Transcript of AI로 인한 혼란, 이제 끝났다
홍정모
0:00Yes, finally this chaos is coming to an0:01 end. All that talk about AI ruining0:03 everything or jobs disappearing. You0:05 know, all that stuff. Especially on0:06 Wall Street, I said it was already over0:08, and in reality, all that talk about0:10 AI replacing everyone has completely0:12 vanished. I believe the direction we0:14 are headed is now clearly set. They0:16 said all jobs would disappear, so I'm0:17 thankful they were wrong. Oh, I'm glad0:19 they were wrong. I'm glad, and they're0:21 glad they were wrong, too. Personally,0:23 I see this as one of the signs that0:25 much of the chaos is settling down. Is0:28 there something that many of you are0:29 greatly misunderstanding? There was a0:31 lot of talk going around that0:32 software-related jobs would be0:34 completely wiped out. The job sectors0:36 originally thought to be blown away by0:38 AI were those high-income, intellectual0:40 professions. Originally, that is. So0:42 people thought developers, lawyers, and0:45 consultants would be gone, and0:47 management actually had such0:49 expectations, leading to hiring freezes0:51 and the like. But as for what happened0:55 as a result, a report has now come out0:57 in the US. In the current situation,1:00 the chaos caused by AI is finally being1:02 cleared up, and we have already moved1:04 to the next stage. And the clear proof1:05 of that is that it has turned into an1:07 infrastructure war. It was originally a1:08 model war. After all the craziness of1:11 competing over models, it has now1:12 shifted to infrastructure. Why is that?1:16 For a while, frontier labs like OpenAI1:18 and Anthropic were ahead, creating a "1:21closed" state where they thought, "Only1:23 we can do this," but that has now been1:26 broken. In any case, the narrative has1:29 shifted to the idea that if you invest1:31 enough time and effort, you can catch1:33 up, so we are now in a spreading phase.1:35 A spreading phase. Running a cash flow1:38 deficit is usually a major blunder that1:41 gets a CEO fired, yet it's being1:43 accepted to such an extent that stock1:46 prices sometimes even rise. This is1:50 because AI technology will gradually1:52 become standardized, and as a result,1:53 it will spread widely to general users.1:56 It is literally moving toward a state1:57 where anyone can use it. It’s1:58 becoming clear that this is where2:00 everyone is looking. There is a2:01 difference between the pre-AI era and2:03 the post-AI era. Now, in the pre-AI era2:05, the era without AI, if you were2:07 developing software...It was a battle2:10 of stamina and patience. But now, those2:12 physical difficulties are almost2:15 entirely gone. So, when typing,2:18 creating loops, or even implementing2:20 algorithms, if I say, "Hey, I know this2:22," and explain how to do it...It writes2:24 the code itself, so those physical2:26 constraints are disappearing2:27 significantly. In the past, the skill2:31 of making software meant typing,2:33 memorizing all the syntax and2:35 algorithms, being good at cobbling2:37 things together, designing2:38 architectures, and pulling all-nighters2:41 for on-call duties when issues arose in2:43 the field, but that's not how it will2:45 be in the future. It’s not a way2:48 where a person has to heavily deplete2:50 their physical or mental strength to2:51 move forward. You’ve all been2:53 snowboarding or skiing, right? What is2:55 the most important thing then? It’s2:56 staying relaxed and not tensing up. You2:58 naturally come down by using gravity.3:01 The laws of physics are there to help3:02 you. So, as you glide down following3:04 the force naturally provided by physics3:07, what should you do? What do you need3:08 to be careful about? First, you have to3:10 follow the basic safety instructions so3:12 that you don't get hurt. What happens3:14 if you get hurt? You instinctively3:16 start to hate it. If you get injured3:19 early on because you didn't follow one3:21 very obvious precaution, you'll be3:23 afraid of it for the rest of your life.3:26 The point is to relax and control your3:27 movements. Say you’ve broken your3:29 wrist before. Then, subconsciously,3:30 your body just freezes up. And then you3:32 can’t do it. If you relax and glide3:34 down, you can perform all kinds of3:35 tricks, but it’s not done by your own3:37 strength. You just need to handle the3:39 control well, right? Even so, are3:42 people who snowboard weak? You still3:46 need to have basic physical fitness,3:47 right? Yes. So I think you can think of3:49 it in those terms. However, once you3:51 get the hang of it, it becomes very3:54 easy. That is why I am continuing to3:56 research deeply into how this can be3:59 taught. You still need to be as good at4:00 the basics as you were with the old way4:02. But it’s not to the point where you4:04 have to solve thousands of LeetCode4:06 problems like before. And for research4:09 positions, they have already started to4:11 move away from looking at LeetCode.4:14 Instead, they want to verify if this4:15 person truly has practical work4:17 capabilities; that is my hunch. What I4:20 feel is, for instance, when a new model4:22 like OpenAI's or Codex comes out, you4:24 sometimes get the feeling that "oh,4:26 this one seems to have less4:27 functionality," right? Where things are4:32 heading now is that, initially, they4:34 were diving in as if they were going to4:37 create a "super intelligence"—an4:40 Artificial Super Intelligence that4:42 knows everything, combining Einstein,4:45 Michelangelo, Leonardo da Vinci, Jesus,4:48 and Buddha all into one, a super AI4:50 that would save humanity. The process4:54 involves dumping so much knowledge into4:55 it that it becomes unmanageable. There4:57 is something called mechanistic4:58 interpretability. It analyzes how these5:01 LMs actually work. While doing that, we5:03 realize, "Ah, it thinks by going5:05 through these specific steps." Let’s5:08 keep the thinking patterns and strip5:09 out the knowledge. Yes. You can only5:11 shrink the model by focusing on its5:13 thinking patterns. So, knowledge is5:14 gradually being removed. The current5:16 trend is that what you knew yesterday5:17 is updated today, so you don't even5:19 know it. Instead, if I have documents5:21 I've researched, I can read them and5:23 get work done. So, what I feel is, "Ah,5:25 they're stripping out the knowledge."5:27 And they are definitely moving toward5:28 increasing the user base. That’s why5:30 they’re pouring a terrifying amount5:31 of money into building the5:32 infrastructure. It costs an insane5:34 amount, but they do it anyway, and5:36 there are all sorts of political issues5:38 to resolve. Despite that, they’re5:40 pushing forward because they’re5:41 convinced. Rather than just advancing5:43 AI tech, the priority right now is5:45 building the infrastructure. Because5:47 the technology itself is slowly hitting5:49 a ceiling. Of course, someone might5:52 make another leap in the near future,5:55 but investing in this infrastructure is5:58 definitely worth it. So, we are in a6:02 time of change. A time of6:03 transformation. Yes. From that6:06 perspective, let’s not get obsessed6:08 with using specific AI models like6:09 Claude or Codex. Yes. Instead, you must6:13 not forget what your goal is. AI has a6:15 tendency to make people very confused.6:18 Since it knows so much and answers6:20 anything you ask, you end up forgetting6:22 the essence while talking to it. Don't6:25 get lost in the confusion; find your6:27 way back to yourself. It is just a tool6:30. It is not a person. It can't earn6:34 money for me on its own. However, it is6:37 a huge help in earning money and6:38 amplifying my abilities by over ten6:40 times. So, it can multiply the6:42 capabilities of a skilled person6:43 several times over. I am certain of6:45 this. As I said again, you must not6:48 forget the essence of this. Yes. When6:52 Google or Naver first appeared, people6:54 thought, "Internet search is convenient6:56," but it wasn't that big a deal, yet6:57 look what happened a few years later?7:00 When bugs appeared, we just copied and7:02 pasted them into Google to solve them.7:04 I stopped thinking for myself. You7:06 cannot resist the tide of the times. AI7:09 is a very broad term, after all. Since7:15 the term "AI" is being used for7:17 marketing and exaggerated promotion to7:20 reach people, it seems you all are very7:24 confused. When it comes to LLMs, is7:27 there even a single person watching7:29 this who thinks they won't use one?7:31 Another thing to consider is that we7:33 don't know what the time lag will be.7:35 That’s why Wall Street keeps warning7:38 about over-investment, looking at past7:40 examples. The prime example is the7:43 internet; we laid down so much of it,7:46 but what happened? We had the IT bubble7:48, and it crashed, didn't it? Then,7:49 after things were sorted out, it really7:52 took off. That was one instance, and a7:54 previous example would be electricity7:56—nobody lives without electricity now7:58. Back in the day, when Edison first8:01 laid power lines, there was that8:03 incredibly turbulent history. They8:06 fought a war over power transmission,8:08 and even after that, while it’s8:10 obvious to use electricity now, the way8:13 of using it was different back then.8:16 For instance, factories are much easier8:18 to run on electricity now, right? But8:21 back then, factories were designed for8:23 steam engines, so it took time to8:25 completely switch over to electrical8:27 systems. So, what people are worried8:30 about in the AI era is, first off, the8:33 lack of power and data center8:35 infrastructure; that part is true. And8:37 I think it’s still a bit too early8:40 for local LLMs right now. The hardware8:42 really needs to get faster. I think8:45 it’s still a bit early for local LLMs8:47 to spread rapidly. But if local LLMs8:49 run well, there’s no reason not to8:52 use them. Now, I still remember when I8:54 went to college. When I entered the8:57 KAIST dormitory, there was something at9:00 the entrance called a terminal. It was9:03 just a terminal. It wasn't a computer9:05 itself; it was a terminal connected to9:08 a server computer. How do we do things9:10 now? We find even carrying a laptop9:12 cumbersome, so we carry mobile phones9:13 instead. So I believe we will9:16 inevitably reach an era where we can9:18 run light LLMs on our phones. Oh, you9:22 actually already have LLMs on your9:23 phones right now. You know, the Galaxy9:25 phones process this and that already.9:27 They are already there. You just don't9:29 know it. Whether it’s Vibe Coding or9:31 anything else, if you create something9:33 and get a result, that's great.9:34 That’s a very good thing. And the9:36 real start is when you aren't satisfied9:37 with the result. This is where I see a9:39 problem with the Korean education9:41 system. It teaches that we should only9:43 encounter things that we like. The9:44 biggest problem is the mindset that you9:46 must always get the right answer. But9:48 what I want to tell you here is this.9:51 If you don't like the LLM result, that9:52 is the beginning. That’s where it9:53 starts. So, let's say I did some Vibe9:55 Coding. But the result comes out in a9:56 way I don't like. That is not the end.9:58 You have been educated with the mindset10:00 that it’s natural for everything to10:02 be perfect on the first try. You feel10:04 that choosing one out of five10:05 multiple-choice options is perfect.10:07 It's never perfect, and even those10:09 multiple-choice answers are just10:11 prepared by teachers to avoid10:12 complaints. Even so, finding a clear,10:15 definite answer in a math problem still10:18 feels neat, doesn't it? But reality is10:21 very complex and a bit messy. That is10:24 what reality is originally like. So,10:26 people who survive, grow, and succeed10:28 in reality are those who are good at10:30 dealing with messy situations. So, if10:34 you're suffering because you're unhappy10:35 with something unconsciously, or10:37 feeling so angry that you're cursing at10:39 Codex, or feeling mentally exhausted,10:41 you really don't need to be; I'll cheer10:42 you on to stay strong. Yes. That is10:45 absolutely not the end. That is the10:47 beginning. You say that because Claude10:49 does everything for you, it feels like10:50 there's nothing left to do, but you10:52 need to change your mindset. You have10:53 to change your mindset; when I was10:55 young, I learned calligraphy. I used to10:57 have very good handwriting. Now, my10:59 handwriting is the worst of the worst;11:00 it couldn't be worse. Because I don't11:02 use it. But does that mean learning11:04 calligraphy was bad? Not at all. I even11:08 learned the abacus when I was five or11:11 six years old, reaching level 3 or 2. I11:16 will never touch an abacus again for11:18 the rest of my life. Do I think11:21 learning the abacus at a young age was11:23 something I shouldn't have done? I11:26 don't think so. I had a habit of11:29 playing with the abacus when I was11:31 young, and it was fun. It had its own11:33 rules, and then playing the piano, and11:35 later typing on a keyboard was fun too.11:38 Then I started programming, and that's11:39 how it turned out. In an era where11:41 typewriters are useless, there's no11:43 need to be perfect at accurate typing.11:46 Typing accurately is extremely11:47 difficult. In the old days, what11:49 happened if you made a mistake while11:50 typing? You’d ruin the paper. But11:53 what’s amazing about human beings is11:56 this. If that’s your job. Then you11:57 really do it without making mistakes.11:59 Human power is incredible. Because we12:00 adapt, we just do it when told. You12:03 shouldn't prepare for the superpowers12:05 of the past, but for the superpowers of12:07 the future. But if you want a12:09 past-style method where answers are12:11 fixed and only expected things happen,12:14 then you aren't the talent of the12:16 future. That’s why I’m very12:17 concerned about that. Once those two12:19 things start to balance out, I believe12:21 hiring will increase significantly from12:24 then on. But the economy has to support12:25 it as well. Once the momentum builds up12:28 and the trigger for growth emerges, it12:29 will just skyrocket from there. Because12:32 once they find a way to make money by12:34 increasing headcount, they hire12:36 aggressively. Yes. First of all, you12:39 must never compromise on developing12:42 your own cognitive abilities. I’ve12:45 seen people suggest that you don't need12:47 to study anymore because AI does12:48 everything. I was quite shocked by that12:50. Think of it like an era where12:52 everyone uses autonomous driving. Does12:53 that mean you won't teach babies how to12:55 walk? Yes. Even kids these days ride12:58 electric bicycles. But would you buy a13:01 toddler an electric bike from the start13:03 instead of a tricycle? That’s not13:05 right. You need to have the foundations13:07. You need to have the foundations as a13:08 human being. Now, the meaning of a13:10 four-year university education is this.13:13 It means you need that level of13:16 education to be a cultured modern13:19 individual. You should not think of a13:22 four-year university education as a13:24 vocational training center. The meaning13:27 of you studying for four years is13:29 simply the foundation of life. That is13:31 why, especially in the software field,13:33 people from various disciplines are13:35 needed. I already told you that I spent13:38 six years in mechanical engineering,13:39 including my four-year degree and a13:41 master's. But the software field has13:43 always been filled by people from other13:46 backgrounds. And what is the reason for13:48 that? It's because they don't approach13:50 software as mere programming skills.13:53 You should approach it as engineering,13:55 but people ignore the engineering and13:57 just leave the coding skills, claiming13:59 that is what a job is. What I always14:01 say is that your major doesn't matter,14:04 and if you work hard at anything, your14:06 cognitive ability improves. That is14:09 what's important. Regardless of the14:11 field. Just because it's computer14:13 science doesn't mean it has nothing to14:14 do with mechanical, electrical, or14:15 liberal arts; that is absolutely not14:17 the case. Even Anthropic hires14:19 philosophers and things like that. Now,14:21 they even become employees at Anthropic14:23. They are philosophy majors, but this14:27 field grows by having a diverse group14:29 of people come together to create an14:31 ecosystem, absorbing, adapting, and14:34 evolving quickly. There's something you14:37 often overlook about Andrej Karpathy,14:40 who is frequently cited on YouTube or14:42 Threads. Do you think those people14:44 reached their positions only with a14:46 top-down approach? If you watch14:48 Professor Fei-Fei Li’s computer14:50 vision course, which is a Stanford open14:52 course. Everything is in Python. It's14:55 all Python, and Andrej Karpathy14:57 actually set the framework for that14:58 course. Everyone flocked to Python. Now15:01, what happened after that? Andrej15:03 Karpathy made a course on building an15:05 LLM from scratch in C, right? When you15:07 see someone who has succeeded, you need15:09 to look at how they started. I’d like15:13 you to spend just one week thinking15:15 about how you truly want to live. What15:18 is it that I want to do? Once you latch15:20 onto that one thing you really want to15:22 do, you start to generate immense15:23 energy. In fact, what you’re actually15:26 doing isn’t that important. So, you15:27 have to find that first. Yes. Once you15:29 find it, things will start to fall into15:30 place.
2,810 words · 429 lines
![[선데이 라이브] 당신도 AI 시대에 살아남을 수 있다!](https://i.ytimg.com/vi/D95UC9U5ShA/hqdefault.jpg)
