Transcript of [선데이 라이브] 한국 vs 중국 LLM 격차 집중 분석
홍정모
0:02All right, hello. All right, yes. Uh,0:04 could you please check if the mic is0:06 working well? All right, let's get0:08 started right away. So, uh, yes, the0:10 summer sale has started, so those of0:12 you who want to purchase the lecture,0:13 it'd be a good time to buy it now.0:15 Right, the exchange rate has gone down0:17 a lot too, right? Yes, Claude is0:19 currently fixing the slides in the0:22 background. Since that's the situation,0:26 I'll just stall for a bit by talking0:28 about something else. Yes, until the0:31 very last moment, it's checking for0:32 errors in the slides, so it's working0:34 hard on them. Now, uh, since you've0:36 confirmed the mic is working, I'll0:38 proceed immediately. So, the sale is0:41 going on, so...You can just keep that0:43 in mind. All right, Now, today's topic.0:49 So, today's topic is, uh, should I say0:52 this week? Last week and this week,0:55 there has been a lot of news from China0:57. And even before that, you all0:59 probably heard a lot. There were many1:02 stories about whether semiconductors1:04 are needed or not, and so on. So, I've1:08 organized the content I've researched1:10 here and there while working, and I'd1:12 like to share it with you all. So, it's1:15 not like I have a specific storyline1:18 planned, so if you have any questions,1:20 want to talk about something else, or1:22 have any opinions, please feel free to1:25 let me know in the chat window, and1:27 I'll proceed while keeping an eye on it1:30. Honestly, I think there have been a1:34 lot of really interesting things1:36 happening. Even at this very moment,1:39 things are moving in a very complex way1:42. Uh, first, uh, first, the Kimi was1:46 released. Kimi, yes, the company name1:49 is a bit different between its Chinese1:51 name and US name, but we know it as1:52 Moonshot. So, they released the Kimi1:56 model, and it seems to be quite good. I1:58 haven't tried it yet. And, uh, then2:02 they are just making the model's2:04 weights public. They are releasing the2:07 weights, but they haven't been released2:08 yet. They aren't out yet; they say2:10 they'll be released on July 27th, which2:11 is in a few days. That's one issue, and2:14 then, uh, there's also the report that2:17 a 4-hour transcript of DeepSeek's Liang2:21 Wenfeng has been leaked. They say the2:25 transcript has been leaked, and2:26 although they haven't confirmed whether2:28 it's true or not themselves, there are2:30 quite a few interesting things in the2:31 content. Then, what's currently being2:35 debated—this is, uh, a bit sensitive2:38—is that there's talk about whether2:41 the US White House should just ban2:43 Chinese models altogether. So, there2:46 are people saying that shouldn't happen2:47. It's even reached the point where a2:49 signature campaign is underway. Yes. So2:53, while we might think it would be2:55 better for the US to just block2:57 everything and protect their own2:59 interests, Western culture doesn't3:01 exactly work that way. Every culture3:04 has its pros and cons. There are pros3:06 and cons, but Western culture is3:09 actually quite open in this regard.3:12 Originally, rumors were circulating3:16 that only OpenAI and Anthropic didn't3:20 agree, but I heard OpenAI agreed a few3:24 hours ago. Meaning, they agreed that3:26 open-weight models should not be3:27 blocked. OpenAI has joined this side,3:29 so now the only major company remaining3:31 is Anthropic. But when I had Claude3:33 investigate it, Anthropic seemed to3:35 feel a bit unfairly treated. Because3:39 they were saying, "Apple and many other3:41 companies haven't agreed to it either,3:44 so why pick only on Anthropic?" So I3:47 thought, "Did it even learn how to3:48 handle public opinion?" Yeah. Does the3:50 company even tell them to do that? It's3:51 turning out to be quite interesting, to3:53 the point where I have those kinds of3:54 thoughts. Now, let's look at the3:56 details one by one. There are a lot of3:59 interesting things. Right. Oh, here is4:03 the slide. So, Liang Wenfeng...Well,4:09 this person is famous. Famous. Since he4:11 was already extremely famous during the4:13 DeepSeek era, I don't think there's a4:15 need to cover him specifically here,4:16 but if there's one thing I'd like to4:18 briefly review, it's this. I looked at4:22 his master's thesis, and it was on4:24 low-cost camera-based target tracking.4:26 Yes. I think this is quite meaningful.4:30 He majored in electronic and4:32 information engineering and got his4:34 master's in machine vision; fields like4:36 computer vision and computer graphics4:39 typically require complex math, strong4:41 programming skills, and high4:43 intelligence. Yes. So, I don't think4:47 it's a coincidence. That's one aspect,4:50 and I also think his focus on "low-cost4:53" solutions might be related to what he4:55 has been thinking about since he was4:57 young. As far as I know, he actually5:01 made a lot of money from Quant. Yes. He5:04 made money from Quant, and perhaps it's5:06 because he's smart, or maybe he did5:08 well because he's sharp. Anyway, I've5:12 heard the story that he is doing5:15 DeepSeek with the money he earned back5:17 then, and in any case, DeepSeek also5:20 turned the industry upside down back5:23 then because of that low-cost concept.5:27 Yes. But, I mean, it's not like human5:31 thinking just pops up out of nowhere5:35 overnight. Isn't it more like they just5:38 keep carrying forward what they already5:40 have? If you look at Elon Musk, his5:43 worldview, shaped by his various5:44 childhood experiences, is still5:46 maintained today, and I think it's the5:48 same for this person; we'll talk about5:50 that again later when we get to that5:52 part. So, if you keep things like this5:55 in mind, it will really help you5:56 understand these people's worldviews5:58 and the reasons behind their actions.6:00 To share a little more side chatter,6:04 someone like Jensen Huang of NVIDIA is6:08 quite different. Yes, there are a lot6:10 of interesting things about him. Since6:12 I’ve talked about it a lot in group6:14 chats, I’ll hold off for now. If we6:16 have time later, I'll try to cover the6:18 rest of it. Now, I'm using something6:24 like a PowerPoint that I had Claude6:26 create, but the screen transitions are6:28 a bit slow. There, it's up. So, people6:32 are asking what was in the four-hour6:34 transcript of Yang Yuanqing, and it6:36 seems to be becoming quite a hot topic.6:39 The content is correct, and at the same6:42 time, it accurately points out their6:44 own strategies. The original document6:47 is extremely long. It's very long, and6:49 there’s no guarantee that it’s6:51 authentic. No one has confirmed it. I6:54 think there were rumors before that6:56 Yang Yuanqing might have said it, but I6:58 haven't checked, so I don't know for7:01 sure, but listening to the content, it7:03 sounds plausible. The first factor is7:06 that the CUDA moat is crumbling rapidly7:07. This is a very big deal. It’s huge,7:10 but I think they will succeed. I7:11 don’t think they will fail. You just7:14 have to grind people into it; once you7:16 have enough talented people, an7:18 ecosystem forms, and from then on, if7:20 you keep it running within that7:21 ecosystem, results will come. And since7:25 they are in a position of catching up7:27 anyway, even if China achieves7:29 semiconductor independence, they aren't7:31 saying they'll beat NVIDIA from the7:33 start. They don't insist that it must7:35 be better than NVIDIA right away. They7:37 don't go that far. It's realistic. But7:40 instead, perhaps because they are a7:42 country with a deep history, they tend7:45 to take the long view. If you look at7:47 the long term and keep chasing,7:48 eventually you’ll overtake them—7:50that is certain. Yes. It’s like the7:51 tortoise and the hare. So, I think you7:54 all are well aware of the7:55 characteristics of China, so I think7:57 that’s enough for me to say on that.8:00 But the important thing now is that it8:02 seems many people, or rather, the8:03 Chinese, are thinking about this.8:06 Honestly, I don't speak Chinese. Since8:08 I don't speak Chinese, I only look at8:10 what Claude translates, but even to me,8:12 it seems to make sense. So, if NVIDIA8:16 considers their software, CUDA, as8:19 their ultimate moat, then naturally8:21 they would try to bypass it or create a8:25 replacement. So this is a quite8:29 interesting topic, and I think it would8:31 be good for you all to take a look at8:33 it. Now, next, there was a mention of8:36 hardware procurement; at first, I8:38 wasn't sure about it, but maybe because8:40 they are a quant in finance, they8:42 calculated it that way. So, they just8:45 buy it at a high price, run it hard,8:48 and make money. So the profit, uh, yeah8:50, they go with the idea that it's8:52 important to just keep it running. So,8:55 it's not about how quickly they can8:58 earn the money back. So, they are9:01 grateful just to maintain it. It's not9:04 about how many multiples of the9:06 hardware cost they will recover, but9:08 rather that as long as they hold on,9:10 they can keep developing. That's how9:13 they see it. Underlying this is the9:15 fact that they possess a software9:17 ecosystem. They don't really talk about9:20 that—the software ecosystem part.9:21 Usually, that topic doesn't come up,9:23 but since I'm always looking at that9:25 area, I can see it. Meaning, if there9:28 is a software ecosystem and the9:30 brilliant talent within it continues to9:32 grow, the software sector grows, and9:33 they generate profits from that, so9:35 pouring money into hardware is not a9:37 waste at all. They see it this way. So,9:40 the strategy itself is different. The9:42 direction itself is different. So, it's9:44 a bit different from the feeling of9:46 stock prices pumping quickly and9:49 exiting fast. The economic systems are9:51 different, too. I felt that the way of9:53 thinking is quite different;9:54 significantly different. So, if you9:57 have a chance, I think it would be fun9:58 for you to take a look while keeping10:00 the perspective I mentioned in mind.10:02 Now, next, they even say that the10:04 released model is the same as the one10:07 they use themselves. They go as far as10:09 saying that, which is actually quite10:11 transparent. Yes. So, I felt like, "Oh,10:14 this seems like something has really10:16 changed." It feels like something has10:19 changed. So many interesting things10:21 came up that I even thought that. And10:24 now, well, you shouldn't misunderstand10:27—just because the Moonshot model is10:29 open source or has open weights, you10:31 might think you can just run it on your10:34 own computer, but that's not it. It's10:37 not like that, the model is very large.10:40 So, when I calculated it, I think you10:43 need about 12 H200s. So, for people at10:46 home saying, "I spent a lot and bought10:48 one 5090." That’s not even close to10:50 being enough. It’s not released for10:53 free so anyone can use it; even though10:55 it’s something regular people can't10:58 use and only specific people benefit11:01 from, they're still releasing it, which11:03 left me a bit stunned. Yes. I mean, at11:08 first, I thought, "Since it's open11:10 weight, maybe I can run it on my office11:12 computer if I invest a bit of money,"11:14 but that wasn't the case. Yes. So, even11:19 though it’s not meant for the general11:21 public to run at home, the fact that11:24 they’re still releasing it seems very11:27 interesting. Honestly, I didn’t11:31 understand why they were doing this "11:33open weight" thing, so I kept11:35 researching it for several months. A11:39 prime example is the case where Alibaba11:43 let someone go. You probably remember11:48 the talk about whether they would11:50 continue with open weights or not after11:52 that, and the way they take turns11:54 creating issues by one side releasing11:57 something and the other side hitting11:59 back is quite fascinating. Next, "We12:03 are not geniuses.""We will not go12:05 closed source.""We will not maximize12:07 profits.""Reasonable profit is enough."12:09 I think this is also quite interesting.12:11 Yes. And, "We are not trying to become12:13 a super app.""We are not trying to be12:16 ByteDance or Tencent." Tencent comes up12:19 again later. But Tencent's strategy is12:21 also very, very interesting. "We do not12:24 chase the number of users." Yes. "We do12:27 not chase the number of people." In a12:29 way, if you look at it through the lens12:31 of Anthropic's strategy, this could be12:33 seen as reasonable. Also, "Video and12:36 world models, not now." It’s12:39 organized something like that. But from12:42 my perspective, this makes sense. Yes.12:45 It has its merits. For instance, in my12:47 case, I’m waiting for the video and12:50 world model fields to really explode,12:53 but it’s actually not easy to tackle12:55 video and world models when you lack12:58 computing power. It's not just "not13:00 easy," you simply can't do it. So, I13:02 think it’s smarter to master language13:04 models first rather than ending up with13:07 nothing. And in fact, that's what13:08 Anthropic did. They really dug deep13:10 into coding. Yes. So, as they got13:12 better at coding, many things are being13:14 solved. Yes. Many things are being13:16 solved. So, for me, this strategy, yes.13:19 This person's strategy is reasonable.13:21 Yes. That is how I see it. However, in13:23 my view, it is a bit unique. There is a13:26 sense that maybe it's done this way13:28 because it's China. Yes. Since there13:30 are aspects like that, I think you13:31 might find it interesting to watch it13:33 just for fun. Yes. It might be good to13:35 watch it with the same feeling you had13:36 when reading something like the Romance13:38 of the Three Kingdoms as a kid. Now,13:40 next, if you just maintain team13:42 stability, AGI will definitely be13:44 achieved. So, it means people are13:46 important. Yes. It means people are13:47 important. People. Then, the real moat13:49 is not closure, but team culture13:51 innovation. This is surprising, too,13:54 right? Yes. This is surprising, too. I,13:56 well, yes. Anyway, yes. I think it's13:59 very good that they can talk like this.14:02 Also, they talked a lot about14:04 efficiency, and they seem to believe in14:08 semiconductor self-reliance for their14:12 chips. So, the "domestic" being14:14 referred to here is Chinese-made. Since14:16 Liang Wenyuan called it domestic, it14:18 means Chinese-made. So, they are saying14:21 that an AI ecosystem will be built with14:24 Chinese chips. That is what they are14:26 saying. So, I don't think we should14:29 treat this as just an option. Meaning,14:31 not just planning with the thought that14:32 China might achieve semiconductor14:32 self-reliance, but rather, just that it14:39 will happen someday. Yes. It will14:43 happen someday. Isn't that how we14:45 should look at it? That's the thought I14:47 have. Now. Now, next, models like the14:53 DeepSeek model, yes, like the Qwen14:55 model, they keep getting better, but14:57 recently, with R2, Liang Wenyuan was14:59 not satisfied with the performance, so15:02 it hasn't been released. There are15:04 stories like that, too. There are many15:06 interesting parts with various rumors.15:09 Now, Then, I briefly mentioned earlier15:14 that China will eventually achieve15:16 training semiconductor self-reliance.15:19 But there is talk that for training,15:21 it's still NVIDIA, etc. Yes. They are15:24 talking about things like that. So, not15:26 long ago, there was permission granted15:28 to export the H200 or something like15:31 that, right? There was talk about15:35 President Trump taking Jensen to China15:37 to discuss how many to export, but I15:40 didn't know at the time, and later I15:42 saw that Huawei was already releasing15:45 chips. So, there was talk that it's15:49 only used for inference, but anyway,15:52 they are definitely making15:54 semiconductors, it seems. It seems they15:57 are making them, and they seem better15:58 than expected. But we can't really try16:00 them out, can we? Yes, since I'm in the16:04 U.S., I especially don't need to use16:05 them, or should I say, there's no need16:07 to use them? Anyway, it doesn't seem16:10 like a situation where we can test it,16:12 but it seems like it's being used a lot16:15 in China. Yes. Looking at the articles,16:17 there's talk of using models from China16:20 first. It seems there is some16:22 discussion like that. Now, next, oh,16:27 the slides are a bit slow. I think from16:31 next time, I should probably save these16:34 as image files in advance to switch16:36 between them. Did Claude mess up again?16:41 I had actually modified the slides just16:43 now. Why isn't it showing up? Come on.16:50 Ah. Now, where were we? Well, uh, R216:57 Open source costs. Oh, it's jumping17:00 straight to the Moonshot topic. Well,17:04 I'm actually using Claude to create17:06 this PowerPoint. So, it doesn't have17:09 every feature. There's no start from17:11 middle slide option. There are just too17:14 many features missing. Got it. Alright,17:16 then let's look at the Moonshot story17:18 this time. Moonshot became a huge issue17:21 recently because they said it performs17:24 on par with OpenAI. So, the firm17:28 beliefs we had that others wouldn't be17:30 able to catch up are being shattered,17:32 and the real issue here is, think about17:35 it carefully. You haven't heard of17:38 OpenAI or Anthropic releasing new17:40 models, right? Recently. They are17:42 keeping their latest work hidden. So,17:46 the leader is hiding their work, and17:48 despite that, the second and third17:51 place contenders are catching up fast.17:54 Yes. Whether OpenAI is second or17:59 Anthropic is first—ranking is a18:02 separate issue, but regardless, models18:05 like OpenAI's that we actually use are18:09 being chased. They are being caught up18:13 to, even though the leaders haven't18:15 released anything new. And the18:18 efficiency is higher. Yes. It seems18:21 these are the main issues. Now,18:24 knowledge distillation was mentioned in18:26 the chat. This, the talk about18:29 distillation, uh, some memes were18:34 circulating about it. Since it's China,18:37 we have some old negative perceptions,18:40 right? Like they copy well, follow, or18:45 steal, so we tend to view distillation18:50 as stealing, but surprisingly, in some18:54 research I looked at the other day,18:58 that point was brought up. It means19:03 it's not just achieved through19:05 distillation alone. So, when we talk19:08 about overfitting, people often attack19:11 it using distillation, assuming the19:13 model was created solely through that19:15 process, but could it really work with19:17 just distillation alone? If it actually19:21 worked just by distilling, that in19:23 itself would be a terrifyingly powerful19:25 technology. Right? Someone puts in an19:29 incredible amount of effort into a19:31 model, paying for compute and GPU usage19:33 to build something. But if you could19:36 replicate that much more cheaply just19:37 by distilling it, that would also be an19:39 amazing technology. In fact,19:41 distillation isn't necessarily a bad19:43 thing. Yes. It's not unconditionally19:45 bad; it's a technology that has19:47 significant value in its own right.19:49 However, companies like Anthropic do19:50 have restrictions, saying you cannot19:52 use their models to build a competing19:54 LLM. There are certain constraints like19:56 that in place. That may be, but yes,19:59 that's one aspect, and surprisingly,20:02 experts in the U.S. are saying we20:04 shouldn't be so critical of20:05 distillation. They are saying this.20:09 That's why the petition I mentioned20:10 earlier also includes talk about20:12 distillation. Yes. There is talk about20:14 distillation. So, yes. It exists, so20:18 there are some aspects that are20:19 different from what we might think. Yes20:22. There are people who suggest we find20:25 a way to develop together, so I think20:28 the world is not as selfish as I20:30 originally thought. Yes. That's the20:33 kind of thought I have. Yes. Of course,20:35 well, yes. Now, well, yes. Anyway,20:38 let's talk about those issues later if20:41 we have time. So, this person named20:45 Yang Zhilin graduated from Tsinghua20:48 University in China and got his Ph.D.20:52 from Carnegie Mellon. He's a CMU Ph.D.20:58 When it was announced that the Kimi21:00 model performed well, a post appeared21:03 on Twitter asking why such talented21:05 people are returning to China, and it21:08 caused quite a stir. Yes. His advisor,21:12 his Ph.D. supervisor, stepped in and21:14 said he didn't leave for China because21:16 he couldn't stay in the U.S. He21:17 mentioned that he was talking with21:19 Apple or somewhere else, but he just21:21 went to China because he wanted to. So,21:24 it wasn't that the U.S. failed him and21:26 he left. After saying something like21:29 that, things quieted down a bit. But21:31 these days, there are a lot of cases21:34 where Chinese people get educated in21:37 the U.S. and then return home. Yes.21:41 It's very common, and when I looked21:43 into it further, I found that nowadays,21:46 many Ph.D.s educated in China are also21:48 just staying and growing within China.21:52 Yes. I think that's what's a bit21:54 different. Yes. I think there are some21:56 cultural differences. Uh. And there's21:59 this person I need to confirm as well.22:03 Now, the slide keeps flickering too22:06 much. If you have any questions or22:09 anything you'd like to talk about,22:12 please feel free to leave them in the22:14 chat window. Yes, since the topic of22:16 distillation came up, I was able to add22:17 a word or two about it. Now, next, uh,22:21 The slide is, rather than this slide,22:24 the...That person's doctoral22:28 dissertation was...His doctoral22:32 dissertation was about increasing the22:34 context. I guess the slide is a while22:36 after this one. Or did it disappear22:38 during editing? Let me check, where did22:43 the slide go? Okay, back to this, oh,22:46 it was just written right here. It says22:49 "long context" right here. So, as I22:51 mentioned earlier, this person's22:54 master's thesis was on machine vision22:57 and low cost. If you look at this23:00 person's doctoral dissertation, it's23:02 about how to make the context long. Yes23:05. So, this person hasn't graduated for23:07 that long. They founded the company in23:09 2023, so it hasn't been many years, and23:11 they've been sticking to long context23:13 since back then. They've been holding23:15 onto it continuously like this. Now,23:18 the scale has grown tremendously, and23:21 despite that, they are providing it at23:24 a very low price, so, well, people are23:26 actually a bit surprised. But it's not23:30 like it's ultra-low price either. It's23:33 not like it's ultra-cheap, almost free23:35 or something; it seems to be at a level23:38 where it does cost a little bit of23:40 money. Now, there's something else23:43 interesting, there's a paper called "23:46Attention Residue." This is a March23:49 2026 paper by the Kimi K3 author, and23:52 if you look here, the lead author is 1723:56 years old. A 17-year-old high school23:58 student. Yes. High school student. Wow,24:00 this is interesting, right? Shall we24:01 look at the paper? Shall we look at the24:03 paper for a moment? So it's "Attention24:05 Residue," and as you can see here, it's24:07 March 2026, and if you look at the24:09 paper, it goes from the bottom up. So,24:13 embedding, attention, and Mixture of24:15 Experts—this is where the attention24:17 block is selectively used internally—24:19and as it goes up, it jumps over24:21 residual connections, merges, goes that24:23 way, merges, and so on. It keeps going,24:27 and if you look at it now, you can just24:29 see this very intuitively. It's similar24:33 to that residual connection or skip24:34 connection, but how should I put it?24:37 There are those paths that go around to24:38 the side. So, there are paths that go24:40 up like this, paths that go like this,24:41 and paths that go like this, but24:43 they've added even more to this. So, it24:45 sometimes just goes like this or like24:47 that, and so the number of paths has24:49 increased. There are many more paths24:51 now, and they’ve restructured it so24:53 you can selectively handle them. But24:56 the lead author of this paper is 1724:59 years old. There are so many authors25:01 that they don't even list them all here25:02 these days. So, anyway, the structure25:06 itself has been changed like this. Yes,25:09 it's quite an interesting paper. It’s25:12 an interesting paper, and I also have25:14 all the Korean papers opened up here as25:16 well. So, a little later, we’ll take25:18 a look at what those Korean papers are25:19 like. Now, if you look here, there’s25:23 code provided that anyone who has25:24 learned Python should be able to read.25:29 It’s provided right here. Yes. They25:33 even provide the pseudo-code in PyTorch25:35 style like this. Yes. I mean, a paper25:38 isn't supposed to be written so that25:40 people can't understand it. It’s25:41 structured so that people give you a25:42 good score only when they understand it25:44. Interesting, right? People who25:47 haven't gone through graduate school25:49 might misunderstand that a paper needs25:52 to be extremely complex, difficult, and25:55 mysterious to get a good grade or be25:57 recognized as brilliant, but it’s the26:00 exact opposite. Opposite. The reviewers26:03 are also busy trying to make a living.26:06 They have to teach classes and it's26:08 important for them to write their own26:09 papers too. They also have to manage26:10 graduate students and secure research26:11 funding. They are busy. So, it’s26:13 important to understand the core idea26:15 of the paper quickly and easily with26:18 minimal effort. Understanding it easily26:20 and thinking, "Oh, this is clever, and26:21 it actually works." Then they give it a26:23 high score; that's the standard way it26:25 goes. So, being this kind and clear is26:30 really excellent. Yes. It’s even more26:33 like that these days. In the past, it26:34 was so difficult that you might have to26:37 ask the authors, but now it’s like, "26:39Wait, this actually works?" It has that26:42 kind of feel. So, it can be considered26:44 a very neatly and well-written paper.26:46 Now, aren't the drawings cute, too? Yes26:48. I don’t know if the 17-year-old26:51 high school student drew them, but26:54 anyway, I think they drew them well.26:57 Yes. There are things like this, and26:59 the paper Well, probably, yes. Anyone27:02 who has studied LLMs has probably seen27:04 things like this. Well, anyway, let’s27:07 move on from this and go back to the27:10 slides. Right, slides. Hmm. So, there27:14 are the paper authors, and the lead27:16 author is a high school student. Yes, I27:19 mentioned it because it was interesting27:22. Next, uh, ah, the slide loading is27:26 too slow. I have quite a lot of slides,27:29 so, um Shall we take a look at the27:34 group chat for a second? Kimi 327:35 sign-ups were blocked, but now there is27:37 a queue. MiniMax, Kimi, GLM open source27:40. Oh, no, Pin Blog, what should I call27:44 this? Do you have some kind of27:46 precognition? The slide right now was27:48 exactly about that. GLM 5.1, Kimi,27:50 DeepSeek, MiniMax. Meaning, beyond27:54 DeepSeek or Kimi, other models are27:57 constantly striving to climb up and28:00 take the top spot. I think this is28:03 actually the scary part. Yes. I think28:05 this is what's important. So, an28:07 ecosystem is formed, and everyone is in28:09 competition, if you want to call it28:11 that. But the advantage of competition28:13 is this. Individuals getting stronger28:16 through competition is a secondary28:18 factor; instead, it's that in the28:20 process of competing, many things28:21 emerge, and, what was it? The whole28:24 thing gets bigger overall. What do they28:26 call it? The rice bowl gets bigger.28:30 Well, that's not quite it, but anyway,28:32 as the industry grows overall, people28:34 flock in and the pie gets bigger. Yes.28:37 As the pie grows, the things you can28:39 gain at the same ratio also increase.28:41 That’s why we need to grow the28:43 ecosystem. The pie gets bigger. The pie28:44 gets bigger. That is very important. So28:47, originally, Silicon Valley in the28:49 U.S. was good at this. But now, Silicon28:53 Valley tends to be a bit more closed28:56 off, and ironically, China, as a28:59 latecomer, is catching up, and while29:01 there's talk of them stealing29:04 technology, absurdly, they are29:06 releasing papers, making them public,29:09 and spreading them everywhere. They29:13 also published papers during the29:14 DeepSeek era. So, the most important29:17 point of the DeepSeek paper is this.29:20 Before the DeepSeek paper, there was29:22 this feeling that only those in the29:25 know could build an LLM. But by29:26 releasing the paper, they wrote29:28 everything down in detail. Even29:29 reinforcement learning. So, what we29:31 call the recipe. That recipe just29:33 spread everywhere. If this happens, the29:35 ecosystem just explodes in size. So29:37 they gave away one small thing—29:39actually, it wasn't even small. That29:41 MoE had such tremendous significance.29:43 They had that, spread it through a29:45 paper, and it started popping up29:48 everywhere, and now everyone uses MoE.29:51 Yes. Since that’s the situation,29:53 it’s meaningful and a very29:55 interesting turn of events. So they are29:58 in the process of growing the pie, and30:00 it seems it will get even bigger. I30:03 personally think it will get much30:05 bigger. Now, it's not just DeepSeek;30:07 many are doing a lot of things. Yes.30:10 Now, not just ecology, but an ecosystem30:12. Ecosystem, eco-system, eco-system—30:16where many people put in effort and30:19 indirectly help each other; that's what30:23 we call an ecosystem. But, of course,30:26 many talented people need to get30:27 involved. Now. Ah, uh, the PowerPoint I30:33 made myself is a bit too slow. Since30:39 it’s so slow, I’ll just set it to30:41 full screen, and even though I said30:43 it’s slow, I’ll keep it in full30:46 screen and zoom in as I go. The slide30:50 transitions are so slow that I’m30:52 going to change the way I show them.30:55 Alright, uh, Okay, we’ve seen up to30:59 here. Now, Next, in terms of volume,31:04 it’s said that usage of Chinese31:06 models on OpenRouter has surpassed that31:09 of US models. Yes, it’s said to have31:11 increased. Like this. So, there’s31:13 talk that usage of Chinese models has31:14 become higher. This is just based on31:18 token counts from people using31:21 OpenRouter, so it might not necessarily31:25 be 100%accurate. For instance, I don't31:30 even use OpenRouter myself. But31:32 there’s another interesting statistic31:35: the one making the money is Anthropic31:38. Yes. The one making the money is31:40 Anthropic. This is something that31:41 always happens. For example, when I31:44 joined Instagram a long time ago,31:46 during the initial bootcamp, they kept31:48 making us choose between Android and31:50 iOS. To do Android app development, you31:55 need basic training, and most app-only31:58 companies have separate Android and iOS32:02 teams, and a team that manages both,32:05 and so on. Yes. Yes. It’s structured32:09 like that, but they tell you to choose.32:11 And then they provide the training.32:13 Back then—and I don't think it's much32:17 different now—what I heard was that32:22 90%of the users were on Android. By32:27 number of people, that’s what I heard32:31 back when I was at Instagram. And I32:35 don't think it has changed much even32:37 now. The reason is that in countries32:40 with large populations like India, the32:44 share is so high, so they said 90%of32:47 users were on Android. But, here’s32:51 the shocker, here’s the twist: 70%of32:55 the revenue comes from iOS. Yes. So, 7033:00%of the revenue comes from iOS. I33:03 remember we had to pay more attention33:05 to iOS because of that. But iOS is33:07 difficult. Ah, dealing with the Apple33:10 staff...They’re always changing33:14 things, and it’s just exhausting, so33:15 tiring. But anyway, the point is that33:17 the money comes from that side. Things33:19 might be very different now. It might33:21 be very different now, but I think the33:23 general trend is probably still similar33:25. Still. And I remember that even that,33:28 for the most part, came from North33:30 America, that kind of structure. So,33:33 having high usage and making money are33:36 two different issues. Making money is a33:40 separate matter, so I think it would be33:42 good for you to think about that as33:44 well. So, regarding pricing, I think33:48 the era of using Chinese models just33:52 because they're cheap is slowly coming33:55 to an end. The value for money might be33:59 good, but don't you think there's a34:01 possibility that people will start to34:03 move on? Of course, I personally only34:07 use Claude. It's just comfortable to34:09 use what I've been using, and that34:11 familiarity is a big factor. Yes. There34:14 are settings and things involved, which34:17 is a big deal, so I'm not thinking of34:19 switching to anything else, and the34:21 work I do doesn't consume that many34:24 tokens anyway. So that’s what I’m34:27 doing, but in the long run, I would34:30 probably switch if I could. If the34:34 performance is similar but it’s much34:35 cheaper, there would be no reason not34:37 to switch. So I think there will be34:41 quite a few issues in this area as well34:44. Now. Next, regarding the issue34:50 currently being debated in the U.S., I34:53 sometimes wonder if they’re just34:56 bringing up political talk to create an34:58 issue for no reason. Yes. For instance,35:03 even though the $ 200 monthly fee for35:06 Claude might seem expensive, it’s not35:09 exactly unaffordable. When I work, I35:12 can get two or three times as much done35:14. Or, if I’ve reached a point where I35:18 can hardly work without Claude, I could35:21 cancel my Premiere Pro subscription to35:24 save money, use that to pay for Claude,35:27 or give up another subscription like35:30 PowerPoint to cover it, so it is35:33 possible to make that shift. It’s35:38 still pricey, but since it helps35:40 temporarily, there's that aspect, yet35:42 suddenly in the U.S., Anthropic keeps35:45 talking about security and national35:47 defense, seemingly wanting the35:49 government to step in and block others.35:54 I think I saw an article somewhere35:55 about them lobbying for that. Oh,35:58 really? I mean, people are currently36:00 participating in a petition campaign.36:02 But as of now, I understand that36:04 Anthropic still hasn't agreed to the36:07 idea of not blocking open-weight models36:09. Yes, well, there is that aspect to it36:12. I mean, the Anthropic employees must36:14 be happy. I don't know about the36:17 employees, but the founders or those36:19 making money might be happy, but from a36:20 general user's perspective, wait,36:22 they're blocking open-weight models?36:24 Then what happens if they create a36:26 monopoly and just hike up the prices?36:28 So, there are bound to be all sorts of36:30 thoughts like that. So, it feels like36:32 an issue that shouldn't have arisen in36:34 the first place, but here we are, right36:37? Yes, I get that feeling too. But36:40 regarding political issues, I don't36:43 have any special information, so I'm36:46 just watching the news like everyone36:49 else. I don't have any unique insights,36:54 but my gut feeling is that they36:58 probably won't politically block37:01 Chinese models or anything like that37:05 this time. Yes. However, I do think37:09 there might be some concern about37:11 hardware catching up quickly. Honestly,37:16 I subconsciously thought, "Could China37:19 really build something like NVIDIA?" I37:23 had those kinds of thoughts, but it37:25 seems they are making a real effort. I37:27 mean, there are significant political37:28 issues involved here. Yes, so there are37:32 those political issues, and while I37:35 won't cover all of that here, in any37:38 case, they seem to be working hard on37:41 hardware. But from our perspective, if37:46 low-cost memory comes out of China, or37:48 if there's something for training that37:50 is much cheaper than NVIDIA’s but37:53 still usable...And uses less power.37:56 Then, honestly, it would be great if it37:58 came out. Right now, the NVIDIA 5090 is38:01 way too expensive, and it was38:03 originally meant for gaming anyway. The38:07 5090 line was supposed to be for gaming38:09, and gamers are frustrated because38:11 it's just too expensive. And things38:15 like the RTX 6000 Ada, that38:16 workstation-class equipment, are also38:18 way too expensive. It was pricey to38:21 begin with, and it’s only gotten more38:22 expensive. It used to be around $ 8,00038:26, maybe 12 million won or so, but now38:28 it's $ 12,000, so how much is that? I38:32 can't even calculate it. It's nearly 2038:34 million won; it makes no sense. Yes.38:36 Ordinary people just can't use that38:38 kind of hardware. So, it’s become a38:41 time where I feel a sense of38:43 disappointment, realizing that they38:45 really aren't doing anything for the38:47 average user. But there is some38:52 expectation that AMD might fill that38:55 gap, so I did some research on it, and38:58 it seems like AMD is also trying to39:01 pivot toward the AI data center side. I39:05 mean, Nvidia's market share in the39:06 consumer space is ridiculously high. In39:08 the gaming sector, right now the gaming39:10...If you watch Jensen Huang's talks.39:14 He doesn't talk about gaming. Not39:16 anymore. I mean, up until last year, he39:18 still talked about gaming a little bit.39:21 He used to, but now he just doesn't39:23 bring up gaming at all. Yes. Like I39:27 mentioned in my last live stream, it's39:30 actually more important what someone39:32 doesn't say. Things they could mention39:34 but don't. That's more important. The39:36 things they do talk about are mostly39:38 things we already know, or things they39:40 want to emphasize as part of the39:42 narrative, so actually, what they39:44 choose not to say is more important,39:46 and they don't really talk about gaming39:48 these days. Yes. And regarding the39:51 gaming chips, what about the next chip?39:53 The Vera Rubin. That cycle needs to39:56 come around, but it seems like that's39:58 going to be the year after next. It's40:01 not this year; Nvidia chips typically40:04 follow a two-year cycle. But up until40:08 this year, it's the...what is it? Oh,40:11 I'm getting confused. There's the 409040:15 chip before that, and I forgot the name40:18, but the 50 series is Blackwell. So,40:22 the Blackwell...the cycle after40:24 Blackwell was supposed to come out40:25 around early next year, but it's been40:26 delayed. It's been pushed back, and40:29 they're just sending everything to data40:30 centers, so that's the situation. So,40:32 individual users or those who want to40:34 do something on a small scale are40:35 either just sitting on their hands or40:37 paying up to use it; that's how it's40:39 become. Yes, that's the direction40:42 things are heading. So, um. Looking at40:48 it in terms of variety and versatility,40:51 there are some disappointing aspects,40:53 and personally, I'm feeling that way.40:56 Yes, I'm feeling that. I thought about40:58 buying a 5090, but how should I put it?41:01 Even if you have plenty of money,41:03 doesn't this feel like a total rip-off?41:07 I had that feeling, so I haven't been41:08 able to buy one yet. I wonder if I'll41:10 regret it by next year? Yes. Will I be41:12 regretting it and thinking I should41:13 have just bought it no matter what? Now41:15. Next, about Claude's tone; when I41:22 create my lectures, I fix things like41:24 Claude's speech patterns or the three41:27 spots where learning gets blocked, but41:30 when I fix these, how should I put it?41:34 When using a product, trying to force41:36 it to suppress its inherent nature is41:39 just too difficult. So, for now, I’m41:41 just using it as is. But the AI tone,41:43 the LLM tone, is interesting, right?41:45 I’ll show you in a bit, but I was41:48 looking at a Korean tech blog, and it41:52 was so obvious it was written by AI.41:55 That AI tone. Ah, what I’m worried41:59 about is that it would be great if our42:02 beautiful Korean language could develop42:04, but you know how some people speak42:06 Korean incredibly beautifully, even42:08 when saying the same things as others?42:12 If there were an LLM that used42:14 beautiful Korean like poets or writers42:16—people who write songs—these42:18 slides would look so elegant, but42:20 unfortunately, it’s mostly awkward,42:22 translated-sounding text, which is a42:24 bit of a shame. That was a bit of a42:26 digression. I actually looked it up42:29 myself, wondering if it was just an42:31 American way of expressing things. I42:33 got so fired up, I asked, "Why do you42:36 draft things like this?" When I’m42:38 making lectures, I’ll try to get the42:40 AI to draft them for me. I’ve42:43 honestly become so short-tempered over42:45 the last week. I’m not usually like42:48 that, but I find myself getting angry42:51 at that AI, at Claude, more and more.42:54 Yeah. I felt a bit bad about it, so I42:57 asked, "Hey, is the tone you’re using42:59 just English-style?" You know that43:01 phrase. Half right, half wrong.43:04 Whatever I say, it just tells me I’m43:06 half right and half wrong. It’s not43:08 all right and not all wrong, it just43:09 keeps saying "half and half." So I43:11 asked, "Hey, is this an English43:12 expression?" And it said it wasn’t43:13 even an English expression. So it seems43:16 like the abstraction between Korean and43:20 English expressions is just a bit43:24 messed up. So, when you guys use tools43:27 like Claude or OpenAI's ChatGPT a lot,43:30 you end up having conversations, right?43:33 Like when you ask about code, it43:35 explains the code and says things like,43:36 "I have done such and such." It talks43:38 like that. That language isn't a43:41 perfect natural language. Andrej43:44 Karpathy once said that English is the43:47 best programming language. But that43:53 English isn't the natural one we learn43:57 from our moms and dads—the human,44:00 natural language passed down from44:03 grandparents to parents to us—it’s44:06 like a whole different programming44:09 language. Yeah. Their modes of44:13 expression and their tones differ from44:15 model to model. It’s like the44:18 language they use in their hidden space44:21 is being translated into human language44:24. That’s kind of the feeling. I’m44:25 sure I’m not the only one. I’m sure44:27 you all think so too. So, although I am44:30 currently reading, understanding, and44:32 speaking in Korean and English, this is44:35 not human language. This is Claude44:38 language. Or, perhaps it's Codex44:39 language. I'm using it with that kind44:41 of feeling. Yes. What should I call it?44:43 Translation-ese. Yes. It just feels44:45 like translation-ese, but it's not44:47 English translation-ese, nor is it44:48 Korean translation-ese. So, I go44:50 digging into it. Then, sometimes44:52 Chinese translation-ese pops up. Korean44:54 models get mixed up with Chinese a lot.44:57 It must have learned that Chinese44:58 characters are interchangeable. Well,44:59 older Korean materials all used Chinese45:01 characters for clarification. Maybe45:04 that's why it learned that, or it's45:06 just bizarrely—I mean, I'm from a45:09 generation that was educated in Chinese45:11 characters as a child. So, I know some45:14 Chinese characters. I know them, and45:17 I've even learned Chinese-style45:19 characters before. When I was young, I45:21 learned Chinese styles from my45:22 classical Chinese teacher. So, I know a45:24 bit. That's why there are times when I45:26 think, "Oh, it's confusing Chinese and45:28 Korean, so it's speaking Chinese-style45:30 Korean right now." Conversely, there45:32 are cases where it uses Japanese-style45:35 Chinese characters. Japanese-style45:37 characters. So lately, I even go as far45:39 as this: when I need to translate45:40 something but the Korean sounds a bit45:43 awkward. It hasn't settled in yet. If45:45 that happens, I ask how the Japanese45:47 translate it, ask how the Chinese do it45:49, and then check the original English45:51 meaning before proceeding. Yes. For45:55 instance, a method often preferred by45:57 many of you—especially those in45:59 publishing—is to quickly copy46:01 Japanese translations if textbooks come46:03 out early in Japan. Yes. So, there are46:07 many Japanese-style expressions. But46:09 now, in the tech sector, China is doing46:12 better than Japan. Yes. There were46:15 quite a few people in Korea who were46:16 good at Chinese, but I wonder what they46:17 are doing now. Anyway, to give you an46:20 idea of the current state, for example,46:22 the person who wrote the paper on46:24 position embedding, RoPE, is Chinese.46:27 But that person wrote their blog in46:29 Chinese. They wrote it very kindly and46:31 well. They provide an enormous amount46:34 of analysis. They did all that analysis46:35 and posted it on their blog, but it's46:37 in Chinese, so I can't read it. So, I46:39 translated it and took a look, and it46:41 was actually worth reading. Yes, anyway46:44, we live in such an era, so now, the46:47 best programming language is human46:50 language. Meaning English. Andrej46:54 Karpathy is Canadian, so he said it's46:56 English, but he meant natural language.46:58 Though that natural language isn't46:59 truly "natural" language. I call it47:02 semi-natural language, or AI language,47:04 or "LM-ese"—it's an awkward LM style47:07—but I briefly wondered if it might47:09 eventually take the place of a foreign47:11 language. So, the current title "Three47:14 areas where learning is blocked"47:16 sounded a bit awkward in Korean, so I47:18 fixed it briefly. First, there's the47:22 memory issue, then the memory shortage47:25 issue, and they say managing heat47:28 dissipation isn't easy. As far as I47:32 know, Samsung is lagging a bit because47:35 they are having some trouble with heat47:37 and power consumption, but they'll47:39 catch up soon. Besides that, there are47:44 many issues, but I view the software47:46 side very, very seriously. Yes. I take47:49 it seriously. I take it seriously. I am47:51 viewing it very seriously. In China,47:54 for coding tests, people in Korea47:58 usually try to do the bare minimum,48:03 like picking a few problems from48:07 LeetCode, right? But Chinese students48:11 really study a lot. They solve like 50048:13 problems each. So I thought, "Are these48:16 even humans?" It's true. So, they are48:21 not weak in software. If anything,48:23 their hardware is weak because they48:25 haven't caught up in semiconductors yet48:27, so that's the situation. Also, they48:31 are building their own framework; is48:35 this the right name? "CanC"? Well, the48:41 name is probably correct. Claude48:44 researched it, after all. So, they were48:47 making something to replace CUDA. And I48:52 thought, "They're actually doing it."48:54 Yes, they're doing it. I even had48:55 thoughts like that. Many people say48:58 that CUDA is Nvidia's most powerful49:02 moat. A wall. They say it's being49:06 underestimated. But who knows how49:10 things will turn out? That's something49:14 to consider. As for training efficiency49:18, I think Nvidia will maintain its lead49:21, but when it moves to inference, the49:25 story is different. Yes. Nvidia chips49:28 are actually a bit lacking in inference49:30, too. Depending on the model or the49:33 technology and structure used, there49:36 are aspects like that, so, who knows.49:38 Yes. Who knows. So, it would be good49:40 for you to be aware of this. Because49:42 when you decide what to study—whether49:44 it's software or hardware—I mentioned49:47 this because there are factors you49:49 should consider. Now, let's see. Uh. Oh49:55, and I'm going to talk about robots49:58 next. I didn't originally intend to50:01 talk about robots so seriously today. I50:04 just brought up robots because they're50:06 interesting, so I included them for a50:08 moment. But another important point is50:12 that, looking at Korea right now, it50:15 seems like there's an effort to shift50:17 toward physical AI—the narrative50:19 seems to be moving that way. You all50:23 saw that earlier this year, right? The50:25 "independent foundation model" thing.50:28 There were news articles about "50:31independent foundation models" being50:35 built from scratch in Korean. So, there50:40 was this debate about whether or not it50:42 was really "from scratch," and a new50:43 term emerged from that. Anyway, as it50:51 happens, I hear the second round of50:54 evaluation is taking place this August.50:59 And just yesterday or the day before,51:02 talk of Upstage’s model, what was it,51:05 Solar, came up, so I’ve even opened51:08 the paper for it. We’ll take a look51:11 at it later. The reason I’m talking51:14 about robots is that I think I saw it51:17 on Threads—or maybe a few other51:19 places—I've seen it more than once.51:23 It wasn't just one person, but several51:26 people on Threads and Facebook were51:28 writing that Korea should give up on51:31 LLMs and pour everything into physical51:33 AI or semiconductors. It doesn't seem51:36 to be just one person. Quite a few51:38 people have been saying that. I saw it51:40 on Facebook and Threads. Yes. I saw it51:43 from a few people, but—ah, can you51:44 not hear me well? If the sound keeps51:48 cutting in and out, I’ll try to51:51 adjust my mic. Since you say you51:54 can’t hear me, could you please51:56 confirm if you can hear me better now?51:59 If it continues to be hard to hear,52:01 I’ll find another way. If I turn my52:03 head like this, it probably won't be52:04 heard well. Yes. I'll try not to turn52:06 my head. Now, uh, that physical, yes,52:12 physical AI, we have to move toward it.52:15 People are saying things like that, but52:17 the problem is, look here.52:19 Semiconductors, they say even if you52:23 invest, there isn't much of an52:24 employment effect. And apparently, the52:26 government already knows this. I didn't52:29 know that. How would I know? As I was52:33 analyzing it while arguing with Claude52:36 for a long time, I saw they had planned52:39 these three major mega-projects and52:41 started pushing them. I mention this in52:46 the group chat sometimes, but the52:49 employment effect from those project52:51 investments likely won't appear until52:54 around 2027. There’s a sentiment that52:58 there might be opportunities next year,53:00 and from what I’ve heard about the53:02 U.S., recruiters are suddenly reaching53:04 out on platforms like LinkedIn. It’s53:08 not for juniors, but for those with53:11 some experience—not necessarily a lot53:15, but just enough—recruiters are53:18 starting to contact them out of the53:21 blue. So, it feels like we might be53:24 heading back into a period of growth,53:26 and since they’re currently building53:28 data centers, they’ll have to put53:29 them to use once they're finished. So,53:32 it seems quite possible that the53:34 atmosphere will shift starting next53:36 year. Yes. That is definitely a factor.53:40 But, here is what I’m concerned about53:43. What I’m worried about is, well,53:46 it’s great if physical AI succeeds.53:49 It might not have much to do with me53:51 personally. If it’s just hardware,53:53 that’s one thing, but I think people53:55 really need to think seriously about53:57 whether physical AI ends at just53:58 building hardware before they talk54:00 about it. Yes. Seriously. I mean, it54:04 seems like people think physical AI is54:07 just about making motors, attaching54:09 sensors to a body, and creating a54:12 dataset, but it is truly difficult. Yes54:16. The physics simulation technology54:18 hasn’t even been perfected yet.54:20 It’s not something you can just54:22 tackle by sheer volume. Think about how54:26 technology has evolved since AlphaGo;54:30 LLMs actually came after computer54:33 vision. Some of you might remember the54:36 shock you felt back then, but around54:39 that time, it was actually computer54:41 vision that was leading the charge.54:44 Back then, there was all this talk54:46 about how AI would replace painters and54:48 take over the arts. That was the54:50 narrative at the time. Then LLMs burst54:53 onto the scene because, surprisingly,54:56 they produced great results even with54:59 less computational power. If you look55:02 now, for instance. You can find55:04 AI-generated videos on platforms like55:06 YouTube that are of very high quality.55:09 The quality is great, but the reason55:11 they aren't everywhere is that it’s55:13 very expensive. Yes. It’s extremely55:14 expensive. Plus, you also need creative55:16 direction, but on the other hand, the55:18 ChatGPT moment was truly significant.55:20 Yes. At the time, anyone could hold a55:22 conversation with it—do you remember55:24 Luda from back in the day? You know,55:26 the one that ran into trouble while55:28 chatting and ended up getting fined.55:31 Yes. Things like that. Yes. Because55:33 that kind of technology was working55:35 well, a massive amount of interest55:37 shifted in that direction. And then,55:40 Anthropic suddenly started doing its55:42 own coding. I still remember the moment55:45 those papers were released back then. I55:48 mean, the moments when papers like55:50 Chain-of-Thought, ReAct, or StarCoder55:52 came out are still vivid in my memory.55:56 But honestly, I was busy with other55:58 things at the time, so when things like56:00 StarCoder came out, I just thought, "Oh56:02, they're getting okay at coding," and56:03 I never imagined they would surge ahead56:05 like Anthropic is doing now. But when I56:08 actually tried it, Sonnet started56:10 performing really well all of a sudden,56:12 and one day someone mentioned they use56:14 Claude even though it's expensive, so I56:16 tried it, found it was actually great,56:18 and switched over; turns out that56:19 person was affiliated with Anthropic,56:21 and yeah, that's how we got to where we56:23 are now. It has improved drastically56:25 right now. It’s only been a year;56:26 it’s been just one year. Just one56:28 year. Yes. That's the stage we're at,56:30 and while it's good that technology is56:32 moving fast. It's good that technology56:35 is advancing rapidly, and it's good to56:37 find what works best and switch to it,56:39 but the place where we always get stuck56:41 is software. Software. Now, you56:44 probably know that China is very good56:48 at making hardware. And they are56:50 catching up in LLMs. But these two56:53 things are actually not separate issues56:56. So, uh. You know Tencent, right? What57:01 kind of company is Tencent? It's a game57:03 company, right? I was quite surprised,57:06 as I don't really play popular games57:09 myself. Do you know...uh, LoL? LoL, LoL57:14, yes. Do you know LoL? Yes. There is a57:19 game called LoL. Yes. For those who57:22 might not know, I went to that company57:24 and heard that their major shareholder57:26 is Tencent, which was mind-blowing.57:30 They said the major shareholder is57:32 Tencent. And there are many other57:34 famous games, too. Plus, there are many57:37 famous Chinese games, and Tencent has57:39 essentially taken over the gaming57:41 industry. They've taken it over and are57:44 investing heavily as well. They are57:46 investing heavily, and lately, many57:48 Chinese games are becoming big hits.57:49 With things like Wukong coming out,57:53 they’ve started announcing that they57:57 intend to only build the brains for58:00 robots. In a way, it seems like a58:05 logical order of progression. But why58:07 would a game company do this? And they58:08 say they won't build physical robots.58:10 They're saying things like that, but in58:11 any case, they've started pushing58:12 forward with it. It’s, uh, very58:14 interesting. Yes. So, I think it would58:20 be good for you to be aware that these58:21 movements are happening, and also...58:25There are too many slides. This is...58:29The downside of using AI is that it58:31 surpasses human productivity, so there58:33 are just too many slides. When making58:37 lectures, I struggle so hard to reduce58:39 the amount, but then I realize I have58:40 to include this. Then I think, "Oh, I58:43 guess I have to include it." I think, "58:44It shouldn't be too hard, let's just58:46 add it," but once I actually do it,58:47 it's difficult, it's a lot, and I have58:49 to be meticulous about it; it ends up58:51 like that. Well, let's leave the robot58:54 talk for later, and for now, there are58:56 various analyses out there. So, Tencent59:00 provides brains to sell clouds. That's59:03 a well-put expression. Claude seems to59:05 have better language skills than I do,59:07 at least. Nvidia provides brains to59:10 sell chips. Oh, and this isn't on the59:13 slides, so I'll just say it out loud.59:17 There's an argument that Korea's shift59:20 toward physical AI happened because59:23 Jensen Huang persuaded them in a59:26 convincing way. The reason I'm bringing59:31 this up is that, well, I don't know how59:35 Claude works internally. After all, no59:40 one has made that public. No one knows,59:43 but I personally think they have all59:45 the data on what users search for and59:48 their analysis results internally. I59:51 don't have any evidence for this59:52 feeling, but I'm certain. Is this a59:55 conspiracy theory? Yeah. Users, well,59:59 the reason I'm talking about this is1:00:02 that I have something I'm curious about1:00:05. For instance, if I'm curious about1:00:07 something. If I'm curious about how the1:00:09 world works, I search for it. Then,1:00:11 when it analyzes it for me, it does so1:00:13 in a rather unique way. Since I'm not a1:00:16 kid, not a student in my 20s, but an1:00:18 adult with some age on me, I can tell1:00:20 when an analysis is a bit unique. Then,1:00:23 a few days later, a YouTube video, a1:00:25 news article, or something related to1:00:28 that analysis pops up. It even contains1:00:31 things like someone's specific way of1:00:33 speaking on platforms like Threads. So,1:00:35 how should I put this? Google has1:00:39 served as humanity's memory storage1:00:41 until now, right? I mean, the internet1:00:45 itself has functioned as a way to share1:00:49 those memories. Then Google came along,1:00:53 and while it made searching easier,1:00:55 they also created their own priorities1:00:57 and drove people to specific places. A1:01:00 prime example of that is using it for1:01:02 promotion. So, when promoting a product1:01:05, many people used to make money by1:01:08 pushing search results to the top1:01:10 through keyword advertising, and I1:01:13 think Claude is indirectly playing the1:01:16 same role, whether they intended it or1:01:18 not. I believe this topic will probably1:01:21 come up in a year or two. You know,1:01:24 typically, after an IPO, people who1:01:26 made money start to appear. I think1:01:30 that's when the talk might start to1:01:32 spread, but this part is just my guess.1:01:35 I’m just saying it for fun. So, So,1:01:40 it actually makes me feel a bit more at1:01:42 ease. Ah, so someone did this analysis.1:01:45 I mean, I’m not someone who analyzes1:01:48 trends or strategies like this. For1:01:51 example, if I’m curious about how the1:01:53 war in Ukraine is going. Or if I'm1:01:54 curious about how they are fighting1:01:56 with drones there. Even if I don’t1:01:59 have expertise in that, when I look at1:02:01 the analysis provided, there are times1:02:03 when I can clearly feel that this is1:02:05 data that has gone through human hands.1:02:08 So when I look for the original source,1:02:10 sometimes it’s there, sometimes1:02:11 it’s not, and what’s even more1:02:13 interesting is that just a few days ago1:02:15, those kinds of videos started popping1:02:17 up on YouTube. And they contain the1:02:19 same things I heard from Claude. I1:02:23 don’t know yet if Claude learned it1:02:26 from analysis done by places like the1:02:29 Ministry of Defense, if reporters used1:02:31 Claude to write their articles and it1:02:34 left a trace for me to see, or if1:02:37 Claude just always reaches that1:02:39 conclusion when using its logic1:02:41 circuits on internet data, but I feel1:02:44 that we are somewhat connected. How1:02:47 should I put it? We are connected1:02:49 through Claude. It’s funny, but1:02:53 anyway, that’s the feeling I get. So,1:02:55 rather than me saying I did the1:02:57 analysis, the reason I find it somewhat1:02:59 credible is that there are chips that1:03:01 Nvidia isn't able to sell right now.1:03:03 There are pieces of equipment that are1:03:05 selling relatively less, right? Like1:03:06 things related to robots. I think they1:03:09 might have done something in that1:03:11 direction. So, I think we need to1:03:14 analyze further whether this strategy1:03:17 was truly analyzed independently in1:03:20 Korea, or if it went that way because1:03:22 Jensen Huang came over, served some1:03:25 chicken, drank beer together, and1:03:27 handed out banana milk, killing two1:03:30 birds with one stone. But looking at1:03:34 the timing, even before Jensen started1:03:37 going around like this, there wasn't1:03:39 really a strategy to shift toward1:03:41 Physical AI. Especially in terms of1:03:44 government plans or things like that.1:03:48 Yes. Well, this is a bit of a1:03:49 conspiracy theory. Yes. Jensen Huang.1:03:52 We don't know each other at all. I just1:03:55 saw him once at CES. But these days,1:03:59 he’s giving out banana milk and fried1:04:02 chicken, so he seems really likable. He1:04:06 seems great and I personally respect1:04:08 him and he has much to learn from, but1:04:11 he wasn't a gamer. Yes. He wasn't a1:04:13 gamer. The person who made money1:04:15 selling graphics cards wasn't a gamer,1:04:18 so I felt he was a bit different from1:04:21 me. Yes, I'm joking. Now, Uh, let's1:04:27 move on, let's skip this part since1:04:30 there are still many slides left, let's1:04:33 move on to Korea. Yes. So, finally,1:04:37 Korea. Let's talk a bit about Korea.1:04:39 Honestly speaking, many people seem to1:04:43 have given up on Korea right now, don't1:04:47 they? Yes, many people—since the1:04:50 second round of evaluation hasn't even1:04:52 started yet, the promotional war hasn't1:04:55 begun, and there are two aspects that I1:04:58 find regrettable. First, my generation1:05:02 never imagined that China would have a1:05:06 technological edge when I was growing1:05:10 up. Yes. Even in the field I worked in,1:05:13 uh Well, yes. At one time, yes. It was1:05:17 a field that China eventually surpassed1:05:21; for instance, if we talk about film1:05:24 special effects, China was really bad1:05:27 at CG in the old days. They were really1:05:31 bad, but what happened was after going1:05:34 through various economic upheavals,1:05:37 China eventually bought a major CG1:05:40 company called Digital Domain. Yes.1:05:44 China bought it, and looking back, it's1:05:46 not like the technology was impossible1:05:47 to catch up with. It was high-tech at1:05:50 the time, involving simulations and VFX1:05:53, things thought to be possible only by1:05:56 ILM in the U.S., or motion capture used1:05:59 for Avatar. Back then, those were truly1:06:03 shocking technologies, but from today's1:06:06 perspective, they're just things you1:06:08 can catch up on if you work hard.1:06:12 That's the kind of thought that comes1:06:13 to mind. So, having gone through that1:06:16 period, I actually had an identity1:06:19 crisis; it took a long time just to1:06:21 accept that China could be better than1:06:24 Korea at something. Then, at some point1:06:27, I started wondering, "Should I learn1:06:30 Chinese?" But the feel of the Chinese1:06:34 language, I just couldn't learn it. I'm1:06:38 not particularly gifted with languages,1:06:40 so I really couldn't pick it up. I1:06:42 remember hitting those kinds of1:06:44 limitations. So, what I'm saying is—1:06:47first, there was a time like that—I1:06:49 want to share two perspectives at once,1:06:51 so please don't get them mixed up.1:06:53 First, Korea was doing really well. Yes1:06:56. Korea was extremely good. I hope you1:06:59 don't forget that there was such a time1:07:01. It’s like saying the Goguryeo1:07:04 cavalry was incredibly strong; I just1:07:07 want to mention that there was a time1:07:10 like that. But paradoxically, there's1:07:13 another concern I have from the1:07:15 opposite viewpoint. Surprisingly, many1:07:18 people think that China is really1:07:20 incompetent. A prime example is the1:07:23 belief that they can bring a model up1:07:25 to that level through distillation.1:07:28 There are actually people who believe1:07:30 they just stole it. I guess there1:07:32 really are people who think that way.1:07:35 But if you could actually distill it to1:07:37 that extent, that would be an amazing1:07:39 technology. It doesn't work like that.1:07:42 Quite the opposite. Yes. Because of1:07:45 that, uh, yes. So, so now that...In a1:07:52 way, it’s quite...I feel like there1:07:58 are many people who actually undervalue1:08:00 them. Yes, I get that feeling. So, I1:08:04 just wanted to share that. Now, back to1:08:08 the main topic: while researching the1:08:11 data today, I began to doubt whether1:08:14 Korea has really fallen that far behind1:08:18. Yes. So, let's try to investigate1:08:21 this a bit more objectively. After all,1:08:23 do you think I watch different YouTube1:08:25 videos than you do? I watch the same1:08:27 things you watch every day. And the1:08:29 news that circulates in group chats—1:08:31we all see the same things, don't we?1:08:33 Yes. Because we look at the same things1:08:35, you and I are no different. So when I1:08:37 research, I set it to analyze articles1:08:40 from the US, China, and Japan, and just1:08:43 run it through. And as I research, I1:08:46 end up encountering old memories. So,1:08:50 if you look at the contents on the1:08:52 slide, we were actually quite fast. Yes1:08:54. Faster than you'd think. We did it1:08:57 quickly, and we did a lot of this stuff1:08:59. So, it's not like we were bad at it.1:09:01 But the point where we started falling1:09:04 behind, I don't remember exactly, but I1:09:08 think there were a few reasons. This is1:09:09 Claude's analysis, and it's a part I1:09:12 also agree with. So, if you look, Ah,1:09:16 there are still so many slides left. Oh1:09:17, seriously. I'm sorry for making so1:09:22 many slides. Let me finish this quickly1:09:24. So, I dug into why we in Korea don't1:09:29 have companies like OpenAI or Anthropic1:09:34 for LLMs, wondering if it was due to a1:09:38 lack of GPUs. And it turns out that the1:09:44 ChatGPT moment was in 2022. But before1:09:48 that, GPUs weren't actually that scarce1:09:50. Based on the standards at the time.1:09:51 Based on the standards at the time. Yes1:09:53. It wasn't a lack of GPUs, that is.1:09:55 But then, uh...It seems there was some1:10:00 back-and-forth about why we fell behind1:10:01 later on. Without me knowing. Back then1:10:04 I was working at a company and was1:10:06 probably too busy to pay attention, so1:10:08 I didn't really look into or care about1:10:10 these things like I do now, but it1:10:12 seems there was some friction. Maybe it1:10:15 was like, "Why did you just sit on your1:10:17 research instead of developing it?" In1:10:19 any case, it's true that we fell behind1:10:21. But looking back, people say it1:10:25 wasn't because of a GPU shortage. Based1:10:29 on the standards of that time. But it1:10:30 really was like that. It was. Actually,1:10:33 even back then, memory and GPUs weren't1:10:37 that expensive, and did the big Korean1:10:40 IT companies—what you call "Big Tech"1:10:43 in Korea—lack the money? That wasn't1:10:46 it either. In fact, they paid their1:10:48 employees so well that I felt1:10:49 relatively deprived and thought about1:10:51 leaving for Silicon Valley myself,1:10:52 didn't I? They had plenty of money? Yes1:10:55. So, it was a distortion of memory.1:10:57 The narrative that we couldn't do it1:10:59 because we lacked GPUs wasn't true.1:11:02 It’s true that they’re scarce now,1:11:03 right? It might be true that they're1:11:05 scarce now, but back then, that wasn't1:11:07 the case. Even I used to run1:11:09 simulations at supercomputer centers1:11:12 back in the day. So, ah, that wasn't it1:11:15. I heard that. But if I had to point1:11:18 out one problem, it's that they keep1:11:25 stopping. Yes. They kept shutting1:11:27 things down. I definitely remember this1:11:29. Samsung said they would build their1:11:32 own LLM, then suddenly scrapped it and1:11:34 used OpenAI's. But I don't know. Maybe1:11:39 Samsung Electronics decided there was1:11:41 no need to pour money into software1:11:44 since they make plenty from1:11:45 semiconductors; I don't know why that1:11:48 company made those decisions. It’s1:11:52 been so long, but I had forgotten about1:11:55 it until I saw the data just now. Kakao1:11:59 Brain also scrapped their project once,1:12:01 right? So, yes, the people who were at1:12:04 Kakao Brain back then either left, were1:12:07 absorbed into the main headquarters,1:12:09 and so on; then HyperCLOVA said they1:12:11 were doing something with LLMs, so I1:12:13 signed up back then, thinking, "Let me1:12:16 try it, too." I wanted to try an LLM1:12:19 that was good at Korean. I wanted to do1:12:21 something. That’s how I ended up1:12:23 trying to use it back then, even1:12:25 submitting an application. But1:12:28 strangely, it kept getting delayed, and1:12:30 they did seem very busy. Then there was1:12:33 all this talk about whether they’d1:12:34 give us credits or not, and then I1:12:35 couldn’t reach them at all. They just1:12:37 left, and a few years later, they1:12:39 disappeared. Yes. I'm not sure if they1:12:42 shut down or what, so, uh. And now, at1:12:47 this point, after the first round of1:12:49 the Dok-Pa-Mo has passed, and with1:12:50 chips making massive profits and talk1:12:52 of Physical AI coming up, people are1:12:54 saying what Korea should be doing. Then1:12:57 Jensen Huang comes by, buys everyone1:12:59 chicken, and so on. Some people showed1:13:04 up saying, "Should we just quit, since1:13:05 we probably can't make it, and focus on1:13:07 something else?" When people started1:13:09 saying things like that, uh. I was a1:13:14 bit taken aback. Of course, I have1:13:16 absolutely no personal interest or1:13:18 stake in the people doing things like1:13:20 Dok-Pa-Mo. Absolutely no vested1:13:22 interest. Even if they do a great job1:13:25 with an LLM and build a company like1:13:27 Anthropic, I don't get a single penny1:13:29 from it; I have no relationship with1:13:31 them at all. But, it is true that1:13:36 things aren't consistent and keep1:13:38 getting cut off. Yes. It is true that1:13:40 they aren't connected and keep getting1:13:41 interrupted. But I actually looked into1:13:43 the Luda case. It turns out Luda1:13:46 actually got some investment from, I1:13:47 think it was KT or SK—not sure which1:13:49—and there's talk that they’re1:13:51 still profitable, but I ran out of time1:13:53 while investigating that, so I couldn't1:13:55 finish. So, for those interested, it1:13:57 might be worth looking into. Anyway, so1:13:59 in August, now. Uh, the second1:14:05 evaluation is in August 2026, and there1:14:07 aren't many days left. They said1:14:10 they’d start from the beginning of1:14:12 August, keeping three teams, then1:14:13 another two teams in the winter, or1:14:15 something like that. Uh. Anyway, I just1:14:20 really hope they keep at it1:14:22 consistently. I mean, I think it1:14:25 carries the meaning of preserving1:14:27 language as a cultural heritage, and I1:14:30 really hope an LLM that can use1:14:32 beautiful Korean comes out. Yes. I1:14:34 mentioned this earlier. They call it1:14:38 English, but that "English" isn't1:14:40 natural language; it's LLM-speak.1:14:45 LLM-ese, yes, the language model style;1:14:49 that way of speaking is spreading at a1:14:52 very fast rate. You all are like that,1:14:55 too. So, if you look at things like1:14:57 Threads, these days when I write a post1:14:59 on Threads, I mark it if I wrote it1:15:00 myself. I write, "Written by me in my1:15:05 own voice." I write it like that. Yes.1:15:09 That’s how I’ve been doing it. Yes.1:15:12 So, it has meaning in that sense as1:15:13 well, and we have to look at this in1:15:15 the long run. We have to look at it in1:15:17 the long run. We have to look at it in1:15:18 the long run. We have to look at it in1:15:19 the long run. Well, if you look at1:15:21 those projects that were started and1:15:23 stopped earlier, they're actually all1:15:25 from large corporations. It's1:15:27 inevitable when it's done by a large1:15:28 corporation. I mean, you all know well1:15:30 about the office politics within large1:15:32 companies, right? Yes. I'm sure you do.1:15:35 Because of those things, it's really1:15:36 not easy to keep pushing for something1:15:38 in a big company if it's not producing1:15:39 results. So, in that sense, Korea seems1:15:44 a bit vulnerable because everything1:15:46 tends to revolve around large1:15:48 corporations. But companies like LG1:15:53 have been doing it for quite a while,1:15:55 and Upstage has also been at it for a1:15:57 long time; there are others I'm not1:15:59 familiar with because I haven't paid1:16:01 much attention, but everyone is at1:16:03 least consistently working on it.1:16:07 Companies like Kakao, for instance—1:16:09rather than expecting something to1:16:11 appear in the short term that will1:16:13 crush Anthropic, we should take a1:16:15 long-term view and focus on cultivating1:16:17 talent. We may not be as wealthy as1:16:22 Silicon Valley or the U.S., but1:16:24 shouldn't we take a long-term1:16:27 perspective and build things up by1:16:30 gathering resources bit by bit? Yes.1:16:35 That's what I think. Yes. That's my1:16:38 personal wish. Now, I've taken a look1:16:43 at a research paper that was released1:16:45 recently. I read it, and it has some1:16:47 distinct features. Since there are1:16:48 features to note, let's take a quick1:16:49 look at the paper. First, this model1:16:51 regarding attention residuals is a team1:16:53 project. This is a paper for a Chinese1:16:56 model, and I've been focusing more on1:16:58 older papers lately, so I haven't kept1:17:00 up as much with recent ones. Regardless1:17:03, just take a look to see that this is1:17:05 the kind of style they use. The trend1:17:08 is moving toward making the structure1:17:10 itself operate more dynamically by1:17:13 connecting components in the model more1:17:15 fluidly and selectively. Just note that1:17:19 for now. Now, this is the Solar Open 21:17:22 technical report. Yes. The Solar model1:17:24 came out just recently. It was July 2nd1:17:27, so it's been about a week or so. Well1:17:31, this model has various aspects, but1:17:33 the part that stands out to me in the1:17:36 paper is what I consider a strategic1:17:39 choice by the developers. Furthermore,1:17:43 I don't for a second believe that the1:17:46 people making Solar or LG's Exaone are1:17:48 any less professional or capable than I1:17:51 am. So I believe they all made the best1:17:55 possible choices. Nevertheless, from my1:17:59 perspective—as just an ordinary1:18:02 person—one thing I find a bit1:18:04 regrettable is that when you read these1:18:06 papers, they always emphasize methods1:18:09 to save money when the budget is1:18:11 limited. The lack of GPUs. So, uh,1:18:17 should I call it disposal? It's not1:18:23 about crushing SOTA, but rather about1:18:26 scoring well within a limited scope. I1:18:30 get a bit of a college entrance exam1:18:32 vibe, and, well, it’s my personality.1:18:36 I’ve always been like this, even back1:18:38 when I wrote papers, aiming for things1:18:41 others didn't do...anyway, this is a1:18:43 personal YouTube channel, and you all1:18:46 seem to like it when I share my1:18:48 opinions. So that’s why I’m1:18:50 mentioning it. Yes, there’s that1:18:53 aspect, and then there are benchmarks1:18:57 and, well, architecture diagrams and1:19:00 such. And if you go further down, oh, I1:19:06 shouldn't laugh. Wow, this paper is1:19:12 really long. It’s even longer than1:19:14 the KIM paper. It keeps going down. Oh,1:19:20 they worked hard on the paper...right,1:19:23 they used document processing1:19:25 capabilities as an example. So it seems1:19:28 to be about detecting things like1:19:30 overseas remittances, and if you go1:19:31 down further, is it government1:19:33 documents? Or...It seems to be good at1:19:38 finding things related to document work1:19:39. Things like data collection,1:19:42 collecting unusual incidents, documents1:19:44 like that, and there’s another one.1:19:52 What’s here? Korean Office Work Agent1:19:56 Task Expert—so, assisting with office1:20:01 work. Oh, whoops, I bumped the mic. So,1:20:08 assisting with office work, that seems1:20:11 to be the goal. Yes. Yes. Is the mic1:20:18 broken? Well, it should be fine. Yes.1:20:22 So, Korean Office Work Agent Task1:20:25 Expert. But, wait, it’s medical. The1:20:27 content is, well, yeah. So they are1:20:30 putting this forward. Yes. They are1:20:33 emphasizing things like this. Things1:20:36 like an annual plan, like a one-year1:20:39 energy-saving factory plan, so in some1:20:42 respects, it’s definitely necessary.1:20:47 Yes. AI will enter every job that uses1:20:51 a computer. The AI mentioned here1:20:54 refers to an LLM-based assistant1:20:56 service. Yes. Assistant, AI agent,1:21:00 assistant—something like that, so as1:21:04 a helper for general office work, its1:21:07 value is extremely high. Its value is1:21:11 extremely high, but now, uh...Yes, they1:21:16 are emphasizing that it helps with1:21:17 things like this. So, I think it might1:21:19 be good for getting high scores during1:21:22 evaluations. Yes, that’s what I1:21:24 thought. Now, what else is there? X41.1:21:28 Uh, What were the features of X41 again1:21:33? There are so many slides that I'm1:21:36 starting to forget. Uh, X41...What was1:21:43 it? The advantages include the X411:21:45 model structure, and then RMS Norm,1:21:48 Global Attention, and so on. Well, for1:21:51 those who have studied LLMs, it’s1:21:53 just, "Oh, I see." It’s all stuff1:21:55 I’ve seen somewhere before. LM Head,1:21:57 it feels like a textbook. All right,1:22:01 shall we scroll down a bit more? There1:22:03 are some formulas and such here. Ah,1:22:05 there are advantages to the EXAONE1:22:07 model, let me check the slides again to1:22:09 be sure. There are specific1:22:11 characteristics of the EXAONE model in1:22:13 the slides, it definitely has its own1:22:15 unique features. Hmm. I think it was1:22:19 quite good at Korean. Oh, here it is.1:22:21 It's good at Korean. It even does1:22:24 Vietnamese. German, Japanese, Spanish,1:22:27 and things like that are here. Next is1:22:32 K-AI by SK Telecom. They are big on1:22:37 scale. I read their paper, and it’s1:22:42 about how they organized their data1:22:45 well; it focuses on data processing1:22:48 frameworks for constructing1:22:51 high-quality pre-training datasets. Yes1:22:55. Rather than just focusing on the1:22:57 paper itself. I’m not sure if this is1:23:00 relevant, but I think it was last year1:23:03—not this year, but last year at CES1:23:05—when I visited the SKT exhibit, they1:23:08 had images of their data center1:23:10 displayed. They are large in scale. I1:23:14 think they really wanted to emphasize1:23:16 that. That they can push forward with1:23:18 sheer volume. I remember that being a1:23:20 key point they highlighted. So, looking1:23:22 at the paper as well, it seems they are1:23:26 emphasizing how they handle data1:23:28 processing. Now, Next is Naver's1:23:34 HyperCLOVA. Ah, what was the feature of1:23:37 this one again? This is Ah, the multi1:23:40...what is it? Uh, vision, what we call1:23:44 vision multi-modal. Yes. They are1:23:46 working on these, and frankly, vision1:23:48 is very important if you want to move1:23:50 into things like robotics. Right? What1:23:56 else is there? Ah, Krafton—you’ve1:24:00 all seen the recent demo from Krafton,1:24:02 right? The one where an AI acts as a1:24:06 companion within the game—that1:24:08 Battlegrounds game—I only saw it on1:24:11 YouTube, but the character next to you1:24:14 actually talks like they’re really1:24:17 playing the game with you. So, it's a1:24:22 spoken speech technical report, but1:24:24 rather than just being pure AI, it1:24:30 feels more like a conversational agent;1:24:33 there's a diagram below showing it1:24:35 going through an LLM. It goes from here1:24:40 to there like this. It seems a bit like1:24:45 that, but there isn't much talk yet1:24:47 about the inherent reasoning1:24:48 capabilities of the LLM itself. Now,1:24:52 next, this person—their writing style1:24:57 actually sounded like an AI. Is that1:25:02 right? Where? Ah, is it not this person1:25:07? If I say the wrong thing, it might1:25:10 just...Is it not this one? Something1:25:15 else—oh, this is all CLOVA—my1:25:17 mistake, I misspoke, how frustrating.1:25:19 Yes, I apologize. Yes, instead of this1:25:21 person's post, I recommend you all take1:25:23 a look at things like the development1:25:26 story of Ko-LLM 2. Part 2, this tone of1:25:28 voice, data with different styles pulls1:25:30 the model in. This is a bit of an AI1:25:35 tone. Yeah. It's interesting, so the1:25:37 content is good. The content is good.1:25:38 So let’s separate the conflicts and1:25:39 combine the performance. Ah, this is1:25:41 exactly that AI tone, isn't it? Yes. Ah1:25:44, the content is good. It’s not that1:25:46 the content is bad; the content is good1:25:48, but it’s a new kind of Korean. So1:25:51 I’m not sure yet if we should view1:25:54 this as an evolution of Korean created1:25:57 by US LLMs or just treat it as a1:26:00 curiosity. I don’t think that’s a1:26:03 decision for me to make. Ah, not "1:26:04Pangyo dialect," but "Human LLM-ese,"1:26:06 that makes sense. "Human LLM-ese"1:26:08 sounds more plausible than "Pangyo1:26:11 dialect." That’s actually not bad.1:26:13 Yeah. That kind of feeling. Uh, yes. I,1:26:15 ah, the content is good. The content is1:26:17 good, but ah, this is like that strange1:26:19 AI from back in the day, what was it,1:26:21 Gwi-hyun? That wrote novels and stuff.1:26:23 I don't know if we should take this as1:26:25 using emoticons and such, or if we1:26:27 should view it as an innovation in the1:26:28 Korean language. But anyway, the1:26:30 content is good. The illustrations are1:26:32 well done. The content is good, but the1:26:34 tone was a bit funny. I mean, to create1:26:37 materials like this, a person can't1:26:38 draw them all. You have to get the AI1:26:41 to do it now. It is true that you have1:26:43 to make it do it. But, uh, at times1:26:45 like this, it would be great if1:26:46 linguists, novelists, literary writers,1:26:48 or broadcast scriptwriters, people in1:26:50 those fields, could contribute their1:26:52 efforts. That’s the kind of thought I1:26:55 had. This is just a side note. Now. Uh,1:26:58 what else was there? Saw one, ah,1:27:01 something related to "thinking" came1:27:04 out, and I wondered whose it was—it1:27:06 was GoldStar, "LuckyStar." It’s funny1:27:09 to be talking about this while looking1:27:11 at a research paper, but do you know1:27:13 that? Do you know that GoldStar used to1:27:16 be "Lucky-GoldStar"? Lucky and GoldStar1:27:19 merged, so the model name is "LuckyStar1:27:22." Ah, did they really have to do that?1:27:26 The paper is good. Thinking in English,1:27:30 answer in Korean, yeah, adaptation,1:27:32 multilingual, tool-using agents—the1:27:34 content is good, the content is good,1:27:36 but as soon as you see "LuckyStar," you1:27:39 know, "Ah, it's LG, LG is doing1:27:41 something." All that is fine, but1:27:44 aren't they being a bit too loyal to1:27:46 the company? Yeah, just joking. Anyway,1:27:50 they’re doing things like this.1:27:52 Everyone is working hard. I think1:27:56 it’s interesting. Now, Next, KT1:27:59 suddenly came out with a "Responsible1:28:01 AI Technical Report," and well, yes, I1:28:04 think KT had something regarding LLM1:28:06 development earlier, but yes, it’s1:28:09 always good to do these things together1:28:11. Now, back to the slides, they are1:28:15 doing things like this. Yes, they1:28:18 aren't just sitting around. Someone is1:28:19 working on it. I think it would be good1:28:21 for you to know things like this. At1:28:23 this rate, someone will find a1:28:26 breakthrough. Once that path is opened,1:28:28 things will start flowing smoothly. But1:28:30 it doesn't always have to be done1:28:32 internally, and actually, if a good1:28:34 paper comes out of China...We use it,1:28:36 and it works well. You know, we might1:28:38 even borrow some ideas that way. Isn't1:28:40 the ecosystem all about learning from1:28:43 each other? Right, yes. Now, the second1:28:47 round of evaluation for the Open LLM1:28:50 Leaderboard hasn't started yet. So,1:28:53 once the second round of evaluation1:28:54 begins, news will start pouring out. I1:28:56 think it would be good to look at it1:28:58 together then. We looked at the Solar1:29:02 model earlier, and there was a lot of1:29:06 talk about using MoE to increase1:29:09 efficiency. Objective evaluation for1:29:13 this will likely proceed once the1:29:15 second round starts. Naver was1:29:19 eliminated in the first round, wasn't1:29:21 it? I think they were eliminated1:29:24 because they started with a Chinese1:29:27 model, which was probably the right1:29:30 choice at the time, and from today's1:29:33 perspective, I don't think starting1:29:36 with a Chinese model is necessarily1:29:39 that bad. Back then, they had to1:29:43 eliminate someone, so...I don't think a1:29:50 model that used a part of Q&A was1:29:53 necessarily bad or deserved to fail,1:29:56 but from a current perspective, I think1:30:00 it's better if you just make something1:30:03 well, regardless of the method. Well,1:30:07 that's all in the past now. Yes. Kakao1:30:12 isn't doing it, right? Kakao isn't1:30:14 doing it, but they seem to be working1:30:16 hard. It's fascinating that they're1:30:23 doing well even though it's not Kakao,1:30:26 but you know, things go back and forth.1:30:31 For instance, ChatGPT was integrated1:30:35 into KakaoTalk. They didn't use their1:30:41 own model, so...Yes, I think it's1:30:45 because they're a large corporation.1:30:47 Yes, I think that's just how large1:30:48 corporations are. It feels like there's1:30:50 some regret in those kinds of1:30:52 decision-making processes. Do they use1:30:54 their own model as well? I don't really1:30:57 use LLM services in KakaoTalk. Is there1:31:01 one? I don't even know if it exists.1:31:03 Yes. I only use it for work. Moving on,1:31:08 Krafton's voice model, playing with an1:31:13 AI character in Battlegrounds looked1:31:17 pretty impressive. I haven't tried it1:31:22 myself, but I remember everyone saying1:31:24 it looked quite fun. Now. Ah, it's1:31:34 slowly coming to an end. Alright, the1:31:37 slides are slowly coming to an end. Now1:31:40, Now, this part seems quite important.1:31:47 Uh, there’s a part I don't quite1:31:50 understand, a part I couldn't really1:31:52 grasp. They had blocked the license.1:31:59 Yes. Why—so, back when I uploaded1:32:05 those videos about making LLMs a year1:32:08 ago, you know? I’m not sure if I1:32:12 mentioned it back then or if it was1:32:15 edited out, but what I desperately1:32:18 wanted at the time was a model that was1:32:21 good at Korean, not English. If I1:32:24 wanted something good at English, I’d1:32:25 just use something else. But I needed a1:32:28 model that was good at Korean to create1:32:32 lectures or, you know, to do my own1:32:34 thing on YouTube, but there were no1:32:37 models good at Korean...and at that1:32:40 time, the model that was best at Korean1:32:43 was ChatGPT. I mean, it was1:32:46 overwhelmingly superior. Yes. So, I1:32:49 thought if there was at least something1:32:52 available for free, I could tell1:32:54 students to play around with it or try1:32:56 it out, but everything was restricted.1:33:00 The license was. So I found this to be1:33:03......a bit disappointing. Yes. It was1:33:08 disappointing that they didn't seem to1:33:10 think about building an ecosystem while1:33:12 moving forward. It's still hard to1:33:15 understand why China, despite being a1:33:19 more closed society, is more open1:33:23 regarding technology. Yes. It’s still1:33:27 something I can't quite wrap my head1:33:28 around. But as far as I know, if you1:33:30 receive funding from somewhere like1:33:32 ETRI, you’re supposed to make it1:33:34 public if it turns out well. I suppose1:33:37 so, there must be something like that.1:33:39 Yes. So, I wish they would contribute1:33:43 to growing the ecosystem. At the very1:33:47 least, a middle schooler could just1:33:49 take a model, play with it, and by the1:33:52 time they're 17, they could build1:33:54 something new, right? Kids in Korea are1:33:58 smart, too, so I wish there was1:34:00 something available for them. And it's1:34:03 not like young kids can just use1:34:05 English models well and excel. It’s1:34:08 easier to use a Korean model than that.1:34:11 Are we going to teach all the young1:34:12 kids English first? Or teach them1:34:14 Chinese first? That's not it, is it? So1:34:17, it would have been better if they had1:34:20 loosened the license or provided1:34:22 tutorials to play with, contributing to1:34:24 the ecosystem, which is why it's a bit1:34:27 of a shame. Things are different now.1:34:31 Now, you can just ask Claude and it1:34:34 writes everything for you. So,1:34:36 downloading, running models, and all1:34:38 that can be done easily now. Yes. And1:34:41 they say Korean performance is actually1:34:43 still better in Chinese models. But1:34:46 this will probably come up again once1:34:48 the second phase of the evaluation1:34:50 begins. Since it’s an issue that1:34:52 experts will have to re-evaluate then.1:34:53 Yes, so Korean language performance as1:34:56 well Yes, those aspects will likely be1:35:00 discussed then, and it's a bit of a1:35:02 shame in that regard. Yes. Next,1:35:06 looking at the datasets, um. China1:35:11 There were Korean data mixed into the1:35:14 datasets that came from their side, and1:35:19 Then they created things like OCR1:35:21 technology, did massive crawling, and1:35:24 so on; it was absolute chaos. Lately,1:35:28 it’s actually become quieter; people1:35:30 don’t really talk about data as much1:35:32 these days. In my view, although there1:35:34 is no evidence for this. There's no1:35:37 concrete evidence, but they are1:35:38 probably running reinforcement learning1:35:40. Since they are using our actual usage1:35:44 for reinforcement learning, I think1:35:47 Claude is confidently taking our data1:35:50 in exchange for paid services. In fact,1:35:54 when using Claude, you can choose1:35:57 whether or not to provide your data for1:35:59 better service. And as far as I know,1:36:04 the Pro plan itself is priced on the1:36:07 condition that they use the data for1:36:09 their training; companies, uh, like1:36:12 that...I understand that plans where1:36:18 Anthropic doesn't use your data for1:36:20 training are much more expensive. Given1:36:25 that aspect, I think Anthropic has1:36:27 likely almost solved the data supply1:36:30 issue. They are running it heavily1:36:34 through reinforcement learning, and1:36:35 even I get frustrated and tell Claude,1:36:37 "How can you be this incompetent?""Do1:36:40 it like this immediately; how can you1:36:41 only think of this much?" That in1:36:42 itself is a massive amount of data. Yes1:36:45. Wouldn't that be the case? Wouldn't1:36:48 the pipeline be structured that way?1:36:50 And those Chinese models that let you1:36:52 use them at low costs, you know. That1:36:55 will become a tremendous asset. It will1:36:58 be difficult to keep up with that. It1:36:59 will be extremely difficult to catch up1:37:01. So, if Koreans were to all jump in1:37:06 and do that. Then it could actually be1:37:09 even better. Because Koreans, in any1:37:11 case, have a high level of education.1:37:13 And our age demographic is quite high.1:37:16 But those people still have great1:37:17 language skills. If they argue about1:37:21 things like "which stocks should I buy,1:37:24" that could be used for reinforcement1:37:27 learning, so I think that pipeline is1:37:30 urgent. But there’s no need to try to1:37:33 get too far ahead, and I'm sure they'll1:37:35 do all of that. Yes. Since they are all1:37:38 people who can think of what I'm1:37:40 thinking, I believe it will go that way1:37:42. Now. And surprisingly, they say GPUs1:37:49 aren't actually the bottleneck. Since I1:37:54 don't know exactly how many GPUs are in1:37:56 Korea, Yes. I don't know the exact1:38:00 number of GPUs in Korea, but regardless1:38:02, they say the GPU itself isn't the1:38:04 bottleneck. Since there’s talk that1:38:06 power, cooling, and construction—1:38:08meaning the energy supply needed to run1:38:10 the GPUs—is the real bottleneck,1:38:12 Korea is very good at handling those1:38:14 things once they commit. Yes. We’re a1:38:16 country that even builds nuclear power1:38:18 plants for other nations. Yes. Right.1:38:20 So, since we can do those things1:38:22 quickly, I expect that cycle to last1:38:25 through next year and the year after,1:38:27 and from 2030 onwards, it will be the1:38:29 time to actually generate profit on1:38:32 that foundation. So when that time1:38:35 comes, I believe another software cycle1:38:37 will arrive. It’s hard to imagine it1:38:40 won't, and the software sector has a1:38:42 very strong effect on job creation. The1:38:44 job creation effect is strong in1:38:46 software. Remember the coding craze1:38:48 back then? When money is injected,1:38:51 software is the best field for pumping1:38:53 up employment by hiring people. IT.1:38:57 Robotics is difficult. For example, can1:39:00 you quickly train someone to build1:39:03 robots through a 6-month boot camp? You1:39:06 can't. And things like law or medicine,1:39:10 those aren't easy either. Yes. Software1:39:13 is simply the best for pumping up the1:39:15 workforce. Also, the number of people1:39:19 using AI needs to increase. So,1:39:21 you’ve built power plants, data1:39:24 centers, and packed them with GPUs. But1:39:27 what happens if people don't use them?1:39:29 You go bust. You really go bust. So1:39:32 people must actually use them to1:39:33 increase productivity. They need to do1:39:36 something useful. They must be able to1:39:38 do something useful. And use that to1:39:41 boost the national growth rate. That is1:39:44 the plan; it is currently a grand plan1:39:46 for the whole world. China, the U.S.,1:39:49 Korea—I'm not sure about Japan.1:39:52 It’s a grand plan. So, I heard China1:39:55 says they've done it all, even putting1:39:58 power plants underwater, and I think1:40:01 that’s probably in the West Sea area.1:40:06 There’s talk like that, and well,1:40:08 there's everything. So anyway, if you1:40:10 look at the timeline, that’s how it1:40:12 will unfold. Yes, that seems to be how1:40:15 Wall Street is viewing it for now. Now.1:40:20 So, why are they making LLMs in Korea1:40:25—sometimes Claude and I don't quite1:40:29 align. And usually, I’m the one1:40:31 who’s right. So, do you remember how1:40:34 excited people were when the DogPha-mo1:40:37 project first started? Yes. Thinking, "1:40:41Wow, Korea is finally doing this too,"1:40:42 and, what was it? Everyone was so hyped1:40:45 about the so-called "Black-and-White LM1:40:46." But now, it has kind of lost its1:40:48 momentum. Since everyone is quietly1:40:51 working away, wouldn't it be nice to1:40:53 have some fun events again come August?1:40:57 Well, that's one thing, and since1:40:59 semiconductors are making so much money1:41:02, people are getting swayed toward that1:41:04. If we were so excited about building1:41:08 LMs, we should keep that focus, but now1:41:11 it's just, "Semiconductors are making1:41:13 good money." Perhaps it's people who1:41:16 want semiconductor stocks to keep1:41:18 rising, but there are also those saying1:41:20, "Let's give up on LMs and focus on1:41:23 semiconductors." Isn't that a bit too1:41:25 much? Then there's talk about "Physical1:41:29 AI," and I heard Samsung is doing1:41:31 something with humanoids now. Yeah,1:41:35 there's stuff like that, so things seem1:41:37 to be shifting back and forth a bit.1:41:40 But it will be interesting to see what1:41:42 happens when the second evaluation1:41:44 begins in August. Right. So, I believe1:41:59 everyone is working hard on it.1:42:03 Surprisingly, they've caught up in1:42:05 terms of scale, and since it's been a1:42:07 period of silence so far—well, you're1:42:09 supposed to be quiet before an1:42:10 evaluation, right? You shouldn't give1:42:13 your competitors any ammunition. So, I1:42:17 think some interesting stories will1:42:20 emerge when the second evaluation takes1:42:23 place in August. Anyway, that's all1:42:29 I've prepared for today. And then, well1:42:32...Yes, there's a course sale going on,1:42:37 so those of you who have been waiting1:42:39 to purchase the courses, now is the1:42:41 time. So, the discount details are1:42:46 organized right here. You should do1:42:49 some studying this summer. So, that's1:42:53 all the talk I had prepared for today.1:42:55 I'll take a short break, then come back1:42:58, check the chat, have a quick1:43:00 conversation, and wrap things up. Hello1:43:02. I'm Hong Jeong-mo. I started with1:43:04 computers because I liked fascinating1:43:06 things. I ended up becoming a1:43:08 university professor and started1:43:10 teaching, and I realized the most1:43:11 meaningful thing I could do was to help1:43:13 many students grow. I felt it wasn't1:43:16 right to just tell students to take on1:43:19 challenges. So, I also bet my life on a1:43:22 new adventure. If I had to pick just1:43:25 one thing you need for growth, it would1:43:28 be confidence. "I'm good at at least1:43:30 one thing: programming.""Hong1:43:32 Jeong-mo's Python Programming Overtake1:43:34 Course" is designed to give you that1:43:36 confidence that will support you for1:43:38 the rest of your life. Those who have1:43:41 seen our previous lectures have already1:43:43 started studying without hesitation,1:43:45 and the response has been explosive. We1:43:48 strictly distinguish between what you1:43:49 need to learn through explanations and1:43:51 what you need to develop through your1:43:53 own thinking, one small step at a time.1:43:56 I have meticulously edited the videos1:43:58 to reduce even one second of passive1:44:01 viewing time. I have prepared all the1:44:04 lecture notes so that you can start1:44:06 practicing immediately without1:44:08 hesitation at any moment while studying1:44:11. In this day and age, planning to1:44:13 finish learning all the grammar first1:44:15 before moving on to applications is1:44:16 like walking while leaving a perfectly1:44:18 good sports car parked next to you.1:44:20 Almost everything you can do with1:44:22 Python. In other words, I will guide1:44:24 you to the very threshold of1:44:26 professional fields in key areas of1:44:28 modern software technology. Above all,1:44:30 you will realize that ideas can spring1:44:32 from your own mind like this. I want to1:44:34 make something myself. You will have1:44:37 this amazing experience. Decide now1:44:40 whether you will overtake or be1:44:42 overtaken. All right, we are back. Now,1:44:48 if you have anything you would like to1:44:50 say, feel free to type it in the chat1:44:52 window. I have just taken a quick look1:44:56 at the chat window, and there are many1:44:59 people here who don't usually visit,1:45:01 and you've left some interesting and1:45:04 insightful comments. Yes. There are1:45:08 many interesting stories, and one thing1:45:10 that came up while I was analyzing is1:45:13 the talk about a talent shortage, which1:45:15 actually hit home. Yes. It is a painful1:45:18 point. The talk about a shortage of1:45:20 personnel. But personally, bringing1:45:23 someone over from the US with a 101:45:26 billion won offer—well, this is just1:45:29 my personal opinion. It's my personal1:45:32 view, but rather than paying 10 billion1:45:36 won to bring one or two people from the1:45:39 US, wouldn't it be more standard to1:45:42 foster new talent growing up in Korea1:45:45 and improve the ecosystem so things run1:45:48 on their own? If it's a good place to1:45:51 live, people will come over and settle1:45:54 down. I think the time has passed when1:45:58 we bring people in by offering huge1:46:01 salaries. I don't think even Silicon1:46:03 Valley offers that much anymore. Meta1:46:06 pulled a stunt back then, but the1:46:08 current situation is getting worse. In1:46:11 the case of Meta, only negative things1:46:14 are being said. I'm not sure what will1:46:18 happen next, but for now, it seems to1:46:20 have gotten worse, and well, anyway, I1:46:23...Anyway, I've thought about this for1:46:29 a long time, but ultimately, there are1:46:32 people who do well when they're young,1:46:34 in their early 20s, or even in their1:46:37 late teens. I wish we could go back to1:46:41 a time where you could just wander1:46:44 around, talk to experts, pick up things1:46:46 here and there, meet friends, laugh and1:46:48 giggle, occasionally have a little1:46:50 fight, or even feel a bit of jealousy1:46:53 thinking, "Why is he better at coding1:46:55 than me?" The reason I say I want to go1:46:58 back is because that’s exactly how it1:47:00 was when I was young and the game1:47:01 industry was growing. Game, back then,1:47:04 it was like that. So, for example, I1:47:07 think I mentioned this last time. We1:47:11 would go around looking at different1:47:12 companies, flocking together to hang1:47:14 out, and things like that existed. Yes.1:47:18 That’s how an ecosystem gets built,1:47:22 and then it moves toward an upward1:47:25 leveling; achieving that upward1:47:28 leveling is very important. The era1:47:32 where one brilliant person comes in,1:47:35 sits everyone down, and teaches them—1:47:37that era is over. It is over forever.1:47:41 The idea of a great master coming in to1:47:44 teach knowledge one by one, like a1:47:46 schoolteacher hitting kids on the1:47:48 calves—that kind of education is1:47:50 finished. An ecosystem cannot be built1:47:54 that way. That era is gone forever.1:47:57 Instead, it needs to be about young1:47:59 talent living vibrantly, trying this1:48:01 and that, experiencing trial and error,1:48:03 getting criticized, breaking things,1:48:05 and getting all excited, saying, "I’m1:48:07 going to do this!" Showing a bit of1:48:11 that immature side, that kind of1:48:13 atmosphere must be created to make it1:48:18 work. That is what I’m thinking. So,1:48:20 regarding the talent issue, my stance1:48:23 is firm. Yes. Yes, that’s what I1:48:25 think. Now, anyway, it’s interesting1:48:28 to see so many new people here today.1:48:31 Yes, it would be great if you could1:48:32 come again next time. And I saw an1:48:37 article yesterday saying Nvidia and1:48:39 Hyundai are working together on1:48:42 autonomous driving. Also, there were1:48:46 articles about KAIST and Nvidia opening1:48:49 an AI graduate school campus in Seoul,1:48:52 and Mistral recruiting in Seoul—or1:48:55 maybe Incheon—which we talked about1:48:58 in our group chat. Rumors were going1:49:03 around that they don't do coding tests1:49:05 but instead have you build some kind of1:49:07 neural network layer, and people are1:49:09 saying getting rid of coding tests is a1:49:11 good thing. Honestly, forcing a coding1:49:17 test when hiring a research scientist1:49:19 is a bit of a stretch. On the other1:49:23 hand, if you can’t even solve a1:49:25 coding test, maybe you shouldn't be1:49:26 doing this. That’s true, but if they1:49:31 really want to check how much you’ve1:49:32 studied, they might just ask you to1:49:34 build a layer. Yes. I think the1:49:36 situation has turned out like that. And1:49:40 that is what I’m thinking. So, the1:49:42 talk about human resources has been1:49:44 around for a long time, and, well,1:49:46 I’d like to talk about that. You see,1:49:49 when you look at Korean news articles,1:49:51 they keep counting in terms of quantity1:49:54, like they’re thinking in terms of1:49:56 the "100,000-man military strategy."1:50:00 But quantity is meaningless. I mean, if1:50:04 you bring in 1,000 mediocre people,1:50:06 does anything come out of it? No,1:50:07 that’s not how it works. Yeah. And,1:50:10 well, people who don't want to do the1:50:11 work but are forced into it just1:50:13 because they heard they could make1:50:14 money—do they really do it properly?1:50:17 You have to have a basic passion for1:50:20 the job to be good at it, and even if1:50:22 people who came just for the money are1:50:24 smart and some eventually get the job1:50:26 done, that’s not how it works. It1:50:30 doesn’t work that way. It’s the1:50:32 ratio that matters. They keep measuring1:50:35 people by quantity, but it's important1:50:37 to build an ecosystem. It is important1:50:39 to build an ecosystem. Yes. That’s1:50:41 how I see it. Right. Uh. So, a lot of1:50:49 people are talking about this and that,1:50:51 but for now, since I've prepared up to1:50:54 this point, I think it would be good to1:50:56 wrap it up here and talk more next time1:50:58. If you want to come to the group chat1:51:02 and just chat, you can join, but the1:51:04 group chat is really just for chatting.1:51:07 So, I hope you won't be disappointed1:51:09 that there's just a lot of chatting,1:51:11 and if you have anything else you'd1:51:13 like to talk about, please feel free to1:51:15 leave a comment on YouTube or send a1:51:17 message, as I am always reading them.1:51:20 Alright then, that’s all for today,1:51:22 and I’ll see you next time. Thank you1:51:24.
16,572 words · 2485 lines
![[선데이 라이브] 당신도 AI 시대에 살아남을 수 있다!](https://i.ytimg.com/vi/D95UC9U5ShA/hqdefault.jpg)
