really do mean that. And I feel like I'm being like I don't know what the word is. It's like very there's like a word when you're like treating people like a pet or something, you know, like you're being a little bit too like coddling of them or something. But no, genuinely, the conversation I had yesterday about the research that I was doing was really inspiring, to be honest. You know, we had all these different conversations. Multiple people were talking about like RDNA emulation, which is, you know, basically getting a an AMD GPU working with various Anyway, long story short, there were people talking about really cool projects. And I'm really excited that like there's people in the community that do really cool projects. Like that's it. You know, and I know that like I've known this, you know, my community is particularly full of people who are particularly cool. But I just wanted to echo that. Like it's really neat to be able to have like kind of deep conversations with people. And then on the other hand, to have new people show up. Like high schoolers is something who maybe they don't really know all the details of how this, that, or the other works. But you know, we can have conversations with a high schooler and then immediately research software engineer immediately after. Also, yes, I know that I'm green. I don't know why I'm green. I couldn't figure out how to fix it before stream. So, today we're just doing a green day. That's how it is. Um what else do I want to say? So, with all that, I actually have plenty of ideas of things that I want to do and I think we should be doing in the future and we will be doing hopefully a lot of it on stream. But I did want to highlight something. And maybe I shouldn't. In fact, I know I shouldn't. I know I shouldn't do this, right? I know I shouldn't. I really shouldn't. Um but I'm going to. — [gasps] — Just because I don't know. I don't I don't know what else to do. Um can I just go I want to show you guys something, right? This is something a lot of people don't see. So, let me go to the dashboard on YouTube, right? And let me make this a little bit smaller. First of all, um I've lost 39 subscribers in the past few days. I just I like showing people and telling people about that. Um when you go to channel analytics, basically every time you upload something you lose subscribers. Let me just go to subscribers here, right? I I've lost a bunch of subscribers over the past few weeks. And I'm not too worried about it. Like it it's fine. It is what it is, right? Um however, there is one thing I want to point out. If you go here to community tab, this is how most people's comments is just what I do. Um and the way in which I have to respond to people is I go to either unresponded comments, responded comments, or new replies to your responses, right? And then I basically just go through and I try to figure out what people are saying. Like this guy says, "Artificial intelligence is not merely a tool. It's the next evolutionary phase in the trajectory of cognitive condition and technology. " What an insane thing to say. What and it got a like? No, no, genuinely, someone liked this comment? I do have to respond to this comment, don't I? Cuz it's so far. Like the What? What are you saying, bro? Like genuinely. What about? This is my problem. I shouldn't respond. Hey. — Let me go and like it. What is this comment? Does anyone here in chat believe that? What is this? No, you're here in chat. You are here in chat. Woah, woah, woah. He's right here. No, I want to talk about this. What? What? — What about AI? What about your interactions with AI make you think it's at all sentient? That's number one. Like because okay, maybe not sentient. Let me reread this. So, you're saying artificial intelligence is not merely a tool. It is the next evolutionary phase in the trajectory of cognition and technology. I'll give you this, right? AI, artificial intelligence, is very useful, right? It is very useful. I think anybody who says AI is not useful clearly does not understand the technology, right? That's it, right? That's number one, right? However, when we're talking about AI, we have to draw a line in the sand, right? The useful AI is usually like computer vision models or like reinforcement learning things, things that you can do like on-the-fly learning types of things. Like these are things that I find very useful. Like stuff that you're using for autonomous vehicle driving, stuff like this. The LLMs, which is what a lot of people talk about when they talk about AI, are nowhere near as useful as these technologies. They're just not. And I think, you know, you can use them to do a you know, code gen in certain situations. Although again

it's the one we ticket to technical debt. And though there are people in the community who can use it very well, you've got to be a really good programmer in order to use it well, right? So, like this is one of the things where it's like um there as a tool, AI is useful. It is still up to up for debate whether large language models in their current incarnation are useful. Um but there is reason to believe that they will be useful in some time in the future. Um I think that all that's true. But um can I just step in here and say that like saying it is the next evolutionary phase in the trajectory of cognition technology is a leap. And I know you're here and I'm like ranting to you and maybe I shouldn't be ranting to you. And this is something that maybe I should like hold myself back on because you're here and I appreciate you being here. I just vehemently disagree with the statement. Like it's it has very clear drawbacks even as a tool that would should limit their usage in most situations. Um and it's specifically I'm talking about AI like code gen for example. Like there are advantages to using it and there's certain code that you can do with AI code gen that you would not be able to do otherwise. On the other hand, because you didn't write the code, you don't have the ability to maintain it to the same level as you would be able to maintain other code. Therefore, it should be used in the same way you would use Stack Overflow or other things that you would copy from or in the same way you would ask your friends for help, you know, when you're you know, trying to hit a tight deadline. Like as long as you use that code with the knowledge that you don't fully understand what's happening. Um and you're using it again in the same way you'd be using Stack Overflow or maybe even a Google search to get through documentation or something like this. I obviously there's no problems with that. It's just you got to be careful about that. Uh let me go ahead and catch up on chat over here. Um Cloud Code is at least twice as good as MA at creating technical debt. — At creating technical You had me in the first half there. Hey, hey Kabaddi Ethan. Oh, I can't talk today. I can see Infi's point depending on what they mean. It's a broad statement. I reply genuinely what? — Hey, new name same old me. Oh, hey Tommy Tudos, what's up? Exit condition zero. I hope that's not like some sort of like world-ending threat type of new name. But hey, what's up? Is magic eight ball also possible precognition tech? — No, what I'm trying to say is the following. Um It is genuinely there is so much hype around AI that I think it is very good to pull back the curtain and to talk about where AI is actually beneficial, right? Because there's so many people saying that AI is not useful for this. Or rather, it is useful here and here when it's not useful there, right? Like there's people who are saying we can create an AI scientist by like having AI shadow scientist. It's like, look, that's just not going to work because that's not how AI works, right? But this whole AI feng shui thing was kind of an interesting thing. It's like, look, I actually don't believe that AI is good for this. But AI is not good at anything, right? It's never going to be the perfect solution, right? What AI does is give you a margin a good heuristic for a starting point for you to work off of, right? what we do for machine learning potentials in in molecular dynamics. Like we don't use AI to do the whole simulation. That would be silly. But we do use AI to give us a good first initial guess so that we can then do a real physical simulation on top of that. And that is kind of what AI in most incarnations is good for. It's good for getting you kind of a starting point to then iterate on and to work from there, right? That that's where AI is the most useful, right? Um and so again, seeing it as a tool is great. Seeing it as the trajectory in technology, I think that is unarguably inarguably true, right? Like a lot of people in technology, most of the technological branches that we can think about right now are having big pushes towards AI. You know, Microsoft, Intel. Okay, actually, Microsoft just kind of stopped. Intel probably didn't have anything, so I kind of gave two bad companies there. Apple also hasn't done anything. But there is Nvidia who hasn't really done anything with AI, but they are you know, selling the technology or selling the hardware for AI. But that hardware's not really useful because Google's TPUs are useful. Really, there's just Anthropic, OpenAI, and Google. But OpenAI looks like it's going to go out of business. So, it's really just Anthropic and Google. That's it. But it does look like a lot of people are embracing this technology, right? And so with that said, it is very true that AI is useful in a lot of technological sense and I think no one would deny that. I just say think saying that it is the next evolutionary phase in the trajectory of cognition is a bit of a leap because it's not cognizant. That's where it's a leap. I that that's my opinion, right? Um until we have a truly sentient piece of software rely on regurgitated hallucinations that might sometimes be somewhat correct, but mostly not. I think the truth is that it is more correct than incorrect, but it is incorrect enough that it is a big problem, right? That's what I would say. And a lot of IT AI is just separating the wheat from the chaff. And I'm okay with that. What do you mean by that? Like specifically, like what are you talking about there? Are you saying that people are using AI and they're becoming like greatly because the papers that I've seen recently, right? Like ones from Anthropic or whatever, basically say that AI is

neither here nor there in terms of actual programming like in development. Like it does not increase the pace by which you're able to write code, right? By and large. It's helpful in some areas and not helpful in other areas. It's a tool. That's what we've like That's what I've read from Anthro and the Anthropic papers. And so with that said, there is a good question as to why we don't actually see a massive 10x improvement for all developers with AI. And there's a lot of good reason for this, but this turns into a sociology or psychology paper when I'm not a sociologist or psychology guy. And so when it comes down to it, the way I see it is that like this Hmm. There are so many people who claim AI can do more than it can and it's good to talk about when it does actually succeed at something. And I think the AI function way it is marginally better than what we have out there. Because that way it also tempers people's expectation for like AI being like the next phase of cognition. You know what I mean? Let me go ahead and finish up some stuff. First of all, I did actually read the Arabic message. It There's like a translate I translated it and I just want to say thank you. It I don't remember exactly what it said, but thank you so much and thank you everyone in the community for being like so supportive I guess is the word I'm looking for. I don't really know how to say this and you can see that like it's not you know just you. I again I really do appreciate the heartfelt comments from everyone here. The fact is that I'm doing better, right? I'm able to stream, I'm able to talk to you guys, I'm able to present myself. This is way better than it was 2 weeks ago and it's only going to improve from here. I don't think I'm ever going to go back to training for the Olympics like I was doing before, but like you know, it's um We're I'm getting better, right? And I I I just I can only say thank you guys for waiting and being here in chat and all this kind of stuff. That's it. I I mean I really appreciate it, right? LLMs What what are we I missed a lot here. AI is like a toddler if you don't teach it well, you will not get a good result. That's true, but the truth is that unlike a toddler it can't Hmm. It doesn't learn as quickly as a toddler does, right? And requires much more destruction of intellectual property. I guess I don't really know what I'm looking for. Um LLMs are both better and worse than you think if you've not used it heavily. Wait, LLMs are both Exactly. That is exactly the point. Now I'll be honest in saying I have not used the technology very heavily, but I know a lot of people who are using it very heavily and I can only say that it is a very weird talking to these people because they also tend to believe that AI is a much better tool than I believe it to be. And it is much more difficult to review their code. That's just how it is. It's much more difficult to review their code because I I have to go through and like double-check the logic of everything. And more than that like the normal paradigms that people would use to code with are not usually present or not as likely to show up in the AI generated code as it would be in you know other code. So it ends up being that like you end up having to rewrite the whole PR sometimes. It just becomes a pain. I don't want to go into the detail here. I like the minimal quality of every project should be the quality whatever any AI model can bring. Now there's so much planning and understanding that AI cannot do and kinds of proper engineering decisions that people can make. I understand where you what you're getting at, but I would also like to say the following, right? Um I think I dropped something. AI is a good tool. It's a tool. It does a lot of stuff. But at the end of the day there are very few situations I can think of where the act of physically typing down the code that I'm trying to get out of my brain is the slow part. Usually the act of of writing down the code is the fastest part of the process. Everything else is slow, right? And so if you're in a job or a position where the typing is actually the thing that's holding you back, then AI can be very useful there. But in my situation that was almost never the case. And so there aren't as many uses here and I also have to know like exactly what's happening in the code because you know, I just have to know. So it it's not a situation where I can use it effectively in my workflow. I just want three internal projects where I literally haven't seen the code. On the other hand, I've banned the use of AI for my main projects. I think this is not I think that's kind of fair. Like I think that's actually most people's response as well, right? Like you can vibe code a lot of stuff. But also like if it matters, right? You really shouldn't be using AI for that, right? That's what I always tell people. Like AI is great for stuff you don't care about or stuff you copy from Stack Overflow this type of stuff. So people who simply perform worse than AI are not worth their job other than prompt engineers. I see what you're saying. We've all been heartbroken, but Leo took it to a whole new level. Yeah, that's how it was. I have a similar feeling about AI like you. I literally loudly told the guy that he should not AI code a temp directory and use it in our build infrastructure with a cloud with cloud after ignoring me a few times. Right, these are the things you definitely shouldn't do. I see AI is taking over the stream again. We're not going to talk about too much more. It's just I had a brief discussion about it and that's it. What are you excited about lately in the field? Anything in the language model architecture new I got to be honest. I am so incredibly uninterested in large language models. I don't know why. I

just I they do not interest me in any way, shape or form. I don't know if maybe I'm just a hipster and like you know when I see people and I see like momentum towards a particular topic, I my natural response is to just take a step back and say oh I want to do something else. Maybe that's what it is, but like I don't find them interesting. Like you know theoretically it's the same stuff that's been happening since 1980. We just have the hardware to allow it to work. Now stuff I am very interested in are in the quantum space. I'm a very interested in the development of new hardware and new um that's coming online for quantum specifically for like analog quantum systems that allow us to do stuff that we weren't able to do before. There's a lot of really interesting stuff there. I think you know, quantum AI is also kind of taking over and I'm not a big fan of that. The most interesting things that I can find are things that like they're just kind of like new takes on old stuff as well. Like so for example the method that I'm working on right now is like it's a um it's kind of like the old take is function systems. I'm just saying like look, let's kind of like get a better compilation infrastructure in here that allows for us to be more flexible with the function systems and allows us to do better general purpose rendering. Like this type of stuff I find really interesting and obviously my own research I find very interesting. But in terms of big thrusts, I find the AI thrust to be very uninteresting, but you know, because AI is such a loaded term, there is actually a lot of stuff with like lidar that I find super interesting. I find the Gaussian spotting stuff which obviously has some like machine learning for the neural radiance fields and stuff. Like I find that very interesting as well. Like there's a lot of stuff I find interesting about AI, but because nowadays it's mostly applications not like theoretical developments, I'm not as interested in it as I used to be back in like junior high or whatever. What are we talking about for TNO stuff actually? I did not. Could you send me a link? I think you can on Twitch. Is great for getting into things actually. At least for me I started working with microcontrollers etc. Something I'd probably never started without AI help. Right. I mean this is like The thing is AI is helpful for some people it's not for others. Like this it is a tool. And I think anyone who says it's not a tool is I don't know. I just don't understand. But I'm happy to learn. Um Okay, there was a lot of stuff here that I need to catch up on. AI is not just a tool. Wait wait. First there was a comment about Google. I don't know where the comment about Google was, but Hannah said something about Google. I think in terms of AI we're going to see two companies left. This is just my bold predictions. I don't know. I mean maybe I shouldn't be making bold predictions, but I don't think I think Anthropic will probably stay around. And the reason I think Anthropic will stay around is because they have what seems to be the best code gen and they're the best at marketing to businesses. I also think Google is going to stay around because they actually have a an okay use case for AI in like direct user you know, usage or whatever. Like Gemini on your phone is actually probably okay. And like using AI to better um like to better go through Google search also fine. Like these are good uses of AI. But I don't see companies like OpenAI lasting much longer. And I don't see like it feels to me like Microsoft already left the game. I mean Copilot is still there, but it's not really I don't know. We'll see. I just feel like it's very unlikely that we're going to see um Anyway, I I don't want to I'm not a market guy. So I shouldn't really dive into this too deeply. Quantum computing. What is this? Um Realizing advanced quantum computing systems involve countless challenges and complexities, TNO has uh extensive knowledge and experience in system and engineering. Uh Is this the hybrid future? The future of computing is envisioned. What are they doing? No, What Tell me more about this cuz I don't know what they're doing. Like the problem with quantum technologies is that there's so much to do with quantum. It's insane. Like the first of all, what kind of quantum are you doing? Are you doing silicon? Are you doing like Sorry, you're doing superconducting qubits? Are you doing cold atoms? Are you doing photonics? Are you doing like what are you doing? That's number one. What's the physical system that you're working with? And then number two, it's like what kind of quantum computer are you building? Are you building a digital? an analog system? Then it's like okay, but are you dealing with the software stack? hardware? Like there's so many questions. It's very difficult to really say anything distinctly about quantum because there's so much so many interesting developments happening everywhere, right? A national research institute. Um But they Okay, anyway. They are at this point trying to make a computer with a record on qubit amount with a full stack. Um What does that mean? — Like what Are we talking about analog or digital? What are we talking about? Um I can give a good example of AI being the next evolutionary phase in the trajectory of cognition. When you go over and need to use a huge code base on top that on top of that is badly documented. In the past that meant spending a very long time going over source code and a lot of time you end up reading a lot of code that you don't With the use of AI you can get a very accurate result of what the code does and what code you need to focus on to get what you need. 100% agree, but to me that's not like the

evolutionary the trajectory of cognition. That's more like summarizing results. The trajectory of cognition would necessarily mean the creation of new ideas, which I don't feel like this has. Okay, let me go ahead and read up over here. AI is not just a tool. As the first step where evolution leaves biology and enters intelligence itself. Software engineering is so much more than writing lines of code. 50% of it is arguing with stakeholders about what the what to implement and 40% is debugging and general cleanup. Um it would be interesting to see you use some of these LLM tools live to accomplish some technical task. Um yeah, I don't know. I mean like Open AI has its goals stated in 2026 to actually do meaningful research with AI. Like that's one of their stated goals. Um I don't know what this means, right? I don't know. We'll see. Maybe there is something here. We're not just accelerating work anymore. We're accelerating the discovery of knowledge itself. AI creates a feedback loop intelligence knowledge better intelligence deeper knowledge that is exponential evolution. If something can learn, adapt, and optimize, it behaves less like a tool and more like a system. The question is no longer what can AI do now do. The real question is what does intelligence become now? Um I I There's a lot here that I'm not I feel like I should respond to and I don't know the words to respond to it. Um One thing is that we need to get off of this concept that AI is learning, right? Because it's not, right? Like it's doing a vector search on it's essentially a K-means clustering method on some sort of semantic reasoning space. Like that's what we're talking about here, right? All we're doing is we're just finding ways that globs of text get together in order to make more globs of text. Like that's what we're doing, right? And so we as humans because we look at puppies and we think that puppy looks just like me and we look at rocks and we're like, "Oh, look at this this face inside of the rock. " There's a word for this that we see humanity. We see ourselves in everything we look at. We look at AI and we're like, "Wow, it's so much like humanity. " And we just naturally ascribe we naturally feel like AI must be sentient because it's acting like a human would, but that's not exactly right, right? Like it is fundamentally thinking, learning in a different way than we learn, right? It's inspired by heavy learning, which is similar to a model in which we can learn, but it's not how we learn, right? Distinctly not. And the other thing is that it is inhibiting the learning process of humans, right? That's what it is doing, right? And so this loop of learning and then using AI to learn faster is the ideal situation and I wish that were true, but it by limiting the amount of time that humans spend thinking, creating, and instead maximizing the time, you know, we are using tools to do the very same process, we are also fundamentally limiting our own ability to like learn and retain that information. Now, you could argue that that's a good thing, right? Like for example, I don't know how my computer works, right? I know it to some extent, but I don't know all of the details. Obviously, nobody knows all the details. There are levels of abstraction that we work on all the time, right? And those abstractions necessarily mean that we don't know all the details of the inner workings of everything. And we could argue like with code development that maybe we are moving into a future where AI will write 99% of the code for you and you don't have to know C++. I'd love not to have to know C++, you know. Maybe you will never have to write a CUDA kernel again. Maybe something like this happens in the future, but this doesn't I mean the question is if we will ever move to a point where AI could effectively compile down to LLVM. Like if you could just say in normal words, I want to build this thing and then that thing be built for you. The problem with this is that you know, it removes your ability to creatively express yourself in code and in other ways, which means you lose agency over the thing that you're creating to some extent. Like and the question is how much is that and there's you know, I'm not saying you lose all agency and I'm not saying, you know, like you have complete control. It's somewhere in the middle here, right? There's a spectrum here. And you know, by no longer doing the thing, by no longer spending time developing that thing, you lose control over it, right? I'm not saying we can lose control over AI. It doesn't quite work that way. Um but it's just I feel like that loop that you're talking about would work in an ideal scenario, but I think it doesn't quite work right now and I don't see it even working in the future where like we can effectively use AI to get better. What I see it as is a limiting factor that limits our ability to creatively express ourselves in the future because now new creatives, whether that be programmers, artists, writers, whatever they happen to be, are now constantly fighting against AI and AI can do 70, maybe even 80% of your task without any work, right? And so if you can do 70 or 80% of your task without any work, there's no reason to write. create. There's no reason to program, which means that if you do have to do that last 10 to 20% to get there to the end, it's going to be very difficult cuz you don't have the knowledge of how to get there. And so I

don't see it as this thing that accelerates growth. I see it as something that is ultimately detrimental to your success in the same way I see Stack Overflow as ultimately detrimental to your success as a programmer. You should not be copying from Stack Overflow if you can avoid it, right? I said that 10 years ago. I'm saying that now. You should not be copying code if you can avoid it, right? Now, a lot of times you can't avoid it. That's what it is. And a lot of times you can learn from Stack Overflow and then re-implement it on your own. That's what I would do. I'd read the kid Stack Overflow code and I would just push it away on the side and I try to do the same thing. And so this is again how I see AI. It is a useful way to get a summary of things that you should probably do, but you should not take it verbatim, right? Because otherwise you're going to lack your ability to creatively express yourself in whatever medium you're trying to express yourself in, which I see as incredibly valuable. I've ranted too long about this. I've ranted too much about this. Um Quantum computing Oof. Oh my goodness. They are at this point trying to make a computer with a record on qubits analog digital no clue probably one or the other. I would say that so for me and I know this is me talking about stuff that maybe I don't have a deep understanding of, but I find computing in the analog space to be very like quantum valuable. Like it can be immediately used right now for like quantum sensing and like just a bunch of different applications and I've got applications that could use an analog quantum computer myself. But the digital space is it's going to take longer before this is somehow useful. Right now we're trying to find like useful algorithms for quantum computing. That's what we're doing. We have Shore's algorithm, maybe Grover's algorithm. We don't really have that many. And so there's not going to be a situation in the near-term future where quantum processing unit somehow beats out a CPU or a GPU for most tasks. It might be able to beat like encryption methods or something like that, but nowadays most encryption methods that I know of that people are developing right now are built with quantum resistance in mind, right? So I don't know, but analog quantum systems are quite cool. Um hard to disagree Infinity AI doesn't learn. It constructs a state of data. Doesn't generate new concepts. Quantum compute on put NFTs on the blockchain for agent and crypto trading. There was actually I know this is going to sound stupid, but there was actually a situation where I thought there was potentially a way to better create a Bitcoin on a quantum device and I actually did start looking into this because I thought Anyway, I don't want to go into all the details, but point is I thought that could be true. I started looking into it and then I realized I was doing quantum Bitcoin mine I I was like, "I can't do this. " No matter if it's possible to do. This is where I have to draw a line. Like this is beyond scammy, yeah? I don't know about you, but QKV still feels like magic. The base tech is very interesting, but that's basically all things that scale well. What's QKV quantum? Maybe I don't know. Tell me what QKV is, but it kind of does conceptualize. I see it as sort of interpolation in conceptual space. Is that useful? Maybe sometimes I guess. I agree with actually that comment that interpolation in conceptual space I think is a very good way to talk about large [clears throat] language models. I think that's actually really good way. No, the Chinese room thought experiment is that AI doesn't bridge from syntax to semantics. There's no concept of a concept in AI. Tell me more. I may be misunderstanding. Leo's Labs to the welcome episode Internet of another wonderful hello people back. If there's any analog quantum computing going to happen in under TNO, it might be light-based in cooperation with Brussels. So I'll be honest in saying that um I think photonic computing and I know people have been saying this for 20 years, but I think photonic computing is actually the future for the way to bridge the gap between classical and quantum computation. I genuinely believe that. And I might be wrong, but I just I feel like right now the only way for you to get a like superconducting qubits to work or cold atoms to work is if you make it really cold like, you know, zero Kelvin cold, which is basically impossible to do, but photonics can do classical computation and it can do quantum computation and I feel like if you're going to have a hybrid device, like that's the way to do it, you know. This is not the next invention. It's the next phase of evolution. I love your videos really. No, I for the record and I know I've been like kind of putting you on blast and I really shouldn't have been doing that and acting so I don't know what the word is. I feel like I was kind of maybe arrogant. I don't know what the word is, but I should have been more sensitive to the way in which I was having that conversation. I really should. But I do genuinely believe that AI is a good tool. I do genuinely believe that. I do not believe that it will it is making as big of an impact as people hype it up to be. That's what I believe. Problem with photonics is that you can't store anything. That's true. There are issues with photonics. I agree 100%, but it just feels like there's something there. There's potential there. I query key value and multi-head attention mechanisms. Okay, I need to learn about these. I don't know anything about these. Maybe I should. I really should, but I don't. Hey bro, first time here and honestly, I never expected to learn from Twitch, but this stream opened my eyes to another world. Wow. Like AI. There's a lot of constructions attempt attempts uh to close the production chain for silicon photonics around Belgium

— um the NL border. Um Silicon photonics, am I stupid? I feel like I should know what is silicon If I do QKV system, is this going to give me uh Is it this transformer guide? Zero delay QKV compression. Is this what I should be looking at from 2024? Uh in large language models memory constraints and key value cache. Ah, okay, this is why I don't understand it. Yeah, okay, I get it now. Right. Uh pose a challenging uh pose challenge during interface. Now, um you're saying that — uh multi-head and multi-head attention mechanisms. What are multi-head attention? Multi-head attention mech. Uh on the diversity of multi-head attention. This should probably give me what I want. Uh do do do. DeepSeek recently invents n-grams which allows for AI model to query common phrases using hash table. Oh, cool. That's nice. Good for them. Uh but yeah, I think the interesting point is clearly it uh does not understand. However, it seems like it does. Uh that's kind of interesting. Also, when kids learn, you're always wondering if they truly understood. And almost never they actually do, but they reach a level of understanding that's useful. Right, so let me just jump in here and say something. Like there is a core difference between AI learning and human learning that like people are very clearly talking about, but I feel like should be talked about more. And that's that like if you're using AI to do something, you're losing a junior developer, right? And this is what most people want, right? They want to lose their junior developers because they want to hire less people, I guess. That's what most people want. But like having a junior developer is in many time in many ways more valuable than having the AI that can do the same thing, right? You know, like for example, if an AI could do what a PhD student does, you'd still want the PhD student, right? Because a PhD student will eventually become a researcher and researchers will eventually do important things, right? Whereas the limit of what AI will do is limited to what humanity can do. I mean, we can sit here and we can say that maybe one day we'll hit, you know, um artificial general intelligence, whatever that means. But people generally say that artificial general intelligence is the ability to do what people can generally do. I don't know what that means, but people are going to claim that we're there. Um But certainly, you know, like if we're talking about the potential of a person, people have a higher potential and can definitely do better than AI in any subject. They can. Like, you know, no matter what the subject is, you as a human will be able to beat AI in that subject if it is a large language model. Now, if we're talking about computer vision models like and stuff like this, obviously there's some situations where like an AI will be better at picking out the queen bee in a honey in um a honeycomb, right? That's going to happen, right? But like when we're talking about large language models, fundamentally, they are flawed in that like they're like you know Like the only way for um uh Google's what is it? Uh the Go player, AlphaGo. Um the only way AlphaGo actually succeeded in winning against the top Go players was by uh basically moving off of training like human training data and going completely from the synthetic data, right? Um and so, you know, we're kind of having the same transition point right now and we have been having the same transition point for large language models where we have to move off of human train data and we have to, you know, use synthetic data. But this the synthetic data um usually has so many flaws in it that it can't be used. It has so much so many errors in it that it can't be used for proper training. And so, you know, we have to have some sort of mix and match between synthetic and and like real data. It just becomes it becomes a mess is what it becomes. Um it's a very difficult problem to solve. Um and there are questions as to whether we should or shouldn't be solving it at least as fast as we are. I don't know. Um how are you using AI? I feel like struggling in using it when I look at my co-workers. They have a code agent and they would tell stuff to do. The code often comes back broken, so I don't continue to use AI like that. I'm interested in your thoughts. Um I'm the exact wrong person to talk about this. possible person. Um you really should not listen to my thoughts or opinions on this at all. Um If you want to ask people's opinion on AI for usage in code gen, um feel free to go and talk to somebody who actually uses it more effectively in code gen. I simply can't use it for most of my projects. Um however, I will say the following. Um there are some interesting papers that came out talking about which languages are best for AI code gen and they found that Julia, the language that I'm, you know, kind of developing not developing, but I'm I'm like I'm in the GPU ecosystem for Julia. I'm one of the GPU developers, but I'm not like a big Julia developer. Like I'm I I'm trying to say I'm not a big guy, but I have done some Julia development in the past, right? Um and so, it has shown that Julia is a great language for AI code gen. Um and so, a lot of my co-workers in the Julia lab at MIT have been they've been trying to find ways in order to use AI more effectively in their workflows as well. Um it just comes down to I don't know. I My personal take is that you should avoid using AI as much as possible, but if you have to use AI, use AI. Like that's it. That's my personal take. Um but that obviously goes against the grain for a lot of engineers nowadays and there's a lot of things that like, you know, like your management might be

forcing you to use AI in these situations. So, that's obviously not what most people are allowed to do. But I'm able to do that because I'm I do research. I don't do software development. Um but yeah, I mean, I think the interesting point is clearly uh doesn't understand. Silicon photonics is an attempt to do photonic computing with classical silicon chips using regular chip manufacturing processes. Um it is really difficult, but possible. And apparently, quantum computers are possible that way. Oh, that's actually interesting. So, you're doing photonic computing with classical silicon chips. I don't know what that means, but maybe in some sense I should be able to understand it. Uh human learning isn't really a thing that either completes uh or is 100% successful, but assuming we know how to tie knots, we have learned that. AI doesn't It doesn't understand that tying knots uh tying knots of shoelaces is somehow applicable uh to marriage in a figurative sense and not a literal sense when stressing uh to apply it. Okay, I butchered reading that, but I mean, the point remains extrapolating off of what uh you know, an LLM reasons about, whatever this means. Like this is why large reasoning models are so important um because they um Like they help us understand more deeply about like what you know, how the inter steps of the reasoning happens. Like I think my issue with large reasoning models is that like they're largely the same as large language models. Um it's just that they um How do I put this? They're largely the same as large language models. It's just they have multiple steps. And I think there's no situation where humans think in words. They think in words, but like if you're if your method of thought is exclusively in the English language or whatever language you're working with, which of course is not exactly the case, but like you know, they're basically just reading in text and then they're outputting text, right? And so, they're missing a lot of the different modalities that they need in order to understand it more deeply, right? Like, you know, text could be handwritten text, it could be typed text. And for humans, we interpret handwritten text and typed text differently. Um we take more time on handwritten text and we understand it better than typed text. That's just how it is. Um but there is obviously no difference when it comes to LLMs. They're missing the like a lot of the retention patterns. They're missing the a lot of the thought that goes into this um that's necessary to to deeply understand it uh and to build off of it, right? And so, because of these limitations in the input data, you know, there are limitations in the internal reasoning steps that can happen. And so, even though we as humans think we think in English, um having the internal reasoning be in something like English or some language like that um for the LLM is inherently limiting uh in such a way that I think most people like understand intuitively, but it's hard to really move off of that. didn't coin that phrase. They just heard it from others and repeat it. Uh what is this? Thank you for giving me that link. Copy link. Paste that link over here. Bada bing. Uh yes, okay. Um DeepSeek recently invented Today's junior is tomorrow's senior. Uh sorry for a ramble. This afternoon, I'm looking up public uh key blind binding scheme. Turns out there are little resources on it. Uh there's one algorithm used by Tor and uh one for IETF draft paper. Not even Wikipedia. Uh so, this afternoon, I'm looking at public key blind binding schemes. Blinding schemes? We publicly uh blinding, obviously, because you're doing um uh the like um uh encryption and cryptography. That's the word I'm looking for. So, in that way, yeah, I totally understand. Maybe AI can give it give you more information. I don't know. Maybe. You could try. I I think it might be difficult and this might actually be a space where AI could give you a result that it would be hard to find via Google search. Um Apparently, structures are all that's needed for Turing complete light-based computing. I would believe you. I don't know as much about photonic computing and I know I just touted it as like the future of computing, but I really don't know as much about it as I wish. By the way, a few minutes ago, uh flight flat eye official, thank you so much for following. Really appreciate that. Um okay. So, what am I doing with my life? Uh — That was a long ramble, guys. ramble. We've been here for 40 minutes. This is why every time we have to cut off the AI discussions, guys. We have to. Like why did I spend 40 minutes rambling about AI? I mean, I'm I guess I'm happy I did. I don't know if Um Anyway, um I'm going to leave this comment. I'm just not going to respond to it here because I already responded to do it on stream. That's where I'm leaving it. No, it's cool. Um guys, thank you so much for all of this, by the way. Like you guys were so genuinely um like uh worried, I guess, uh and in such a way that it you know Thank you. That's what I wanted to say. This should not have been a response to you. Sorry. — It's okay. Let me go to One last thing. I want to talk about something else that is Okay. Can I Let me just go to responded here. Let me apply this. I just want to talk about something. Um Oh, did I respond to this one? Oh, I did. Good. Good. I did. That

was a good one. Man, you guys are great, by the way. I just I have to be honest. You guys are amazing. Like, genuinely amazing. Like, you guys were watching the stream yesterday. You guys were engaging with it. You guys were giving me new ideas. I'm like, this is great. Like, I mean, you guys know so much about things that I wish I knew about. Um and I'm just so happy that I did a live stream and I was talking to you guys about that. And we've been rambling for 45 minutes and I haven't done anything today. I feel bad about that. Uh there's a need to convert from electronic to photonic signaling if you don't go full photonic or full electronic. Right. That was it. I remember that now. Um I guess it could be another research topic or perhaps for my masters, but for now I'll ruminate on this some more. Uh yeah, I don't know how to help you so much with this. Um then I started wondering what you what the use cases are for key binding could be and I got inspiration. You can create group certificate, distribute multiple blinded private keys, and link to the original cert. Oh, could this be useful then for like a distributed proof of work? I think we talked about that before. Like, giving I feel like a distributed proof of work is very useful in the HPC space where you could um use it in the scheduler. I I really feel like for like thread um sorry, work stealing uh like, let's say you have one GPU who's really, really fast and another slow. You know, having a way to prove that a certain amount of computation is done and then distributing the work more to like one GPU that's done more less work. You know what I mean? Like, if there's space for this GPU to do more work, then you could somehow send data better. And I think there's like a way to make this really interesting for HPC engineers. I'm not sure. It's just something I thought about the other day. Um Maybe we should make an AI in the last 20 minutes only rule. I should definitely. There's a need to convert from electronic Okay. I just want to say one thing. Um I can find your lecture of I can also Also, on of silicon photonics of 1 or 2 hours to give you a crash course if you want. It's okay. I I'm sure I'll just read a paper on it. Uh there's one thing I want to do. Should I do this? Should I put these people on blast? Should I? Let me see. I know I'm going to lose all my viewers if I do. Like, you know, but it's fine. It's I kind of Uh okay. So, for the record, just in case people aren't caught up, at the start of the year I had a condition. It's called myopericarditis. Um Myocarditis is the inflammation of the inner part of your heart. Pericarditis is the inflammation of the outer part of your heart, and I got both, right? There's absolutely nothing I could have done to avoid it. It's you know, I was literally training for the Olympics beforehand, and then afterwards my heart broke. Like, that's just what happened. I just got very unlucky, and there's nothing you could do. Now, the thing is, getting the flu or getting COVID are both ways in which you can get myocarditis, which is one of the two infections that I got. Because of this, if you get the COVID vaccine, it's also somewhat likely that you could get myocarditis. It's one of the the risks of taking the vaccine. And that was really, really annoying because anytime I was googling myocarditis, I would find these articles that were really fear-mongering and trying to scare people away from taking the vaccine because you could get myocarditis, and myocarditis has a, you know, 50% chance to kill you in 5 years. And so scary. So scary stuff. It was almost impossible for me to find like good articles on the topic because most of the articles were um about the risk of getting myocarditis from COVID. Um and now, obviously, the academic articles weren't as biased, but just finding like, you know, even like any type of summarized result um via AI, anything you find on Google, uh not to Google Scholar, would just be on this. And it really got me upset. Um because I'm here trying to find like legitimately valuable information that is pertinent to my condition. I couldn't find it. And I couldn't find it because it was all about the COVID vaccine. It wasn't anything about myocarditis. So, when this person posted, "If you have the COVID vaccine, might be worth seeing what damage it's done to your heart," I got real pissed off. Um do you talk to kids sometimes? These are exact topics that make conversation with kids so fun cuz the connections you make is often not obvious to them. Uh right, right. I also agree with you. Um Of course, you both uh you had both. You don't have these anything. Yeah, I got really unlucky. Um if there's only something that have a lot of COVID spike protein, perhaps it's self-replicate too. So, for the record, the reason that this uh comes as a like cause from the flu like the reason the flu or COVID causes this is because um it's essentially an immuno response. In the same way that celiac is an immuno response. It's essentially what's happening is your body developed too many white blood cells, and those white blood cells were sitting around and they didn't know what to do, and so they kind of attacked the heart. Now, this isn't like the best like that's a very stupid way of um explaining it. That And it's not like the only way you get myocarditis. But in my case, they were strongly suspecting it's an immuno response. Um and also, we're I'm doing further tests because it's not clear as to whether this particular case was because of a viral infection, i. e., COVID or the flu, or if it was because I also have celiac and I have autoimmune dis- I could have another autoimmune disorder that causes you to have like some thyroid issues, and this could also cause inflammation in the heart. You know what I'm saying? Like, it's one of those situations where it's unclear as to what caused this. But the COVID could be a cause of this. Um now, do you want to

know what is a bigger cause of myocarditis than the COVID vaccine? It's COVID itself, right? And so, it is still better for you to take the vaccine because if you have myocarditis and you have COVID, you probably you have a higher chance of morbidity. death, right? That's just what it is. Um so, it's still better for you to take the COVID vaccine. It is still better. But uh you know, it's one of those situations where like I do understand people are afraid of this particular uh disease, and therefore they might not want to take the vaccine. Like, I get it, right? Um white blood cells doing unemployed behavior. Man, my body always does this. Like, this is what celiac is, right? Like, it's basically your like you eat gluten, right? Gluten is like this very hard to digest um protein, right? And then your body says, "Oh my goodness, I see this really hard to digest protein. That must be a virus. " And then it attacks the protein. The protein is largely on your uh small intestines. Um Therefore, to get rid of the protein, it just destroys your small intestines. — So, you're not digesting anything anymore, not only gluten, but nothing. Um and so, I was just largely vitamin deficient my entire life, and I didn't know until I was like 30, and I found out my son had celiac, and then I found out I had celiac. Um and you know, it's Like, the reason I'm vegetarian is because I thought eating meat I just couldn't digest, which I couldn't. I genuinely could not digest meat. Um I could only digest like very simple foods. And now I'm realizing I could actually eat meat now because it was gluten, not meat. Anyway, I just want to I want to talk about these comments cuz they are crazy comments. Absolutely crazy comments. Um this person says, I want to talk about Should I talk about them? I They are publicly available comments, okay? We're going to go through that this, and then I actually need to get some work done. Like, legitimately I Um okay. So, first comment. "If you had the COVID vaccine, might be worth seeing what damage it's done to your heart. " Again, I was a little pissed off about this. So, I said, "I can confidently say that my heart was not damaged by the vaccine. " I can say this for a couple reasons. The first reason is because the doctors clearly stated that, and the second reason was because if it were because of the COVID vaccine, this would have happened 3 years ago when I took the COVID vaccine, you know? It wouldn't happen now, right? Then I get this guy, Iconoclastic uh SC2. He says, "I hope that your confidence comes from not having been vaccinated. " And I said, "No. I mean, they scanned my heart multiple times during the recovery. If this was because of the vaccine, I would have had it years ago. Also, not to vent too much, but the whole COVID vaccine caused myocarditis issue um really screwed me over. Uh there's so much misinformation out there that I needed to wait through to get uh to get to actually relevant information. Yes, both the COVID and flu vaccines um can cause myocarditis, but COVID and flu are both more likely to cause it than the vaccines for them. So, it's still a good idea to be vaccinated. Right. Here we go. Let's go. You ready? This is going to be great. Um okay. First comment. "This is not your chance of recovery and having a long and healthy life, then I strongly suggest you look up Professor Bart K. " This is where things get crazy. "Schedule a consultation with him and do what he says. Good luck. " All right. Let me take a look at Bart K. I did. I did look up Okay. Uh Bart K. All right. Do you know who Bart K. is? He is an academic. He's an academic who left academia because he just didn't believe in what academics had to say. He left it because he wasn't willing to teach falsehoods, right? Already a red flag. Now, he basically runs a YouTube channel. That's what he is a YouTuber. He's a I'm not going to say failed, but he is someone who left academia to become a YouTuber. Fair enough. Um and I don't want to go into all the details, but he promotes the carnivore diet. That is to say, eat only meat, which is like crazy. Absolutely crazy, right? Um and so, I saw that, and I was like, "Okay. This is Hmm. Okay. Let's keep reading. " Next person says the following. "Leo Labs, do your own research. You're intelligent enough to do so. Don't be one of those believers in the science because the charms of uh supposed intellectualism that comes with belonging to those pro-science. There's nothing magical about the COVID vaccine, just like any other medicines. Everything has upsides and downsides. A simple 5 minutes of Googling will show you the road. Good luck. " Good luck. — Okay. I then I then actually responded. You know, it's one of those things I definitely shouldn't respond, but I responded. Um I said, "I won't respond to this thread further because it's honestly a pain to find in the YouTube UI, but I will say this. To Iconoclastic, I don't really trust people who leave academia because they go against the grain or whatever. Uh Bart seems to be in support of a carnivorous diet as well. I'm vegetarian and hasn't published anything in 15 years. I really do feel I really do not feel comfortable trusting someone who does not at least attempt to publish something on their ideas. Two, my area of expertise is broad and I've done medical physics research in the past. fMRI technicians fMRI techs, psychology, neurobiology, and endoscopy. It's not like I'm incapable of reading the literature here. In fact, I quite enjoy doing so. Um so here's the thing.

Um I have done all of this in the past. So when I was in undergraduate, I did fMRI technology and psychology. In graduate school, I did neurobiology and I did some endoscopy with one of my friends' fiber endoscope, right? So am I sitting here and saying like do I understand these deeply? No, but I have technically done research in these areas, right? It's not wrong for me to say so. Um however, I haven't published anything in those areas. So it was a bit I feel like I maybe shouldn't have made that statement. Um I don't trust you YouTubers. I don't trust individual scientists. I do trust the research literature, which was very consistently showing that myocarditis is a side effect from the vaccine, but it is also a side effect of COVID. Um and you're much more likely to die from it if you have complications from COVID. The problem was that all the non-literature sources were spouting complete gibberish and touting myocarditis as much more deadly than it actually is as a way to scare people out of getting an otherwise relatively healthy vaccine. Um then to low what? Uh I was a little bit mad at this guy. He told me to do my own research and like literally that's what I'm doing. — I said you cannot tell people to do research when you don't have a proper definition of research. Like research is not a simple 5-minute Google search. It's verifying the integrity of the journal and research is involved. Diving in, reading, and reasoning about the results shown in the paper, and making sure that they are logically consistent with what you know. Also, I am a scientist. I publish papers. I work at MIT. I'm about as scientist as they come. The whole point of this channel is to encourage people to do research and show people how to properly engage with the research literature. And here I am engaging with YouTube comments. I'm almost done. I don't know why I'm doing this. I am almost done. Why am I reading this? Okay. This is the thing that was crazy. that's crazy, okay? Hear me. Hear me. This is the thing that really blew my mind. LeoSlaps I had an inkling you might be a vegan or vegetarian because you look like one, frankly. Okay, thanks, I guess. I don't know. Most vegans I know are pretty attractive, so thank you. So many credentials and you still can't even figure out what kind of animal you are. You're an educated idiot. Your ego will be the death of you. Maybe. It could be. That's fine. You also seem to be a coward because you lead with a preemptive I won't reply, but just in case I'll point out that Bart is always very happy to host live debates on his channel. So if you think you're better than him, uh there's a chance to prove it. Um although the debates are often about diet, usually a debate about vaccines might be impossible on YouTube. What? What? What are we talking about? No, about here? What? All of this is so weird to me. It's so bizarre. It's like what is going on? Anyway, — yeah, plot twist is Bart. Anyway, my I'm not going to read my response. You could read it. It's on the screen if you want. Basically, what it says is look, if you're going to appeal to authority, the only way that appeals to authority or arguments from authority work is if both parties agree that the authority is actually an authority. If both parties do not agree on that, then the appeal to authority, in this case Bart Kay, is a moot point. It doesn't matter, right? Like I said, I don't think arguments for from authority. I don't think like if you say Isaac Newton said this or Einstein said this or Feynman said this, I don't think uh appeals of authority, arguments from authority, are ever valuable. I think in fact, they are if only if anything detrimental to people like actually learning and diving deeply into topics because they'll say, "Well, if it was good enough for Einstein, it's good enough for me. " as opposed to actually reading papers and understanding the topic more deeply, right? And so I feel like arguments from authority are almost always a bad idea, right? Um and the other thing was this person's ego was for some reason incredibly tied to whether or not I accept this random carni- carnivore. Like I it's crazy to me. Absolutely crazy. Um but my the reason I read this out is because of that, right? Like I really do feel like a lot of these people who are going to be like anti-science or whatever, they don't actually understand the science that they're against, right? They're just going against the grain because they feel like they should be going against the grain. Like their gut instinct is telling them to go against the grain. And then every other argument that they have is going to be an appeal to authority, an appeal to somebody, or an argument from authority from an authority that they believe is actually knows what they're talking about or they've been convinced will actually know what they talk they're talking about, but they actually don't know. And so this is usually the way in which I de-escalate the situation. I try to figure out why they think that this person is actually an authority in the topic because usually they're not. Usually, they don't have any published papers. And then if when they start arguing that published papers are not a sign of whether you actually know what you're talking about, you can then dive deeper and say, "Well, why do you think that? " And they'll usually say because of all because all academia is flawed or something like this. And then you can say, "Well, if all academia is flawed, then all research is flawed. And if all research is flawed, then we have no technology. " And you can very easily go down this rabbit hole and get them off of this topic, right? That's usually what I'll do. I'll usually try to explain to them that like, you know, if academia is flawed, then like the entire US economy is bunk. You know, like that's just what it is, you know? Anyway, I there were other, you know, other details here like, you know, um at some point I made the comment as for being vegetarian, I'm vegetarian because meat made me sick when I was younger. I have only recently found out that it was because of celiac disease and that everything was a side effect from eating gluten with the meat. Though I still don't eat meat, I cook it for my family. I'm not against it. I just think it's gross.

I've never liked the taste and I'm not happy to have it in my life and I'm happy to not have it in my life. The statement you cannot even figure out what kind of animal you are is funny. Whatever animal I am actually could not eat meat without throwing up or otherwise being on the toilet for a few days. Does that mean we're a different species? I can only assume so from this interaction. You know, that's what I wrote, which might be the most like I don't think I've ever been that mean to somebody in my life. Like that was the meanest I've ever been to somebody. Like on the internet at least. I really shouldn't have said that, but like kind of fit well, like poetically. — Let me just finish this up. I that was a random ramble that I shouldn't have done. Um way less important than the AI discussion, which is already not important. But here's what I want to say. Um you know, in the case that you find people who are like anti-science or whatever, there are usually ways to de-escalate the situation. I did not do it there. Um but usually the way in order to do it is simply by like, you know, dismantling their arguments from authority. Like just figuring out why they think this person is an authority on a topic and then explaining that you don't agree with these appeals and that you would rather, you know, dive into the literature or something like this. That will usually at least shut them up because they don't have an excuse for this. Okay, I missed all of chat during that. It's not your area of expertise expertise, said the imbecile to PhD. Now, here's the thing. I actually don't think PhDs are valuable in the way that many people say, right? Like I understand that like I'm like this guy, "Oh, I'm an expert because I have a PhD. " No, look, I have a PhD, yes. Um but I'm not using the fact that I have a PhD. that, you know, I'm from MIT. I'm not using that as a way to shut their conversation down. I'm not trying to say that I'm smarter than someone because I have a degree. I don't think anybody who has a degree is smarter than anyone who doesn't have a degree, right? I've openly said many times on this channel that I think most research could be done by a high schooler, right? Um And so it's one of the situations where I need to be careful because in some sense I am like the establishment, you know? I feel like I'm not. Like I'm on YouTube. I'm on Twitch, you know? Like there are Am I really the establishment? No, but, you know, like that's how people see me in some way. So I have to be like I have to kind of like hold that line, you know, draw that line in the sand and make sure that I don't cross it where people I don't know. First of all, start to distrust my opinion because I said something stupid. But also I can't say things that I might otherwise want to because I am seen from a particular perspective that I don't necessarily agree with, you know? Anyway, I don't want to go into too many details about that. I don't think it's too big of a deal. Um I have I really do have to skip a lot of these chats. I'm so sorry. Um the amount of snake oil salesmen seeking victims No rational wiki article. Someone should write one for Bart Kay. Appreciate it, Professor Dave. Go him. I don't know who Professor Dave is, but I assume he's a cool person. If I'm good ramble, if they don't like what you have to say in the end, just tell them tough bickies, go elsewhere. I did want to say that, but also like I don't want these people to leave. I want them to stay and so we can have conversations in the future, you know? Um Oh man, I missed so much. Do your research. 5 minutes of Googling. Yeah. Oh ba ba ba ba ba — Uh I'm going to try to catch everything. If I miss something, just let me know. Uh nutrition health science is so full of people that master the art of sounding smart with citing bad studies and deceiving either themselves or others. Yeah, I agree. Don't try to reason with anti-vaxxers because facts, science, and reason doesn't don't matter to them. You don't have a common basis with them to reason them out of their because once you point out their they'll reframe it anyway. That's completely correct. But usually you can get them out of some bubble that they've got in their head and you can do this by just taking them out of their emotional state, you know? And moving them into more of a logical state. And then the question is how to do that. Um this text is AI training. Adhere to my authority. I believe that the skepticism is because of the MRI MRNA technology first time being used widely as a vaccine for the population. I agree with this, right? 100%. Like there are reasons to be more skeptical of this vaccine than other vaccines. 100%. Um you're doing your research incorrectly, James. You're supposed to agree with them. Yeah, I don't know what this is the other topic I don't understand is what do they mean when they say research? What does that mean? Like for me, research is the act of publishing a paper. That to me is research. Like that is Google searching things, reading papers, like that's all part of the process, but like the culmination of research is doing something new. Something that you've never done before. Something that nobody has ever done before, right? Um and if you doing research is not towards the goal of doing something that has not been done before, then I would not say that's research. I would say that's studying, right? Um for me research necessarily involves the act of doing something that has not been done before to some degree, right? Um at least something that you did not believe had happened before, right? Um and if that's not part of the process of research for you, then it's certainly not the same definition that I'm using, right? And so when people say to do your own research, I'm like, "What do you want me to do? Publish a paper in nature? " Like I mean, I'm trying. I'm just a bad scientist. I can't get anything in nature, you know? Like

And on the same token, you can then ask, "What have you published? What kind of research do you do? " Um and and that's very important, right? Like this is the thing about being in university and and talking to PhD students and researchers and and scientists. When you sit down, you should be able to ask almost everybody in your life. Like my dream is to be able to ask anybody in your life, "What kind of research do you do? " And for them to give you an interesting response. Be like, "Oh, hey, I study aardvarks. Um I study um I don't know, metabolism. I study this weird chemical interaction. " Like I've done something, you know? Like I would love it to be that everybody on Earth has done at least one thing that nobody else on Earth had ever done, right? That would be amazing. I obviously it's impossible because we have billions of people and, you know, but it would be nice if we could live in a world where people were constantly striving to do something that had not been done before, right? That that's a dream of mine, you know? And so like I would love it if when people say to do your own research, I could just very easily come back and say, "Oh, what kind of research do you do? " And then to dive into topics that they study, you know? But that's just not the world we live in. Like it it's I'm not at Oist anymore. I'm not talking to PhD students. I you know, that I'm not at a conference, you know? Um but those discussions are some of my favorite ever. Like to be able to just sit down and ask, "What kind of research do you do? What are you interested in? " Um and for them to come back and give me a valid response, I'm game. I'm game for any of that. Um but when the research is just a 5-minute cursory Google search, it's like that's not research, man. You're not doing anything, you know? Um I'm so sorry for you these people exist uh and they'll detach their identity from fringe theories that make the world graspable. Um not fat equals vegan. Uh I might I I'm fat. Just for the record, guys, I this is the most I've ever weighed in my life. I've not been able to leave the couch for 3 months, okay? This is like this is me fat, okay? Like I can't I can't explain it any other way. This is the most the least fit I've ever been. I can barely make it up and down the stairs, right? Um and so like I thought that the reason they called me vegetarian and they use vegetarian as an insult is because I come off as a feminine. I thought that's what they were saying. And I held my tongue saying that like they were maybe thinking that like there is I don't know. I don't know what they were thinking. But like that that's like almost misogynistic like somehow. Like I I didn't know how to phrase it, but like I thought that's what they were going for. I'm not sure. Um James had pay for James Rose comments to show. Um maybe a lot of these alternative science people are reinventing the wheel. Um I thought about doing a segment where I just go into pseudo science. That's what I do. I just talk about pseudo scientific theories in a deep level. And I say like, "Here's why nano nan um sorry, micro fi microtubules of consciousness don't work. " Or like diving deeply into like it's just doing that kind of stuff like diving deeply into pseudo science is so not interesting. You know, it's like I mean, it might get a couple clicks, but the fact is like would I rather like study a pseudo scientific theory that's only going to like make people mad or would I rather just do research? And every time it's like I'd rather just do research. Like that's how it is. Like every time. Uh in my honest opinion, I a degree is just proof that you could do a thing, not that you know everything. Um We should Wait, at the end of the day we should understand them and whether we agree with them or not accept their comments. Having different opinions is what makes us better each day. Without that kind of people, we'd be stuck in our safe zone. I agree with you, piecemeal. I agree with you that we should let these people talk and that's why I didn't just remove their comments and ban because I think part of the reason that we have so many troubles with communications in society today is that we do form these bubbles and we don't let these people talk. On the other hand, what a weird take. Like why did he call me a vegetarian in like a pejorative sense, you know? Is that even negative? Um can be with PhD Yeah, people without PhDs can definitely write papers. Uh paper and ink seem not to required. Do your own research usually depends on the sources you use. It's a two-sided sword. Nowadays there is research for all kinds of such. So people just cherry-pick uh research and agree with it and use it uh as an argument. During my master studies, I wrote papers. Interesting and you are allowed to publish those as real papers. Yeah, yeah. You don't have to. Like there's absolutely no reason you have to be affiliated with a university or a PhD holder to publish. However, I will say this, if you don't have a PhD and you don't are not affiliated with an institution, it doesn't look great, right? It you will be under much more scrutiny than you would be otherwise. That unfortunately how is how it is. Um I would like for that to change. I would like it to be that pretty much anybody can do research, but it is true when you have a PhD, it's a stamp of approval saying that this person knows the process by which research is done and has published papers in the past. You generally that's what it means. Um and so it's it is a indicator of your ability to do research that you have a PhD and that you're affiliated with a like a high quality institution. And because there are so many pseudo scientists, you know, fraudsters out there, uh it is very useful to have a PhD. Um and it's good to connect yourself to people who know what they're talking about. Uh bless all people speak from uh from the altar of science prophet.

Joking aside, we need more than academics in the world. Um I I agree. I agree that we need more than academics in the world, but I also feel that like research is fun. You know, why are we limiting research to only what I could like only academics? Like why can't you as like a high schooler, you know, work with an academic, work with a researcher um to do something that hasn't been done before, right? Like that's cool, you know? Um if you meet journal requirements and have enough funds, you can publish them. Phrenology, I'm not going to do phrenology. I don't even really know what phrenology is to be honest. Um Professor Dave explains he made good tutorials and debunks. He's well known for being extremely spicy toward trolls and conspiracy theorist. Um For example, watch his debate with James Tour. Uh you can only reason with someone who understands science as a process of learning. Uh collaborative nature of your style of continuous ramble is great uh to overwhelm opponents. — I recommend Professor Dave uh or uh Minute Man. Uh — Your style of continuous ramble is great to do overwhelm the opponent. I didn't realize I was using a Pokémon move here. Like I didn't realize that this is like uh super effective against certain people. Like In an actual conversation, I do my best to listen, you know? Like I'm not just like talking all the time. Um phrenology was a racist science about uh skull size and intelligence between races. Well, if you do skull size as a measurement of intelligence, then like you're basically saying that women cannot be intelligent as men because their heads are smaller. What a stupid thing to say. You'd also be saying that koalas are smarter than I don't know um uh elephants, you know? Because koalas Oh, no. Koalas are smarter than uh dogs. Because koalas have a bigger head and a bigger brain. But it's obviously not true. Is that true? W- I mean, squirrels. Then koalas are smarter than squirrels. Something like mice. Mice are very smart um and koalas are clearly less intelligent on average than mice. Um and you know, obviously koala skull sizes are big, you know, compared to a mouse skull. Like that's how it is. Um Uh research is fun and but challenging. If you lack data, no matter how good you are, you have no use to academia. Um yeah, I mean, skills are very useful. All research needs data, though. Though yeah, I mean, like you know, I do theoretical physics. That's kind of what I do. Potatoes are more developed than humans because they have two more chromosomes. — Might be true. I wouldn't know. I know nothing about potatoes. But if you're a potato researcher, I would love to talk to you. Do we have a potato researcher in chat? Um I would love to talk to a potato researcher. What is the interesting research in potatoes right now? What does that mean? — Guys, I rambled for too long. This has been an hour-long ramble session. I had kind of promised myself that I wouldn't do ramble sessions. The new me, the new Leo's Labs would not do rambles. And yet here I am rambling. Um Uh okay, one last thing I need to say because I'm on YouTube, might as well. Um let me go to dashboard. Uh I am obviously losing uh losing subscribers. That's fine. I don't mind. Uh but one thing I want to point out is the following. Uh youtube. com/rambleleos. I'm thinking about just like Uh can I find it? Here it is. I had this second channel called Leos where I was trying to put like regular content, um but I just didn't do it in at the end of the day because I just didn't do it. It's not I don't know. I would rather do like live streams and clip the live streams and I would then make like individual videos every day. Um these are very easy videos to do. These are like Charlie style videos. You know what I mean? Um but uh I think I'm just going to take this profile picture and use it on everything cuz I like that profile picture. This is a completely different ramble. It's not related to anything else. But I just thought it was a really funny and good profile picture. And um I was thinking about just changing the channel from Leo's Labs just to Leos. Just have everything be Leos. Change this channel to be Ramble Leos. So if I still want secondary content, I can put it there. And then um just kind of moving everything to be more about like me as a person and not about like um — [clears throat] — like not be like Leo's Labs. Like I'm going to you know, I'm trying to hire people or something like that. I want to just be Leos. Like it's just this channel is my channel. And if I'm going to do other research with other people, then I want them to be on their channel, you know? Um that's kind of how I want it to be. I'm just letting people know that. I enjoy the rambles. It makes the stream feel familiar to me and I enjoy listening and partaking in the craziness. Well, apparently I do too many rambles. So um what is QKB compression for mitigating KB crash, KB cache, and network bottlenecks in LLM inference? Um I'll be honest. I have a blind spot. I have a big blind spot when it comes to LLMs. I know a bit about like reinforcement learning and different uh like, you know AI setups and stuff like this. But when it comes to large language models, I just I'm so incredibly uninterested in them that I haven't spent the time understanding them, right? Which is a problem, right? Like I usually don't have these blind spots. Usually like when I see a new area of of research, I'm like quite interested in it. But specifically when it came to LLMs, I was so bored with the topic the moment they came out. Um and I and nowadays like everyone focuses on them and it it's crazy. And so I feel like I should at least have a better understanding of it. Um and I don't know what to do to get that better understanding. I thought about maybe

working on a paper that is related to you know, because there was this paper that came out a while ago about the efficacy of different programming languages when it came comes to like LLM code gen. And they found that Julia was surprisingly good for LLM code generation, right? It's much better than Python or C or all these other languages. Um and it's the question was why, right? And the question is that because Julia has better abstractions or is it because Julia is a new newer language and there's less bad code to learn from? I don't know. Um but maybe that's a worthwhile paper to look into. The problem is it it's a very difficult problem to discuss because you have to like find a way to qualify what an abstraction is. You have to find a way to qualify like what like I don't know. There's a lot of things here. Like oh, I hit the microphone again. Sorry about that. Um there's also questions as to like Sorry. How do you make a good test suite for um like AI agents and stuff. I don't know. There's a lot of things that I could maybe be interested in. But the problem with all of these are like it's using AI to generate code and looking at the code results and it's just it's such a boring workflow that I don't know if I could get into it, you know? Um It's hard getting interested in LLMs with all the hype circle cycles, right? Have you tried writing your own small neural network? Perhaps doing that would give you a baseline to build off of. The thing is I've already done this. We actually have it in It was one of my first projects on Twitch. You might have actually been here when I did it. It was called BIAS, the uh biologically inspired architecture simulator. It was an analog electronic neural network um is what it was. Um It was nice. It was fun. I enjoyed doing it. Um but uh you know, that that's how I learned neural network stuff and basically we implemented a heavy and learning circuit in analog is what we did. Um and then we simulated it in spice. Um I'm not saying that that's like the you know, like I don't know. I don't know what I'm trying to say. Like it's just I went into OIST thinking that was going to be the thing that I was most interested in, right? Um I genuinely thought I was going to be the most interested in uh like this intersection between neurology and uh computer science, right? Uh creating biologically inspired neurologically inspired circuits and stuff like that. See was the thing I went into OIST wanting to do. And I left doing quantum science. Um That's just what happened, you know? Um and I still wonder if maybe I should have done the neurological stuff because now AI is so hype. Um but I also recognize I would not have GPU experience. I'd be a completely different person if I did that. Um so I don't know. Attention the key ingredient in LLM architecture. Uh also got the de facto standard in DL vision. Uh it's also very useful in times and time series um and signal processing stuff. Wait, wait. Deep learning vision. Uh yeah. I need to learn more about this, right? So the thing is when you implement a neural network, I've implemented like a basic neural network, not deep learning stuff. Um and so that that's kind of where my knowledge stops. And I should probably learn more. Definitely, 100%. And so I have all these discussions about AI and all this kind of stuff, but the truth is I don't know. I mean like I'm maybe I'm an idiot and maybe I should be doing it, you know, more of this kind of stuff um to have more of like a an understanding of this kind of stuff. But the thing is like if I were to make do a research project related to AI, the problem is I also want to make content related to that thing. And I just I don't want to make content related to AI, guys. I just don't. Like every time I bring up AI, it turns into a 1-hour rant of people like having all their opinions and you know, it's good to have their opinions voiced, but it's just I don't It's not interesting. I'd rather have the discussions about like I don't know, computer graphics methods or something, you know? Like that to me is more important. I was also not very interested in LLMs, but there's a lot of interesting stuff going on there that's useful all around. I agree. 100%. But I feel your sentiment. Like I agree because like I said, everything outside of LLMs I find interesting, but LLMs are not to me. That like, you know, everything with computer vision and stuff is useful. Okay. Guys, is he calling himself Professor Bart Kay when he's not a professor? You can't call yourself professor if you're not a Anyway. Anyway, let's keep going. Pioneering quantum computing. We talked about this as well. Guys, guys, guys. I have a problem. The Twitch title for this stream says research stream. Might even have a period at the end. I can't tell because the UI is so small up there on that screen. Um I've done no research. I've done nothing today but ramble. Um that is a problem. — That's the problem. They I feel like they weren't even good rambles. They were just normal rambles. Um let me at least open up the paper. And let me just kind of like I don't know. Look into it and stuff. Um next thing you'll tell me is Dr. Phil is not a real doctor. — You're right. You got me. I'd be more interested in those biochips that Australians developed. What biochips are we talking about? Um actually, I had an idea of contacting Maya. Um she is a uh she recently opened up like some sort of institute uh for like I don't know. Uh it is like animal rescue institute. Um she's a Twitch streamer. But

potentially there is there was a collab there was something that she mentioned in terms of tracking animals that I found kind of interesting and potentially there's a research project there that might be an interesting collaboration um — [clears throat] — that I thought would be interesting. I don't know. I would just as you know, like I there's opportunities that could open up, you know, with more people. Uh but I'm not there yet. Back in 1998, I wanted to write a natural language processor. I was 13 and without internet. Uh I banged my head against it for 3 years. I learned it was a full-on research field and I got depressed about it. I've done that as well. I think everybody has done that. Who has not done that, you know? Like um copy link. Uh yeah, I'm positive we did that. Um My reason wanting it uh My reason for wanting it was wanting to make a better Zorg. What is a Zorg? Actual intelligence think beyond silicon. Uh join the new era of computing and deploy your technology directly to real neuron. Huh? No, no, no. What? What is this? What? Join the new era of computing and deploy your technology directly to real neurons locally on the CL1 or distributed on the cortical cloud. What? No, no. Take a step What? A novel protocol for the efficient generation of all three major hippocampal neural subpopulations from human uh pluripotent stem cells. So they're building they're literally growing a brain. And you compute. How do they get the data out? A quantum quantifiable information processing hierarchy provides a necessary condition for detecting agency. Uh I want to know a computational perspective on neuro AI and synthetic biological intelligence. I want to know how they're reading the data. Can What can brain cells on a microchip tell us about intelligence? Intersection between the biological and digital, synthetic biological intelligence and organoid intelligence, active inference and behavior and intentional behavior. Where is the paper that shows me their um No, no. This is all stuff I don't care about. Uh organoid intelligence, uh biological neurons develop uh compete with deep and reinforcement learning and sample Aha, that's what I want. Perfect. Uh 2024. Ooh, what a new paper we're reading here, guys. All right, let me open this bad boy up. Where did it open up? I don't know. I think I'll never find it. Where did it go? Oh, here. — I want to look at this. I'm sorry. Where was this sent for publication? Where was this published? Let me do a quick forward citation here. 2024 says must be cited um somewhere. So let's go to scholar. And let me find the journal here. Um cited by nine. So not very popular and it doesn't look like it was actually published anywhere. That does not bode well for the efficacy of this research. Does not bode well. I'm just saying. Nine citations and uh no actual publication as of 2024 means they couldn't publish anywhere. Um and this to me seemed like their seminal paper, like the one that people should care about. Because we go over here to um my internet browser and we look at the rest of these papers that they have. Um Where is it? Where does it come up? Cortical here. Uh this is the only one I found where it's doing a direct comparison between uh like actual technology, right? So that's the only one that I found that indicates that they can actually do some form of computation um on this this uh architecture. Biological neurons compete with Okay, so how do biological systems and machine learning algorithms compare in the number of samples required to show significant improvements in completing a task. Um we compared a learning effi- efficiency of in vitro biological neural networks um to the state-of-the-art deep learning reinforcement uh Why are they using reinforcement learning algorithms? Oh, deep reinforcement learning. I was stupid. Um in a simplified simulation of the game Pong using dish brain. What? They call it dish brain? What is this? They're playing Pong is what they're playing. This is just a This not a paper, first of all. I see why it was only an archive submission, and it looks like they're playing Pong, which is fine. Oh, maybe I'm dumb. Maybe it just didn't load the rest of the paper. Um Apparently, also run Doom on it already. Human brain cells can play Pong in a Petri dish. Cool. I mean, that's nice. Attach electrodes and shock the brain cells into submission until they self-organize. Self-organize makes me think of like uh What's it called? The um You got a company, and then the company you have the workers organize together into a What? Uh not a cult. What is it called? The workers organize into a something, and then they attack their boss. I thought William has a cool video using brain cells to play Doom. Really? I

mean, cool. Um it's less of cloud neurons and more of remote incubator. Um I just don't understand how you get the data out here. Union. That's the word I'm looking for. Yeah, when you say uh they revolt, I'm thinking of like a union type of thing. I just don't understand how they Maybe I need to look at this dish brain thing. Uh planning system that integrates roughly 800,000 living human cells. Images? No. This isn't what I'm looking for. Um Here, dish brain plays Pong. Biological intelligence. Um Six citations. Ah, [sighs] I don't know. This is what I'm looking for, guys. This is not what I'm looking for. I just want the paper that tells me that like this is how this architecture works and how you get data from it. Like that's what I want to know cuz I My understanding is that like, you know, like neurons don't have a clear It's not clear how learning works in the brain. So, I don't understand how they're learning in this system. Like it's not clear like what the hippocampus is for. I mean, it is. It's like for long-term memory retention. We know that objectively, but we don't know exactly how it interacts with like the other layers of the brain, you know? And so, it's I don't understand, you know, ask Grok. I could, maybe. Um I could. Okay, we'll talk about this later. Is Grok good? Uh you know, when I was mentioning the AI architectures, I didn't I'm didn't even think about Grok. Um but I heard Grok is actually useful for some things. All right. See projects, papers, SIGGRAPH. Oh my goodness. Uh let me do uh ocular_main. pdf. Oh, I have to say something, too, don't I? Um dish brain paper gives a link to cell. com might have what you want. Yeah, I should have probably just Googled that, shouldn't I have? Uh dish brain paper. Uh in vitro neurons learn and exhibit sentience. Uh anytime you put sentience in a title in a simulated game world, uh integrating neurons into digital systems may enable performance infeasible with silicon alone. Here, we develop dish brain, a system that harnesses the inherent adaptive computation of neurons in uh structured environment. Um So, just a bunch of questions I have. Like first of all, neurons are slower than electricity. That's number one. So, if we could, you know, more accurately This is why I want to do the bio I still biologically inspired architecture simulator. Can I just go back? back and look at bias? Where was it? It's such stupid code. I wrote it in C, man. See, why do I suck? Bias, a biologically inspired architecture simulator. When did I write this? I wrote this It was some stupid project. Um Uh this project follows from the development of IBM's TrueNorth chip and Compass simulator for the TrueNorth chip. The idea is relatively simple, to build an analog cognitive computer with some, if not all, of the bells and whistles of TrueNorth. So, like in my opinion, if you figure out how brains work, then we don't have to build a new brain, you can build an electronic brain. But we don't know how brains work, so we use matrix multiplication for that. Right, that's what I'm talking about. Um basically, there's a reward signal to neurons every time the neuron did something good, they receive a reward. The input output is voltage to specific place. I see. My question is the following. The plasticity of neurons, the way in which neurons move around, is very important for the way in which learning happens, right? Neurons, like they literally move around and form the networks that way, right? Like that's what they're doing, right? And so, if you get a brain to learn something, right, via these electric shocks that you're doing, um I do believe it is possible to do that, right? Um but the thing that I'm confused about is that first of all, that feels different than how we like, you know, obviously, there's like brain waves in your system that you can read via EEGs and stuff like that, and like it's a collective pattern of a bunch of neurons, and that's how you do um error correction because you have a bunch of neurons all signaling at the same time, or not same time, but in a similar fashion that allows, because there's so many neurons firing at the same time, it allows for you to do um some form of uh error correction in your head before like things come out into the world. Um So, like it's a collective action of a bunch of neurons uh through these, you know, somewhat wave-like pulses that allow for you to enter certain states collect um data back from the brain in regular cognition, right? And so, my question is, if you've taught a brain to learn something in some way, is it only able to exhibit that behavior forever, right? Like how do you reprogram it, I guess, is my question. Like if you program it to play Pong, do you now need a new brain to play Doom, you know? I guess that's my question. Because like the way these networks work, it becomes, you know, you have this plasticity, the neurons moving around, and then once it forms that network, it's kind of a little bit hard

to rearrange that network, you know what I mean? So, I'm maybe stupid. Um uh some random YouTuber uh basically did the same uh with a rat brain. It is relatively simple. Uh you do reinforcement learning by going giving it endorphin. Yeah, that makes sense. Uh basically, there's a reward signal. Uh the neuron is aligning by itself. There's no direction given. Also, neurons are like themselves very undifferentiated. Right, exactly. But what I'm saying is that like the You might be right. Maybe I am underselling the plasticity of the neurons. I just know that plasticity uh goes down with age. Like as you get older, um the neurons stop moving as much, right? Um that's just, you know, younger kids have more plastic brains, you know? Uh but I guess if you always have young cells, then you don't have to worry about that. Uh learning phase during sleep. I mean, there's multiple learning phases that happen, um and certainly some of that happens during your like sleep, during the REM cycles in sleep. Um but uh you know, I learning here could mean a bunch of different things. It could mean like what we're talking about for um like machine learning learning, it could be human like learning something for a particular task, it could be something in between. So, it's a loaded term here. Um maybe you're confusing fully formed and structured brain versus clumps of rat neuron stem cells. Yeah, I think that's exactly what I'm doing. I think I'm trying to get them to replicate the brain, but all they're doing is getting neurons in a pot, right? That's the difference, I think. Um Uh make two brains, one for Pong, one for something similar, but different enough. Then try to figure out how to get from one to the other, uh and potentially how to mold move the brain to do that. Uh motivate Kids have millions and millions of neurons as they get pruned as they age. Make bigger brain uh that has the capacity to learn. We should do that for knowledge stream. I think you've uh pointed me in the right direction. I have like seven neurons. That is seven more than I have. I'm pretty sure I'm actually just a robot. Um okay. So, uh interesting. It's interesting. Thank you guys for showing me this. Um now, ah Can I just be honest? I walked in to this stream anticipating that I would get work done. I was feeling a little bit sick. I thought to myself I thought to myself, "It's okay. I will get over it. " Uh but it's okay. The stream title is like I don't know what I'm doing. So, it's fine. It's fine to have a random one-off stream where we just kind of chat about things. It's fine. It's fine. It's fine. It's okay. Uh let me Dag nab it. Um One thing I want to do is all right, but this is I want to take a look at this AMD mesh papers. Mesh shaders paper um because there's a follow-up paper to this that was published a little bit later, um and I want to take a look at it. So, real-time procedural generation GPU work graph. So, as of yesterday's stream, um we were talking about a lot of related projects to the one that I was doing, right? There are a lot of related projects. And I was thinking to myself if I was wondering if I should retheme the entire paper so that it's more related to data compression and less about the function systems. But the fact is that this paper is proposing multiple novel things. None of them are like super novel, but they're interesting in their own way. Um this is the first time the paper that I'm proposing that um you truly have general-purpose rendering with function systems. So, that's number one. I think that's what I should focus on. And if I want to do another paper on the data compression aspect of it, I should do another paper on that along with more compiler uh stuff that I do to make it more efficient and stuff. So, that that's going to be a future paper. Um but I do want to very quickly look at this on Scholar, and I want to look at future uh cited by. Here we go. And I'm going to go from here. Real-time meshlet decompression, I create GPU kernels for expression tree generated runtime, random art. Actually, expression trees might be what I Oh, but expression Never mind. It's useful, but not what I'm looking for. Uh real-time procedural surfaces using GPU mesh shaders, GPU work graphs. The thing about it is there's so many people talking about these GPU work graphs. I feel like maybe we should be doing this, right? Like I'm not implementing a compiler. That's not what I'm doing. I'm restructuring my code in such a way that I can put I can just throw it at the OpenCL compiler. I'm just I'm trans uh I'm transpiling it into an OpenCL kernel, I think, is the correct way to put this. And then I'm letting the OpenCL kernel do the compilation, right? The thing about OpenCL is it allows for you to do the compilation when you want and that's what I need. I need the user to specify when the compilation needs to happen, right? And so the discussion of the GPU work graphs, I think is very valuable, but if I start going down that route, all of a sudden I have to write my own compiler and I don't want to do that. I I I've written compiler-esque stuff before in the past and it is a pain in the butt and you have to deal with so many edge cases and there's no guarantee it's going to be faster than what's already out there. It's just it's a pain. Um Well, since you some less than half of what I'd hope for. — Uh if every viewer has a neuron has neurons donate some neurons to Leo's you'll have half a neuron at the end. Avocado Violet, hey Avocado Violet, thank you so much for the raid. Um I'm going to do what I always do and ask how was your stream? I hope everything's going okay, but I also recognize that

myself, I don't actually like it when I um uh when people talk to me after like raiding and stuff like that. So you can choose to respond, you don't have to respond, but thank you so much for the raid. I really do appreciate it. Um I really genuinely do. Okay, so we were looking at this real-time procedure. I'm going to open up the PDF cuz that's I just you know, I want to open up the PDF. Let's take a look. I have free time. Okay. So and then the next follow-up paper, the one that I have to find I think I put into the paper pile. So I think I'm looking at this in a dumb way. I think if I just go to the paper pile, I'll find it. So I'm going over to Discord. Where are we? Where is Discord? Discord updates E Um where is the paper pile? There's a paper pile. I believe it is here. Let me see. Let's see. Let's see. One second. Paste here. And let's see what we've got. Is it going to work or am I stupid? Both could be true. Okay. No, this is massive parallel rendering complex closed-form implicit surfaces. Not the one I was looking for. I was looking for another one. Uh and I don't think I can find it. That's okay. Let's get the forward citations. Um oh, maybe I can. Let me go to my channel dashboard. comments. Let me go to uh responded. That's exactly what I want. And then I want to go do do do do do here. Uh and then I want to go to Where is it? Uber shader massive parallel rendering of complex closed-form implicit surfaces. Is that what I'm looking for? There should have been a follow-up paper here. Who sent me the follow-up paper? Someone paper. I can't find it. Okay, I I'm not going to docs people. Okay, regardless, let's take a look at this paper anyway. It's fine. Okay. So this is the paper that was published uh last year or two year and a half ago. Real-time procedural generation with GPU work graphs. So um the reason this is interesting is because uh what they're focusing on here is basically I mean, we talked about it yesterday like data minimization in a sense, right? They're trying to make it so um the rendering of meshes the drawing of meshes is done on a minimal set uh of GPU data. Like we don't have that many mesh points that we're storing in memory, right? Um so we present a system for real-time proce- procedural generation that makes use of novel GPU programming model work graphs. Uh the nodes of a work graph are shaders which dynamically generate new workloads for connected nodes, right? So this thing is actually quite similar to what I'm doing um in that like we allow for users to more easily create uh their own like basically is a meta-programming aspect. Um this greatly simplifies the implementation of recursive procedural algorithms on the GPU um combined with GPU ray tracing and procedural mesh shaders um our system makes use of this graph structure to tackle various common problems with procedural generation. Our system is very easy to implement requiring no additional data structures from what would already be available uh in a modern rendering engine. We demonstrate the real-time editing capabilities um on representative examples. We augment the scene with a teaser image and the teaser image with 79,710 instances in 3. 17 milliseconds on an AMD Radeon. That's actually quite slow. That's quite slow. We augment the scene in a teaser image with 79,710 instances with 3. in 3. 74 milliseconds. Am I misreading this? What's the FPS here? Am I stupid? I feel like I'm stupid. I'm just not mathing correctly. 1 divided by um 0. 375 That means they're only getting 250 FPS. That's not great. Um like I mean, it's not bad, but like you know, if you're if this is your demo and this is what you're showing, right? Don't get me wrong. It's a nice picture and actually if you look at the scene here, they've got shadows and stuff like that. So yeah, I mean, I'm not saying that that's bad. I'm just saying that like I was expecting less time to generate this is what I was expecting. Because I know with my specific image or the stuff that I'm doing, I can generally get um within the thousands of frames per second with the method that I'm doing and it's in a very similar field. Um I know I'm not generating in real time, but each kernel takes about uh 1/1000 of a second to run. Uh so I know I can get that with my render rendering method. My rendering method is not efficient at this point. Um so I know I can do better than that. Uh so with that in mind, I was expecting this especially coming from AMD to be much more um how do you put this? Much more I not better. I don't know what I'm looking for, right? Uh what's a procedural mesh shader? Can you come up with new objects on the fly? Yeah, that's exactly what a mesh shader is. Um So a mesh shader So here's the thing. If you want to draw with math, you want to do math-based stuff, right? Um and you don't want a mesh, what do you do? Right? Well, you've got basically one method that you can do. Well, two methods. You can do ray marching. Sorry, tracing uh which you know, you'll just draw your objects out in some form of closed-form object and you can do ray tracing. And in general like if you're doing spheres or squares or something, you can do that with just ray tracing and you don't need a mesh for that. Although you probably want However, there's the ability to do ray marching uh which is slightly different algorithm, but still involves taking rays from your camera and shooting them

onto the scene um and then uh you can model your objects in a mathematical space known as the signed distance field, right? With signed distance function. And by modeling your objects in this way, it allows for you to march through the scene and draw them on screen uh without having to use a mesh, right? And so this allows for you to do meshless rendering, right? And so you know, if you're looking for ways in order to draw things with math, which I was doing a couple years ago, you might think well, this is it, right? Um and that [clears throat] generally is it for the literature right now. My paper proposes a new strategy, but one way in which you can draw with math is actually creating mesh shaders. So in the GPU space, you've got fragment shaders, you've got vertex shaders and a whole bunch of different types of shaders, computational shaders and stuff like this that all fit into this graphical workflow. Vertex shaders move vertices in a mesh. Um uh Fragment shaders shade the fragments, that is like the little uh the little bits of the mesh that are eventually put onto screen uh and um then the other do different things. Like compute shaders are basically computational kernels. Uh but there's a new shader out there as of 2022, I believe. Maybe actually maybe I'm wrong. Maybe it was earlier than that, but I remember the papers from 2022 um called a mesh shader. The mesh shader allows for you to generate the mesh on the fly. Um it's in a weird it's a weird shader. It's hard to use um and doesn't allow you to do like really topologically non-trivial objects as well as you would like. Um but it works. It's okay. It's fine, right? Um geometry shader allowed you to do similar things, but geometry shaders nowadays are more or less um they're more or less uh deprecated. People just don't use them as much as they used to. Um So — [clears throat] — Also, uh let me go ahead and finish up. There's some comments here that I missed. Uh nice suggestion. Hope you're doing hopeful hope you're doing debunks. You have the potential for it. I just think the debunks just to finish up the conversation we were having earlier. Like I feel like I could change the way in which I do stream to be someone who debunks, you know, ideas. Like I I said it before and like I kind of like the idea of going through pseudoscience and just like figuring out what's wrong with it and then uh and then just talking about it and like explaining at a deep level why it's wrong. But I just don't feel like that's me. I It's unnecessarily negative and you know, I don't see pseudoscience as bad exactly. Like my opinion about pseudoscience is mixed. Um I know I've been rambling all day. We'll just ramble a bit more. I don't actually have a negative opinion about pseudoscientists, right? Because at the very least a pseudoscientist is actively engaging with the scientific process, albeit coming to conclusions that I don't respect. Um but that is better than someone who just blindly follows the science um because you know, or blindly follows the pseudoscience, right? Because there are people who follow both. But at least the person who is actively trying to understand the world is actually world, right? Um even if they're promoting something that is not correct at the end, they're trying. And that means that you can transform a pseudoscientist into a scientist. I mean, look at me. When I started doing research as an undergraduate um not ignoring the stuff I was doing in high school uh and like I started actually trying to do research for real, I was very interested in this concept called the microtubule of consciousness. It is definitely pseudoscience. It's completely wrong. It It's It was a method to understand how consciousness works by um understanding brain waves and stuff like this. And there's plenty of ways in which we can um prove that it is incorrect. Right? Or maybe not prove, but we can point to it being incorrect. Um but it had a lot of Nobel Prize winners. Not a lot, but there was one Nobel Prize winner who was very into it um and with that in mind, it kind of felt like, you know, if this Nobel Prize winner understands it, then probably it's okay and you know, it just a lot of reasons why I wanted to get into it when I was younger. Um And from that I learned a lot. The ability to go through a pseudoscientific topic and explicitly understand why it is pseudoscience is very valuable um and I you know, I think as long as you're not constantly going down the rabbit hole, I don't have a problem with it as much as other people do. Like I don't think it's as bad as people say it is, you know? Um Similarly, I don't think the replication crisis is as bad as people think it is. I think it's fine. Like, you know, the way it is right now, yeah, it kind of sucks that we can't replicate papers to a certain extent. But on the other hand, if I'm being honest, replication happens on citation. It happens on people using that work for future work, you know? And so, as long as your paper is well-cited, it's not as big of a deal as people are saying it is. Is pseudoscience necessarily wrong or not rigorous enough to prove true? That depends. I would say pseudoscience is necessarily wrong, whereas uh you know, something like string theory is just not proven to be true, right? Like, I would not say string theory is pseudoscience, whereas I would say flat Earth is pseudoscience. You know what I mean? But string theory does not have enough is not rigorous enough to be proven true, even though it is quite rigorous. There's just no way to experimentally validate it. Um so, yeah.

Anyway, that was my rant on pseudoscience and why I at least right now I don't necessarily want to do debunks. I don't want to like channel the energy we have into negativity. And if I have x amount of energy per day, then I don't want to spend that energy on you know, proving people wrong. I'd rather myself to be right. You know, so you have to prove it false to prove it pseudoscience. I think I think there is part of that um to be true. Yeah, I mean, pseudoscience is something that goes against the grain. Um yeah, that you know, the grain being what scientists generally believe to be true. Um Okay, let's keep going. Knots and strings fascinated cats and humans alike. That That's a fair point. Okay, let's keep going. Um Okay, so creating highly detailed vir- I'm so sorry. Creating highly detailed virtual worlds manually is a labor-intensive task. Procedural design tools like Blender or Houdini accelerate the process. Um the generated geometry, however, is typically baked into the polygon polygonal format. Right. So, data duplication for rendering, all baked data has to be stored on disk. Tool disparity, scene statis- uh staticity. Sta- staticity. Anyway, uh point is like they have a whole bunch of good reasons why you want to mathematically generate things. I don't think we have to go too deeply into this um to know that they tried their best, they did a pretty good uh job. My question is I I've got a couple questions. So, the first of which is I know there's someone that used this in the future. I know there's a forward citation here. Where are we? Um real-time CPU GPU tree generation. It's probably this one. It's probably the one I'm looking for. Uh let me go to the PDF here. There we are. Enter. Open this bad boy up. Real-time GPU tree generation. Yes, this is the paper I'm thinking I'm looking for. So, here's the thing. When it comes to mathematical art, I don't know why, but all of the methods to render mathematical art um generally tend to draw plants. I think I know why, because plants don't have to look exactly correct. Uh they can look good enough. Um and they're very easy to procedurally generate cuz you can follow like very uh clear patterns in order to draw a tree. Uh and because it's not like a human face, it means that you can um uh it doesn't it fall into that uncanny valley, right? Uh but I do want to take a look at this. So, trees for real-time media are typically created using procedural algorithms and then baked into a polygonal format, uh requiring large amounts of memory. We propose a novel procedural oral system and model for generating and rendering realistic trees and similar vegetation specifically tailored to run in real time on GPUs. By using GPU uh work graphs with mesh nodes, we render gigabytes worth of tree geometry uh from kilobytes of generation code every frame exclusively on the GPU. Contrary to prior work, our method combines instant in-engine uh artist authoring, continuous frame-specific level of detail and tessellations, highly detailed animation, and seasonal details like blossoms, fruits, and snow, generating the unique tree geometries of our teaser test image uh and rendering them uh in the G-buffer takes 3. 13 milliseconds. Getting a little bit slow on Radeon 7900 XT. So, couple things here. I don't care about like this looking like fall or winter or spring. I don't think that really matters. Um the fact is this doesn't look great. Like, it's fine, but am I wrong in saying that for a 2025 paper, this is not how I'd want trees to look? Like, the winter trees look fine, but you know, the summer trees and the fall trees, like they're specifically going at a distance cuz otherwise it doesn't look as pretty. Um they just don't look as realistic as I would like one of these things to look, right? Um going to head out, fetch a package and perhaps even a pizza. See you, James. Was enjoyable. Good luck brain- braining the brain words. Yeah, see you. Thank you so much for popping in. I know I want pizzas soon, too. Have a good one. Yeah, thank you guys so much for stopping by. Like, seriously, I really do mean it. Thank you guys so much. I really do appreciate it. Um okay. So, the reason I brought this up is because I think I do want to cite this paper as well. Um 3. 13 milliseconds. This has me wondering Okay, let me write down the things that I know I need to do for this paper. So, let's go to Zernov PP. Okay, let's just if nothing else happens at the end of the day, let me write down a list of things I have to get done. Um what do you mean unrealistic? Those are great-looking trees. They're pretty good-looking trees. Like, they're pretty good-looking trees, but like I mean, like maybe I'm wrong. Maybe I just haven't played a video game in a long time and I don't remember what trees look like in video games. But here, let me write down a list of things I definitely need to get done. So, one, I have to get down Why is it red? — Okay, what am I doing? FFS Greedy, thank you so much for following. Really appreciate it. What am I doing with my life? So, this paper needs a couple things. One, I need to re- for- structure. Cool. Two. Um

I need to uh figure out how many uh bytes are actually in the space image. Uh the space image being this image in here. I suspect it's like five, but I think it's more cuz I there must be some allocated in the function system that I'm not paying attention to. Um okay. So, that's what I have to do. These are easy to do. Now, stuff that's hard. Uh hard, I want to do a signed distance field uh ray marching. Uh plus uh my composable function system stuff. So, basically, I want to render something with some other method and then use my method to generate a smear frame uh for smear frames. Cool. Easy peasy. Uh number four, uh I want to do some form of uh meta programming. And I think I what I want to do is lip sync. Cool. Easy peasy. Um not easy peasy, it'll take some time. So, these are all things that I want to do. Um and I definitely need to do before the paper is out. Now, when it comes to this kind of stuff, the other thing I mentioned yesterday is rigging. Um that is we take a hand and we just show how to move it around. This should be easy enough to do with a hand. Okay, so that's three things we got to do. Uh four things we have to do. Okay, now for the uh the other stuff. So, this is the um what's it called? Like, the stretch goals. Goals? Okay, so number one, real time. Now, here's the thing. I know that this code should work in real time. Um however, it doesn't work in real time because currently I'm using OpenCL. Right? OpenCL is a compute framework. Uh to pass that over to a graphics framework like Vulcan is not super hard. I just haven't done it yet. Um the other thing is I actually wanted to do a separate paper on the real-time rendering of this because the way this works is I generate a point cloud, right? And if I want to generate this point cloud in 3D, well, I'm just doing simple transformations and so it makes a lot of points. Like, let's say you have an object, like the front of the object uh and the back of the object will both be drawn. But uh what will end up happening is I'll overdraw the number of points cuz, you know, in theory, I don't need these points in the back, right? I could get rid of those points in the back. Um but I don't know how to get rid of those points in the back right now. I mean, I can if I just do the right transformation, but finding that right transformation is kind of a pain. Um so, um what I want to say is that I wanted to do a paper on real time afterwards. And the reason I want to do this is because I also wanted to combine this with um some auto uh diff stuff. So, basically, the way this works is you draw a circle and then you add additional transformations to that circle to draw whatever else you want. But in practice, what you could do is since you know these transformations, you could do some form of automatic code generation uh to make it so uh you can create this initial distribution of the circle that has the exact number of points that you need so that each point will land in a pixel at the end. Does that make sense? Like, if you take a circle and you stretch it like that, well, that means you need 2x uh in the x direction the number of points as you have in the y direction. You see what I'm saying? And if you warp it like this, you know, you or you wiggle it like this, like there are different things that you need to know. Um uh we should in principle know what that initial distribution should like should look like. And so, the act of creating a generator um to minimize the run time should totally be possible. So, I think the reason I'm bringing this up is because this paper does very much mention like the 3. uh 3. 13 milliseconds and all this kind of stuff. And I think that's very valuable information to have. But I also feel that like if I'm going to do any type of paper talking about real-time rendering, I should definitely mix it with the auto diff stuff. 100% should So, I don't think that's something I'm going to do right now. I think this is going to be for later. That's going to be another paper. And that's totally fine to leave as another paper, I think, personally. Um I might get laughed at Siggraph because it's not real time, but I don't think so. I think it's okay. Um Obviously, there's other stuff I have on the the uh the list of things to do. Uh one of those is quantum. I think there's a good quantum formulation that I want to do in the future. And then, what else is there? Uh a bunch of other papers. Oh, and a good Gaussian splatting example. Okay, so but I think really all of these are going to be uh because this has to do with digital twinning as well. So, I think all of these, these additional things, uh, these are just going to be extra papers. I could definitely see these as different papers. I think for this paper I just have to show these things. Um, this one shows its interoperability with different methods. This talk talks about its uh, possibility for data compression. Um, this talks about uh, Quibble as a metaprogram framework. And then this talks about just it it's just a possibility. I don't know if reking is actually that necessary, but considering that people care about it, I think it's worth bringing up. Okay. So, this helps me reformulate how I want to cite these papers, right? Um, in particular I just need to mention in the related work that such examples exist. Um, and they're useful. Um, but I think this is going to be like a future work type of thing. Um, and I should just definitely mention that in future work. Uh, Come on. Uh, if you got the differential then you can theoretically draw it as a Gaussian. Uh, what do you mean unrealistic? Uh

those are great looking trees. Yeah, maybe I'm maybe I'm being too harsh on the trees. Guys, I'm sorry. Meanwhile, [snorts] fluid dynamics are run at uh, minutes per frame. Yeah, this is exactly how it is. It's rare for you to find a real-time fluid dynamics simulation. And I'm coming from a computational science world where like nothing is running real time. So, I also I'm kind of from that perspective as well. The trees what the tree looks like is irrelevant. That Okay, you're right. You're I was being too harsh on the trees, guys. I'm sorry. trees. I understand that. Okay. So, let's go ahead and commit. Let's commit a bit, okay? Um, not I'm not talking about committing code. I didn't do much today. Today was a what am I doing type of day. That's fine. I want to say that tomorrow I'm going to do something important. So, which of these topics Mhm. Lazy evaluation for the win. That is actually very a very useful comment. Okay, so I have to do something. Right? I am going to do something. We have Wednesday, Thursday, and Friday left in this week. We're going to get one of these done, okay? Which one do you want to get done? Lip sync, uh, sign distance fields, um, uh, or hand motion with reking. Which one do you think is most relevant? I think probably probably we want to do ray marching, right? Like that's the most interesting, I think. Um, of all of these I think that's the most interesting. Um, Yeah, I think so. Um, so let's just kind of go into sign distance fields and and try to learn them next week or tomorrow. And that that'll be the goal for tomorrow. Um, uh, and then let me go ahead and put all these in like to do statements so I don't forget about them. So, let's go into vi main. pdf. Okay. Ah, you're kidding me. I opened up a PDF file. Tech. All right. Uh, and let me just go up here and just put a quick comment here, to do. Um, and then we're just going to paste all this in there. I did that didn't work, did it? No, that didn't work at all. Uh, this Bam, control C. Where is my terminal? Paste. There we go. Uh, and let's comment this out. So, I just I know it's here. I'm going to look at this and we're not going to worry about this. There we go. Done. Done. Open this. Okay. So, I think tomorrow we're going to dive into sign distance fields. We're going to do some form of a visualization with that. Um, I think — this guy is a VIP PDF tech. Imagine if I'm directly editing the PDF. I could, you know? I'm sure people have, but not me. Okay. So, then in that case, let me go ahead and read through the related work section as I have it now and kind of restructure that as it is. Um, Man, I Can I I'm going to do something really quick. I'm sorry. I shouldn't do this, but I'm going to. Uh, I got a message on Slack. I don't normally get messages on Slack, but I had to make sure that it's not something I have to Okay, it's okay. It is what it is. I'm not too worried about that. Good. That's fine. That is fine. All right. Was it this button? That button? Perfect. Okay. What are we doing next? Uh, I'm going to go through the related work section. That's what I'm going to do. E elated. Perfect. Um, Okay. Uh, related work. Okay. Uh, we have to talk about AMD mesh shaders and function systems. So, do some graph deep graph layout deep dive visualization at the moment using Vulcan and C and I graph. Pretty fun and scales to large graph. Actually really cool. Vulcan So, I assume you're using Vulcan C, right? Or you're using like uh, the Python interface or something like that. The thing is and I mean this genuinely the one main reason I have considered How do I put this? Vulcan is an area where I could consider AI code gen to be valuable. And yet I'm terrified of doing AI code gen for Vulcan. I'd love to support your journey and help you grow your audience. I can connect you to a popular Twitch broadcaster who streams to thousands and has nearly 1. 2 million followers. This is just Is it just spam, right? Uh, MoonMoon100. Uh, chat warning. Uh, just delete. It's fine. Uh, I want to see if it scales zero overhead. Yeah, I get you. But is it really that much overhead in like C++ Vulcan? It was a scam. That was definitely a scam, but I just I had to double check. Okay. So, I'm going to go ahead and read the related work as it is right now. And then we're going to talk about where we want to fit in these additional papers. So, iterated function systems and related approaches have been an active field of research uh, interest for decades with many methods showcasing their versatility and usability in various aspects. As such, composable function systems uh, build upon and extend Uh, I don't like this. I already don't like it. I think I need to modify every composable every CFS so it's actually a compute composable function system framework. Here's the thing. When I say composable function system, people are going to think it's a math thing. But this isn't a math thing as much as it's a

programming thing, right? And so I really should say I just don't know how to because it really is for composing different function systems. That's what we're doing. We're composing sign distance fields. We're composing uh, mesh generation. We're composing all this stuff together into one big thing. So, it's composition of function systems. That it it's a good name, but I think it should be framework. I think I should definitely say framework. So, I'm just going to go through all of these and I'm going to modify all the CFS to C F S F. Uh, so it's compose composable function system framework. And we're just going to do it. Um, yes. Uh, a simple CFS that C F S F. I Can someone give me a better name than CFSF? Like it it's a bad name. Like it's fine. Um, It's okay. — [clears throat] — It's fine. Am I dumb? No, I'm okay. — All right, let's keep going. Uh, elated work. Okay. So, let's go down. — [clears throat] — and extend uh, several other existing techniques to in the literature. Uh, do do. For example, in Kuns et al. the authors propose extending iterated function systems to a new technique they call the iterated multi-function system, which allows for greater flexibility in the creation of fractal-like objects. However, their approach is limited in terms of allowed mappings. The fractal flame algorithm um, is also similar in that it allows for users to inject functions into a pre-existing iterated function system structure. In contrast to both of these methods, this work focuses on building a metaprogramming framework to allow users to flexibly create specific methods for composing functional units and create general purpose non-fractal artwork. Several other approaches limit themselves to affine mapping where no such limitations exist in the work. Um, I think that is a weird sentence to put at that point in time. Um, uh, do do. Paste. In contrast to all of these methods — there are also methods to for creating for rendering that are uh, do not use uh, iterated function systems and instead use function systems systems. Let's see. There are other methods for rendering that do not necessarily use iterated function systems instead use function systems um, as a um, as a How do I put this? Function systems composition function system composition framework. Uh, yes. So, I have I think the composable here is more important to put at the start, right? Um, but yeah, we'll figure out the naming stuff later. So, there are methods for rendering that do not necessarily use iterated function systems instead use function systems as um, — [clears throat] — a series of transformations uh, on Hmm. as on object primitives um, to trans to transform pre-existing No, to and say use function systems more generally. Um, and I need to cite this guy. This guy here. There we go. That's fine. I'm okay saying it that way. It's okay. It's not the best sentence, but I don't care. Um, This work focuses on In contrast to all these methods this work focuses on building a metaprogramming framework to allow users to flexibly create specific methods for compos composing functional units and creating general purpose non-fractal artwork. Um, I think that's fair enough. I think it's good enough to say. Um, as a general purpose rendering framework, uh, composable function system frameworks for general purpose rendering. Uh, general purpose rendering uh, Composable function system frameworks must also contrast themselves to other methods such as meshes. Mesh mesh. Sell contrast itself? As themselves? I don't know. Uh, what do we have here? Hey, what's up, Al Chris? How are you doing? Um, would it make sense to just have the system framework so you could have composable function system Wait. Oh, wait wait. What do we have here? Composable function framework. It feels very general, doesn't it? But in principle, you're right in that I'm not necessarily using function systems all the time. Um, like the core utility that I'm providing here, like that hasn't been done before, uh, let alone just the idea, which is, I believe, interesting. It is kind of this thing. It's a better meta-programming framework that allows for people to, like, basically inject [clears throat] code in without using function without using inline functions. Like, in principle, that's all I needed to be able to do, and I couldn't do it in Julia, in a lot of other languages. For some reason

I had to write my own framework to do this, right? So, that's number one. Um, but this [clears throat] isn't in principle hard to do, and I'm not encouraging people to use my code as much as I'm encouraging them to, uh, just, you know, that this is a valuable tool for rendering. Like, it allows for you to do much more complex things that you could do otherwise do. Like, basically what I'm saying is it would be nice if you render your mesh onto one of these layers so that you get additional transformations off of that, right? Um, it No, it's So, the difference is you can use an L system in this compositional framework. Uh, and I show that here at the start. Uh, using iterative function system instead. So, the idea is that, like, you know, you will use the iterative function system at step one to generate the primitive, and then you can do additional transformations on top of that. Where the additional transformations could also be further iterative function systems, if you want. It doesn't matter. Um, you kind of just allow yourself to do whatever you want to do. It's just that, like, the goal is that, like, you generate that primitive in any way you want. It could be via iterative function systems, any other method you want, and then you do additional transformations on it. So, it's a combination of a bunch of different things. And the reason this is complicated is because when you're generating an iterative function system, You know what I'm trying to say? When you're solving an IFS, you have to do this usually do this via the chaos game. The chaos game calls functions at random, and because it's calling functions at random, it's very hard to inline those functions, right? And so, this is why you have to create a separate utility to do so, right? Um, and so, it's kind of like a combination of these two things. It's first doing the iterative function system. That's easy. Second is adding transformations, also easy to do, right? Um, that's what the old paper did that we were talking about yesterday, right? So, both of these are easy to do. It's just the combination of the two of them hadn't been done before. That That's all this is. I didn't expect In a sense, I didn't expect this to be a paper, but it turned out that I couldn't find a way to do this outside of doing, like, you know, I writing the paper. It took me a long time to figure this out. So, we go here, and let's just keep going. So, uh, ba ba bum. Where are we? Related work. Um, as a general-purpose rendering framework, uh, CFS composable function system frameworks must also contrast themselves to other methods such as meshes, ray tracing, and ray marching with signed distance fields. Um, in contrast to traditional mesh-based rendering approaches, this work does not necessitate the storing of any vertex points or textures in memory. It is capable of generating N-dimensional scenes with point clouds that are spot on the screen. Um, objects and object locations do not necessarily need to be stored in memory, either, and can instead be directly embedded into the function systems themselves. Um, if users wish to use mesh-based methods, it's possible to do so by generating the triangles and creating the mesh through software rendering approaches. Uh, however, the software implementation of, um, compositional function system frameworks, um, written for this work primarily relies on computational kernels acting as shaders, uh, and lacks certain hardware features like rasterization and ray tracing. That's all true. I think that's fair. Nobody says React framework, they just say React. Same with Next. js. You're right. Maybe I could just leave it CFS. Um, and just say it's a framework. That's fine, too. Um, that's fine, too. I It's just the reason I'm saying framework is because if I say FS, then people naturally assume it's a math thing. I just don't want them to think it's a math thing. That's it, because this isn't really a math thing. I don't talk about how these function systems like I don't want anyone to prove anything in here, right? Um, it is also possible for users to implement ray tracing ray marching methods as shaders to objects created with, uh, CFS as to be manipulated by, for example, non-affine transformations. This is particularly interesting in that it extends the capabilities of ray marching with signed distance fields by enabling users to create complex mirror effects without needing to march through non-Euclidean spaces. An example of this interoperability shown in figure five. Okay, so here, what I have to mention, I think, in the previous paragraph, um, is the AMD mesh shaders. So, let's see, what do we have here? Um, uh, it's possible to do uh, objects and object locations do not necessarily need to be stored in memory, um, in this way. Uh, it is similar to, uh, mesh uh, shade to um, approaches using work graphs and, uh, mesh shaders for real-time rendering, and let's just go ahead and cite these guys. Here. It's a perfect location to cite them. And then, I also want to cite their uh their nature paper. Uh, nature being tree generation paper, not nature as in like the journal. Uh, Man, imagine if this was a nature paper. Imagine that. Uh, what do we have here? Scholar. Bam. Let's get out that citation. Um, I'll cite them. I'm liberal with my citations. I don't know why I I'd cite their tree paper, but I will. It's fine. Uh, let's go here. Um, and we're going to CD in the projects, papers, SIGGRAPH, uh, VI literature. bib. Go down here. Paste that bad boy there

and then grab this one. Oh, I see. They published one for one year and one for the next year. That's fine. There we go. Perfect. Done. Had enough with that. Okay, uh, I'm going to go ahead and PDF LaTeX this bad boy. PDF LaTeX. Huh. PDF LaTeX main. I stop. Why do I suck so much? Why am I bad at life? There we go. And then, BibTeX. And then, PDF LaTeX once. And PDF LaTeX twice. And that should give me Yes, perfect. Exactly what I want. So, I have one more paragraph here. Um, finally, as this work introduces Quibble, a GPU meta-programming framework, um, it is important to differentiate it from other existing graphical and computational programming interfaces. Simply put, there are no interfaces that allow for flexible and efficient calling of function pointers on the GPU by default. Um, this means that implementing generic user-defined functions on the GPU can be troublesome for many applications. I shouldn't say most. Um, most. Um, and I it's not clear what I'm saying here. Uh, this means that implementing generic user-defined functions on the GPU can be troublesome for implement for many applications. Uh, when uh the compiler cannot uh, what am I looking at? There's a word here that I can't remember, but it is introspect cannot um, it can't find the function to inline. Uh, I might as well say cannot properly uh, flexible efficient calling it This means implementing generic user-defined function I'm going to do function calls. That's what I should say. Uh, user-defined function calls on the GPU. There we go. I think that solves the problem. Um, Quibble exists as a meta-programming framework to solve this problem by taking advantage of features from both graphical and computational workflows. Unlike graphical interfaces like Vulcan, OpenGL, DirectX, and others, Quibble focuses primarily on the creation of high-quality computational kernels instead of the traditional graphical with, for example, vertex and fragment shaders. Unlike computational interfaces like CUDA kernel abstractions, SYCL, and Kokkos, should I cite CUDA? I cited everything else. I should at least cite CUDA, shouldn't I? Who do you cite for CUDA? The programming guide? Um, scholar CUDA programming guide. I'll cite CUDA, I guess. It Who Who do you cite for CUDA? What if I just go to scholar? And I do CUDA. Is there like a paper that introduces um, is this it? Uh, can I sort by cited by? How do I sort again? Sort by relevance. CUDA by example introducing It looks like CUDA by example looks like the most This common book. So, maybe I just cite these guys, even though I haven't read it. Yeah, that looks like what I'm going to do. I mean, everybody knows CUDA, right? So, just going to cite these bad boys. BibTeX. All right, let's go. Uh, do I still have that other terminal up? Yes. Uh, here I am. Can I say something that's mildly annoying? That no one else in the world, I'm fairly positive, absolutely world cares about. Fairly positive, I'm the only one, okay? VI is not Vim. They're different, okay? They are different. Um, I use the I. I don't use Vim. I usually don't have problems with Vim, but there are some minor nuances to Vim that I don't like, which is why I use VI. Um, very simply put, sometimes interactions with my keyboard and, um, and like Vim keybinds or whatever, um, keystrokes or whatever, uh, they're just slower than they are in VI, and I don't use any features that you would need in Vim. So, I just use VI. However, very recently, as of the past few months, Arch did something weird that I never expected them to do. They took VI out of uh the um the package manager. You can now no longer download VI on Arch. Um So, you have to use Vim and use it in VI mode. That's what they do. You can still You can still, you know, download Vim. That's fine. But, if you use VI, it's naturally just going to use Vim instead. Like, it will automatically do that. This is what all the other things do. So, now every time I do I always have to have Vimrc now. Um and things are annoying, right? When I open up this file, this color down here. I don't know if you can see it, but this color there's a little blue tilde. I don't want color. I You know, if I wanted color, I wouldn't be writing code here. I'd be writing code in a different text editor, okay? It It's annoying. It's annoying is what it is. Um but uh I'm the only one on Earth that I think cares about this. There's literally nobody else who actively wants VI instead of Vim. Um CUDA? I can't believe I'm citing CUDA. Wait, why am I citing CUDA? Who cares? There you go. Uh Quibble instead focuses on compiling a large custom kernel from user-provided functions at runtime. This

allows for users to quickly and efficiently iterate on their function set uh without having to rebuild their primary executable in some other language such as C. Uh though several modern languages such as Julia, Rust, and C++ provide flexible metaprogramming capabilities, it's difficult to use such tools uh directly for this application as we instead need the stage compilation approach that's commonly found in graphical languages like Vulkan and OpenGL. Uh for this reason, Quibble is more similar uh to kernel fusion frameworks and is the first programming framework that's flexible enough to allow for function system-based approach for general-purpose graphics. That is all true. However, if I'm going to cite CUDA, I should also cite Vulkan and OpenGL. Like, is that how I don't know. Um commonly found in graphical languages like Vulkan and OpenGL, though also present OpenCL, uh which is the um primary uh library used Uh Should I say which Quibble transpiles to? That's what I should say. So, if I'm going to cite CUDA, I should probably also cite Vulkan and OpenGL, right? Do people cite this? Um Okay, what do we have? Uh making real trees with GPUs would make it on nature. That's true. Uh kernels is a term uh for combined and exchangeable functions that might better fit your naming since we are in the GPU context anyway. Uh what specifically So, kernels to me have a distinct meaning, right? Um And I used them I think in the way I wanted to use them. But, could you explain what you mean when you say I should How do you want me to use kernels? Recent updates broke a lot of things in Arch. Uh yeah. It's okay. I don't mind it too much. Uh I lost my syntax highlighting, why not? Yet, a new Vim is definitely better. My thing with syntax highlighting is the same reason I don't like the IntelliJ like, you know, like auto-complete thing when you're writing code. Like, when you're writing code, I don't want my text editor to tell me like what I should or shouldn't write. You know what I mean? Like, I don't want to write the word for and then for that auto-complete I is equal to I like I is equal to, you know, zero um I plus thing. I don't want that I don't want auto-completed code. I don't want that, you know? Because like I know what I'm going to type. I have in my mind And you telling me what I'm supposed to type makes it so now I have like just a nanosecond a millisecond of thinking about what I'm going to type that I didn't have before. I'd rather it not be there, right? And the same is true for syntax highlighting, you know? Like, I'd rather just not have it because now I don't have to think like why is this color blue? Why is this purple? Is it supposed to be purple at this point? I don't know. Like, just I don't want these thoughts. I want to just write code, you know? Um and it's just how I've always done it. And I I you know, this is why I'm the worst person to talk about AI. I am genuinely. Every time you speak about exchange function and exchangeable function, use kernel. Okay, maybe. An exchangeable function. Um So, what you're saying is this is more of a composable kernel language, which is true, than a composable function system framework. I think that is a fair point. Um Uh but auto-complete is faster to type and also syntax highlighting is easy to read, making your eyes focus on the important bits. Yeah, but I I don't know if that's necessarily true. I mean, maybe it's true for you, but like, you know, my eyes just glaze over the colors. You know, like I I sometimes can't read error messages because the error is in this bright red and I just I don't look at bright red. I only look at white. So, if you give me bright red in the middle of a terminal, I'm just going to skip over it and assume it's not worth me reading. Um meanwhile, if everything is white, then everything is equally important and I'm going to read all of it, you know? Um I do agree like I agree that in principle auto-complete is faster to type and that it in principle better to read. I agree with you when you say that. But, also like again, I'm never in the situation where the manual typing of code is the slow part. So, I'm okay with losing like a microsecond or a nanosecond of time or like even a couple seconds of time or even a couple minutes of time if it means that I enjoy my development experience more, right? That's just how it is. Um Vulkan programming guide, is this the way? I guess so. I guess this is the number one cited paper, I'd say. There's only 108 citations. Well, that's fine. Doesn't really matter too much. I'll cite them. Uh cite BibTeX. Why not? It's okay. I'll just cite everyone. It's okay. I You know, no reason not to, right? Um Cite and we're going to cite a Vulkan. 2016. Vulkan's been out for 10 years. That's something. Um scholar OpenGL. I mean, these existed before like people actually actively cited software, you know? Um I'm using

What am I using? I'm using version 3. 0 and 3. 1. So, I guess I'll cite No, no. I That's for OpenCL. For OpenGL, I guess I'll cite this one. Man, lots of people cited OpenGL, but very few people cited Vulkan. Am I dumb here? Am I missing something? I must be missing something, right? Why are there so many sites for OpenCL, but there's no sites for OpenGL? Uh oh, sorry, the other way around. There's so many sites for OpenGL, but there's no sites for Vulkan. That is weird to me because I thought Vulkan kind of made a pretty big impact. That was dumb. That was dumb. Maybe I got the wrong citation for Vulkan. Um Uh Neider, 1993 open. Come on. That's fine. Um Next. X paste. There we go. And the last one's OpenCL and I specifically want this. Uh OpenCL specification, OpenCL programming guide. Um I guess programming guide is what I want. But, the OpenCL specification is probably also fine. Um maybe the specification is fine. This is the one that most people are citing, so let's just grab that one. It's okay. Uh up here. Open Where are you? Come on. I'm an idiot. There we go. All right. Let's do that one as well. — I do think it's kind of silly to do these citations though. Like, definitely not It's It's fine. Who cares? All right. Uh why do you even cite OpenGL and Vulkan? Everyone in the field knows what they are. I know I shouldn't I shouldn't, but I'm doing it. The thing is I had I had citations only for kernel abstractions, SYCL, and Kokkos, right? Because I thought these were ones that um that, you know, graphics people might not know because they're computational stuff. But, if I'm citing them, I might as well cite everything else, you know? Who cares? It's I mean, citations are free, right? Um so, doesn't really matter that much. Uh BibTeX PDF Uh PDF Yeah, I mean, you know, it The citations are free. It doesn't matter. Yeah. Um And uh I would rather have more information than less, you know, because at the end of the day people go to your um uh go to your bibliography, you know? But, my bibliography it's getting a bit long now. It's not that long. It's obviously a respectable bibliography size. But, uh it doesn't matter. It's okay. I mean, no one cares you know, there's no like bibliography measuring contest to see who has the biggest bibliography, you know? That doesn't matter. Um okay. So, that means I can remove the to-do up here. And I believe the related work seems fine now. It's okay. It's good enough for me for now. Um You're not reading those documents you cite anyway, so it defeats the purpose, IMO, but that's just my two cents. For the record, I have read the CUDA programming guide. Um Okay, so for the record, those were not exactly what I did cite, but I read most of those. The only one I didn't read was the OpenGL one. Um uh I've read uh the CUDA programming guide, which I think is the one I cited. Maybe I could have the specification. Um and I also read these papers. 100% I read those papers. Um then the question is what didn't I read? Uh the ones I didn't read was the one from OpenGL. That's the one I didn't read. Um but, you know, I have read the OpenGL specification and I believe this was the Vulkan programming guide, which I have also read. So, I mean, you're saying I didn't read them, but I have actually read those. And it's weird if I have the other ones and not OpenGL. Um So, I think the CUDA one I usually cite the programming guide. So, it is a bit weird that I cited the book instead. Um but who cares? Like I mean we're really I think this is not worth the time to think about any further. Um for my setup he's Notepad++ while I have autocomplete it only completes when I enter when enter is pressed. If I don't like it, I press escape and the dialogue disappears. I understand what you're saying but I just don't like it telling me like I don't like when I'm typing, you know, I don't like to say four and then I don't like this box that appears, you know, um uh I equals zero I less than 10 um I plus or something. Like I don't like it when it shows me oh you should type this next, you know, um like I don't like that part because then like while I'm typing the for loop, I now have to reason about am I just going to type this or am I going to press enter? I have that extra bit of reasoning that I have to think about. It's not much but it's enough that it annoys me. Uh so I don't like I just don't like the intellisense thing at all. Oh no, multiplexer. I actually agree with you. Your statement you're not reading the documents you cite anyway so it defeats the purpose.

I 100% agree with you. Usually I say that you should only cite papers that you have actively read um and you have seen me in the stream cite papers that I haven't read too deeply. Like for example the tree paper I didn't read too deeply. Um I looked at the other papers off stream a bit but the tree paper I didn't read too deeply. I just kind of saw it as a good example and I said yeah, it's a great example let's go ahead and put it in. Um it's fine. Um but uh as far as the uh the other stuff, you know, I you know, I agree. You should only put in stuff that you read. Um but I also understand that citations are a metric of not success but uh of payment I guess you could say in academia and if you don't you know, being liberal with citations is nice because it means that you're being you're helping other people in their career. Even if it's just like a small you know, gesture of gratitude. It's like hey look, thank you for publishing this thing. It was helpful to me um in some way, you know, um That's all it is. Um okay, so related work is good. Um the next section that I have it's fine. I'm not saying related work is great. Um so brief introduction to composable function systems for rendering. Um Simply put a function system is a set of functions with defined inputs and outputs. I might be going to Okay, I see where you're coming from. It makes perfect sense. Yeah, again I feel like I'm not critiquing you you're not critiquing me. We're just talking. Just for the record I want to make sure that I'm not like I'm getting to the point in the stream where I'm a little bit tired and like now my sentences come off as like very um how do I put it like I'm angry unnecessarily. I'm not angry at all in any way, shape or form. Uh let me check something really quick. Um Okay, so um we've had some weird viewership on today's live stream. I'm just saying like the viewership on Twitch has been great, I think. Um the light is yellow. I I've the light the background light is yellow. Yeah, I've got So, um you can't see it. Can I bring it down? I can't move the camera up very easily. Um can I show it over here? It barely doesn't fit. No, okay. I have like lights over the top. I'm in a box right now and I've lights over the top. Um maybe I can I move this up a little bit? No, I can't otherwise it's going to mess everything up. Um no, I'm not sick right now but I did mention at the start of the stream I'm a little bit green. Can I color correct? Let me try to color correct really quick. I'm not very green though. Very green for some reason. Can I lower the brightness of the dim minus. Ah. It was just very bright. And then if I do that that's better. It's much better. Yeah, that's great. That's fine. I can also like change I can like do this. No, that Wait, what can I do? These lights have like a whole bunch of different actions. Wow. What a brilliant That's not bad. Wait, what am I doing now? If I turn them off then — it's just slightly better with them on. Let's go back. What a waste of everyone's time, isn't it? It is fine. It's fine. I didn't realize it was uh No, don't do that lights. What are you doing turn off and on? No, it stay on eight and stay. There we go and I can increase the lightness just like that. That's okay. That was very green. This is fine but if I turn it off is that bad? Yeah, it's definitely better to have them on. What's this button do? Oh, it's a timer. Okay, perfect. Done. All right. Great. Thank you guys. I appreciate that. Um — All right, so what did we do today? We did nothing today. I feel like every stream I'm like what did we do today? Nothing. Um but we did something today. We we got through the related work section which is fine. We had a lot of conversations about AI which we need to not do very much on stream but we did. Um and uh um yeah, it's okay. Like you know, we're getting there, right? Um again I'm still getting better health wise, right? Like that's my primary focus right now is just starting to feel better. I have intention of getting a video out until basically the end of May. Um I know that sounds really dumb but like I want to get the paper out kind of at the same time the video comes out cuz I've got like an animation that I want to work on and I also want to put that animation like images from the animation in the paper. So, right now I'm getting the paper done. I'm getting the concepts done. We're going to kind of work through a lot of these like complex things and then after that I'm going to uh start actually like you know, animating and mixing in different research projects with it. There's going to be a lot of different things that I'm going to do. Um so, I think it's been 2 and 1/2 hours. Yesterday I kind of crashed after stream so I think this might be my limit for now. Um yeah, I think I should cut it here. Um I'm going to say the following. Um Number one, thank you guys so much for being here. I know I ramble too much today. but I have to be honest in saying that yesterday's stream and the day before that I loved those streams. I genuinely enjoyed um like every conversation I had with you guys um and like the engagement with uh not only like the research that we're talking about now but the research we were doing on the archive last sorry last Friday. Um it was great. I I'm I just you know, I'm not we are not really moving in a new direction in this channel. Like not exactly uh but you know, I'm making content for the first time in a very long time and I

know this time zone's not good for everyone. I know it's not. Um I I'm going to try to fix that in the future. I want to do this stream at this time zone and then another stream at another time zone every day and then somehow combine the VODs. Um I think there's a way to do that. Um but uh yeah, we're at the stage now where I'm just trying my best to get back into things and I really want this paper out and I've got a whole bunch of other research ideas that are coming up. Um so, yeah, we'll get there. Um my goal is to get this paper um out and uh a video about it in a month and while I'm working on um getting this paper out and stuff like that my goal is to then um start to get other research projects in there. So, starting next month um again this is you know, basically 2 weeks away. Um my hope is that it's I want to start like incorporating these other research projects because you guys have given me a lot of ideas. There is the uh cryptography stuff. Um there has someone has given me a uh a concept of something to do with proofs of work which I find really interesting. Um there's some quantum stuff that I want to do. Um there are a whole bunch of different people talking about different compiler design aspects. There's so many things so many interesting things that we could do. Um So, if you're a researcher and you want to talk to me about like doing research whatever that means, just you know, reach out whether be here on Discord or anything else. I I want to talk to you guys um and I'm happy about that kind of stuff. Um does Twitch chat get authorship? I might put you in the acknowledgements uh but if if anybody in Twitch chat does something like worthy of getting a pub like worthy of authorship, yeah, of course you get authorship, you know? Like if you guys actively engage with the research like of course you get authorship. There's no question. Um is yellow for me as well if that helps. Yeah. Well, it's fine. It is what it is. Anyway, that's it for now. 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 will see you guys around next time for the next episode of Leo's Labs which will be hopefully tomorrow and hopefully we're going to start working on the sign distance fields and that's what we'll do. Um like actually meaningful work. We'll see. Um That's it. Thank you guys so much for coming in and uh YouTube people I will see you now and the Twitch people we'll see if there's somebody to raid. I still kind of want to raid people but I'm not sure exactly how to do that. So, I'm going to say bye-bye to YouTube for now. Bye-bye YouTube. Um