Transcript of AI가 못하는 것에 답이 있다
홍정모
0:01If you watch The Lord of the Rings,0:03 there's this scary, powerful villain0:05 who says no man can kill him. He0:07 interprets "man" as any human male. So0:10 he claims no man can kill him. He says0:12 no one can kill him. But then a0:14 princess shows up, having fought while0:16 pretending to be a man, and reveals she0:18 is not a man. "I am no man." And then0:21 she just kills him. I’m sure you’ve0:24 all seen this scene. It’s a very0:27 interesting example of how a twisted0:29 prophecy actually works. Now, you’ve0:34 heard a lot of talk about how AI will0:36 ruin everything, right? That there’s0:39 no need to learn coding because of AI.0:42 Or that there’s no need to study this0:43 or that because of AI. You’ve0:45 probably heard these kinds of things a0:46 lot. But I’ve always felt quite sorry0:48 whenever I heard such talk. I’ve0:52 thought that those kinds of statements0:54 are close to these twisted prophecies.0:58 In the generation slightly younger than1:00 mine, it was a trend to say that you1:01 could get into college without studying1:03. The scary thing about these twisted1:07 prophecies is that the words themselves1:08 are not technically wrong. Get it? You1:11 really can get into college without1:12 studying now. But the problem is, a1:16 situation arose where many colleges1:18 don't offer much of an advantage even1:20 if you go. In the end, looking back,1:23 only those who believed it suffered the1:25 consequences. Only the kids who1:26 believed it lost out. Well, what’s1:29 done is done. So, here is the prophecy1:32 for the current generation. Whatever it1:34 is, there’s no need to study. It’s1:37 being pushed as if AI will do1:38 everything for you. This is the deeply1:42 distorted prophecy of our time. And1:44 this prophecy will likely come true,1:46 too. That’s why I find it even1:48 scarier. For starters, if AI really1:51 does everything, then there’s no need1:53 to code. Because you won't be able to1:56 get a job as a coder. If you can’t do1:58 it, you simply become unnecessary. So,2:00 people who can’t do anything are no2:02 longer needed. If you just slightly2:04 twist those words, it becomes "people2:06 are not needed." It is true that LLMs2:08 have been improving significantly2:09 lately. But does that mean they do2:11 everything well? No, that’s not the2:13 case. When a new technology emerges,2:15 there are always limits it can never2:17 overcome. For example. For example,2:20 let's say you're building a car. If2:24 you’re building a car, you cannot2:26 defy air resistance. No matter how well2:29 you build a car, you can’t make air2:31 resistance disappear. Is a car that2:34 runs while ignoring gravity possible?2:36 That’s impossible. Yes. So, every2:39 technology has its own fundamental2:41 limitations. You can make improvements2:44 by overcoming those fundamental2:45 limitations in various ways. For2:48 example, if you design the exterior of2:50 a car to be sleek and aerodynamic, you2:53 can significantly reduce air resistance2:55. But does that mean air resistance has2:58 disappeared? Not at all. In fact, for3:05 internal combustion engine vehicles,3:07 there's a set limit where things like3:08 fuel economy, efficiency, and heat3:10 generation theoretically cannot be3:12 overcome. In that situation, innovation3:18 happens as something like electric3:20 vehicles emerges into a new territory.3:23 The same applies to AI. It’s3:27 impossible for something with3:29 superhuman intelligence to appear by3:30 skipping over those set physical limits3:32. The points you are seeing now are3:37 from the official lectures that you can3:39 watch for free on the Entropy website.3:43 Even outside of the official Entropy3:45 lectures, for those who know a bit3:47 about LLMs or computer science, or3:49 understand probability models, these3:51 are actually obvious concepts. They are3:53 self-evident. I mean, yes,3:54 understanding the principles is exactly3:57 like this. Now, initially, because3:59 people didn't know much, they thought4:01 that if they just threw in enough4:03 computing power and data since it was4:05 working so well, performance would4:06 surely increase infinitely, but it's4:08 been confirmed that we cannot deny4:10 these essential limitations. It4:13 doesn’t mean it won't get better. It4:15 does get better. But to transcend those4:19 fundamental limits, we need4:21 next-generation technology; simply4:22 refining current technology won't lead4:24 to some massive leap forward. While4:29 people are being scared that AI might4:31 lead to the extinction of humanity, if4:33 you actually go to the companies '4:35websites, they start by teaching that4:36 LLM-based AI services or AI assistants4:38 have limitations due to these inherent4:40 characteristics. They start by teaching4:43 this. The first point here is that it4:46 only predicts the next token4:47 probabilistically. The limitation of4:50 current language models is that you4:52 have no choice but to train them by4:53 saying that for this prompt, you could4:55 answer this way or that way. Perhaps in4:59 the near future, some geniuses or5:01 scientists might work hard to solve5:03 these limitations in a different way.5:07 And even with current technology, if we5:09 diligently refine datasets, run5:11 intensive reinforcement learning with5:13 user data, and fine-tune it carefully,5:15 it could be mitigated to a significant5:17 extent. But there is a general5:23 consensus among scholars that it won't5:25 suddenly become a superintelligence, as5:27 if some genius is doing everything for5:28 us. Now, in terms of knowledge, there5:37 was a time when we believed AI could5:39 hold all the world's knowledge, far5:40 more than any one human ever could.5:46 However, if you look at it now and use5:48 tools like Claude, you've probably5:50 noticed it fluctuates, where one thing5:52 gets better while another gets worse.5:56 So, while it can fetch and analyze5:58 external data, it definitely isn't5:59 memorizing all the knowledge in the6:01 world. Now, let's move on to working6:04 memory. Many people wonder if it can6:06 remember and process much more6:08 information than the human brain, right6:10? And think, "Wait, isn't it actually6:12 getting much better?" People thought6:15 that, but now we know that the more you6:16 feed it, the higher the probability of6:18 it getting confused. Yes. From that6:23 perspective, because of the nature of6:25 generative AI, it's difficult to6:26 control it exactly the way you want. So6:29, if you don't like the result on the6:31 first try, can you just do it a few6:33 more times? It gets quite a bit better.6:34 But if you look closely, you'll6:36 eventually find a critical flaw6:37 somewhere. Then there are the habits6:39 left by training, or biases caused by6:41 the datasets. In some cases, those can6:44 act in a very critical way. AI cannot6:46 take responsibility. You cannot assign6:49 tasks that require accountability to AI6:51, and I believe laws will be created to6:53 ensure a human remains in the loop.6:56 That’s what regulation is. Society6:59 and organizations are structured in a7:01 way that someone always has to take7:03 responsibility. It's not that AI can't7:06 do anything. There are definitely parts7:08 where it is useful. But that doesn't7:10 mean it will improve infinitely. It has7:13 its limits, and it will likely keep7:15 improving as it approaches those limits7:17. Recently, OpenAI claimed they solved7:21 the Navier-Stokes equations, but it7:23 turned out they didn't actually solve7:24 them. They just threw an incredible7:29 amount of computing power at it to find7:31 instances where it didn't work well.7:33 They found cases where problems7:35 occurred. There are many instances like7:38 that where some misunderstanding7:40 occurred. There are several examples,7:43 and the one you see now is from7:45 February 2025. So, an AI company7:49 claimed their AI system could train a7:51 hundred times faster, but it turned out7:53 they were just gaming the metrics7:54 instead of actually achieving the goal.7:58 So, eventually, they admitted it. They8:00 acknowledged that they had made a false8:02 announcement. Problems like this keep8:04 occurring, which means, in the end, we8:06 still need humans. There are8:08 limitations. There are simply clear8:10 limitations. The important thing is8:11 that, ultimately, only those who are8:13 deeply trained can recognize this. What8:16 is left for that person? First,8:18 judgment. What should we build?8:20 Deciding what to build. This is8:22 something humans must eventually do.8:24 They have to do it. Next, importantly,8:27 AI is good at execution now. It8:30 executes well, but it cannot judge for8:32 itself whether something is perfect,8:34 and it is the human who decides what8:36 qualifies as perfect, and it must be a8:38 human; finally, there is accountability8:40. Accountability. This is a survey of8:45 132 Anthropic engineers from December8:47 2025. They made it public themselves.8:51 In it, they consistently responded that8:54 people do not use Claude for strategic8:56 judgment or design decisions that8:57 require organizational context,8:59 organizational preferences, or taste.9:03 There are areas where it shouldn't be9:05 used. We tried various things, but9:06 realized, "Hey, you shouldn't use AI9:08 there." Earning money has many meanings9:10; for instance, wanting to work at a9:12 company. From the perspective of9:14 wanting to be hired, or wanting to9:15 switch jobs. Wanting to increase my9:17 salary when changing jobs. I want to9:19 approach this from those perspectives.9:21 Where does money tend to flow? Money,9:24 you see. Price and value are determined9:27 by substitutability. If you are easily9:30 replaceable, your market value is low.9:32 If you are very difficult to replace.9:34 If it's hard for others to do what this9:36 person can do. Then, your market value9:38 goes up significantly. Scarcity9:40 determines value. Yes. Anything anyone9:43 can create is bound to become cheap.9:44 Attempting to do simple coding manually9:46, just to be fast. That is something9:48 you shouldn't do. Yes, that's right.9:50 Yes. So, what I should do and what I9:52 shouldn't do—that is what I want to9:54 talk about today. The reason I'm being9:58 so passionate right now is that until9:59 recently, there were so many claims10:01 that there was no place left for humans10:03, which caused many of you to feel10:05 frustrated. But at this point, there is10:08 definitely a place for humans. There10:10 definitely is. It’s like this now10:12 because of current AI policies, but if10:13 we think about the next ten years,10:14 isn't it possible that might change?10:15 I'm going to take it easy for 10 years.10:16 Yes. Life, year by year, is incredibly10:19 important. Growth needs to happen fast,10:22 year by year, or even every six months.10:24 Especially when you're young, and it's10:26 an era where AI tools make that even10:28 more possible. You know the saying, "10:30Plant an apple tree even if the world10:31 ends tomorrow." I think there's real10:33 meaning in that. I didn't understand it10:35 when I was young, but in a way, doesn't10:36 it mean to ignore those talking10:38 nonsense about the world ending and10:39 just live your life? You can deal with10:41 10 years from now when the time comes.10:43 You just need to be the kind of person10:44 who can handle it when the time comes.10:46 Do you understand? When change happens,10:48 you could say, "That's got nothing to10:50 do with me, I don't care.""I'll just10:52 keep doing what I've been doing to make10:53 a living." Or, you can just live your10:56 life putting in your best effort and10:57 moving in the right direction. It's11:01 definitely important right now. But the11:03 point is, the timing is good right now.11:06 Generally, money flows toward the11:07 bottleneck. So, AI can certainly write11:10 code for you. It can write it, but you11:14 can't tell if it's correct or not. They11:17 said AI would even do code reviews.11:19 That's why they said humans wouldn't be11:21 needed anymore. But once people opened11:22 the hood and experienced trial and11:24 error, they realized leaving everything11:26 to AI would cause big trouble. Anyway,11:29 as it stands now, the bottleneck has11:31 finally become reviewing the code11:33 written by AI. And the people who are11:36 good at this will see their value rise.11:39 Then, is being good at reviewing very11:41 different from being good at writing11:42 code? There are overlapping parts and11:45 some that don't overlap. But you can't11:49 review code if you don't know how to11:50 write it at all. So, when code is11:53 abundant, the eye that knows how to11:55 read it becomes more valuable. From our11:58 perspective, there is clearly a vacancy12:00 in what we call LLM AI. If you let go12:04 of the attachment to convenience and12:06 seize this vacancy, things become much12:08 easier. This is where the problem12:12 arises. The part about juniors—the12:15 entry-level employees. When we say "12:17entry-level," you might call them that,12:19 but they are actually people studying12:21 in school looking for their first jobs.12:24 This is the problematic part. You12:26 entrust coding to AI. Then, your own12:28 coding skills decline. And if your12:30 coding skills decline, your ability to12:32 supervise also diminishes. AI won't get12:35 any better just because I push it12:37 harder. Therefore, if you use AI solely12:41 as a substitute for coding, you are12:43 highly likely to fall into a vicious12:45 cycle where your own ability to use AI12:47 effectively diminishes. The current12:50 system is very concerned about this.12:52 The talk coming out of the U.S. these12:54 days, what they are worried about is12:56 this. A junior developer becoming12:58 senior-level is simply a matter of12:59 doing the work. It's just that people13:03 aren't doing it. Because AI does13:06 everything for them, they've stopped13:08 doing it, and there's a lot of talk13:09 about this being a serious problem. So13:13 you need to study the basics, write13:15 code yourself, read documentation, and13:17 struggle through debugging to use AI13:19 well, but people are falling for the13:21 hype that AI will do everything,13:22 skipping these steps, and wanting to13:24 jump to the next stage, so the chain is13:26 broken. It's broken. The ladder is13:30 still there. The ladder is still there.13:32 The ladder—the ladder to improving13:34 your skills. The ladder to using AI13:35 well is exactly where it has always13:37 been. And actually, the rewards for13:39 climbing that ladder are even greater13:41 now. It’s true that hiring for13:42 entry-level positions has decreased13:44 significantly. But strangely, the13:45 starting salaries for the new hires13:47 that are being brought on have actually13:49 gone up. This is also strange. Yes.13:51 This is also strange. But that’s in13:53 the AI field, on the side that builds13:55 the AI systems that use AI. Yes. That13:57 area has seen increases. So, if you put13:59 off studying the basics because you14:01 heard that AI does everything, that is14:02 a huge mistake. Yes. The direction of14:06 the light, the highlights, and the14:07 shadows don't have any consistency here14:09. I intentionally created an example of14:13 what it can't do well, so as soon as14:15 you see it, you'll probably realize14:17 right away that the shadows look14:18 strange. Another classic example you've14:21 probably seen is when the number of14:23 fingers looks weird. These simple14:26 things—people can notice these simple14:28 things right away. Even if you aren't a14:31 graphics major, you can easily spot14:33 something like this. But as you get14:35 into increasingly specialized fields,14:37 the errors become subtle. And things14:40 that are valuable aren't common; they14:42 are rare. So, to find those subtle14:45 errors, you ultimately have to study.14:51 That doesn't mean you have to draw14:52 hundreds of pictures just to study this14:54, right? If you just understand the15:01 principles accurately, the days of15:03 having to suffer through endless typing15:05 and drawing are largely gone, and I15:06 hope you can adopt a good perspective15:08 to take advantage of this wonderful era15:10. Since many people caught up in15:17 distorted prophecies are just sitting15:19 idle, if you enter the field now, you15:20 can easily get ahead. So, if you15:24 mindlessly click to approve things when15:26 it’s unclear whether they’re right15:28 or wrong, something truly terrifying15:30 happens. Yes. But here's the thing.15:33 Paradoxically, those who have actually15:35 done the work from the ground up know15:37 exactly that this could become a15:38 problem. If you've worked in the field,15:40 you'll know this. In the past, the15:42 approach was specifically about piling15:43 up knowledge in your head. But from now15:46 on, simply accumulating knowledge like15:48 that will decrease in value. You can15:51 just offload that knowledge to AI. Once15:55 you clear the neural pathways for15:56 critical thinking, you can grow very15:58 easily by leveraging AI. That’s why I16:02 keep saying it’s such a great time.16:05 That’s why I keep forcing you to16:06 think in the first part of all my16:08 lectures. Try thinking for yourself. I16:11 struggled a lot trying to adjust the16:12 difficulty levels while doing all of16:14 that. But regardless, for now, there is16:16 nothing more valuable than equipping16:18 yourself with the ability to think and16:19 clearing those mental pathways. We’ve16:22 now reached a consensus on the study16:23 method. At the introductory level, do16:25 it the old-fashioned way. You need to16:29 move toward understanding basic16:31 principles, being able to direct LLMs,16:33 and developing a general mindset that16:34 isn't tied to one language but allows16:36 you to work across many. Basically, at16:41 the starting stage, it’s back to the16:43 old way after all. There’s a recent16:47 trend that says it's better to be a16:49 generalist who knows a little bit of16:50 everything. I absolutely do not think16:53 so. Because someone who has dug deep16:57 into a field and someone who hasn’t16:59 have completely different attitudes17:00 toward life. It has become an era where17:04 the experience of diving deep into17:06 something is incredibly important. But17:09 now, if you use AI, it’s very fast17:10 and easy to transition deeply into17:12 other fields as well. That’s why17:15 digging deep into one thing is much17:17 more valuable. That’s the era we are17:19 heading toward. Just hovering shallowly17:21 between this and that at a mediocre17:22 depth. Yes. That’s just asking to17:26 compete head-on with AI. Many of you17:30 have probably experienced that as AI17:31 goes deeper, it starts to hallucinate.17:34 So, if you move to areas where AI17:36 struggles, you’ll definitely win.17:39 There’s no need to mess around here17:40 and get things tangled up. So, I highly17:43 recommend digging deep into just one17:45 thing. But honestly, it was like that17:47 in the past, too. It was like that17:49 before, but it was just too hard to do.17:54 But now, if you need to memorize things17:56 or find reading English papers17:57 difficult, you can get help from17:59 language models since they explain18:01 things so well, making this a truly18:02 great time to study. So, when I18:07 recommend graphics, I do it for two18:09 main reasons. First, it's the best18:12 field for studying something in great18:14 depth. That's because when you dive18:17 deep, the results are visible. It’s18:20 fun, but because it’s visual, you can18:22 study by utilizing our natural18:24 strengths. Since you can see it, if18:26 something looks off, you immediately18:27 realize, "That shouldn't be moving like18:29 that," and it really hits home. Yes.18:32 That’s one reason, and secondly,18:34 it’s a field that clearly targets the18:36 weaknesses of AI in the era of Physical18:38 AI. Korea, in particular, is pushing18:40 very hard into Physical AI right now.18:44 They’re pushing it almost as if the18:45 nation’s life depends on it. If you18:49 are skilled in a deep field, especially18:51 in dealing with 3D space, you never18:53 have to worry. For one, the most18:56 certain part, in my opinion, is this.18:59 For data structures, algorithms, and19:01 basic programming, it's good to start19:03 by practicing with problems you can19:05 solve entirely on your own, using 100%19:07of your own brainpower. Yes. I19:10 recommend that. That much is certain.19:12 Getting AI help from the very beginning19:14—That isn't good. It’s like giving19:17 a baby a walker from the start just19:18 because autonomous driving exists.19:21 That’s just not right. That doesn't19:24 mean you have to solve extremely19:25 difficult problems without AI, but you19:27 need to adjust that boundary. In my19:31 lectures, the practice problems are19:33 designed with plenty of hints so you19:34 can solve them without even needing to19:36 search the internet. But if I just hand19:39 out a difficult, twisted problem that19:41 someone is supposed to get wrong and19:42 say, "Solve this using only your brain,19:44" that wouldn't work. Distinguishing19:46 between those cases is what an educator19:48 does. That is the job I do. Hmm, the19:50 most important thing is this.19:52 Everything I’ve been talking about19:54 for hours today is an effort to19:55 convince you of this one point. Your19:58 foundation determines the speed at20:00 which you learn new things for the rest20:01 of your life. It’s a bitter truth20:05 that the value of knowledge itself is20:06 declining. And although it’s not easy20:11 for me to keep studying new things,20:13 I’m making a living because I have20:14 the ability to do so. And in the future20:17, it will only accelerate. As I kept20:20 saying earlier, there are areas where20:22 AI learns well and areas where AI is20:24 slow or cannot learn at all. One thing20:28 is that the areas it can't learn are20:30 ultimately those that require deep20:31 thinking. And moving forward, new20:34 technologies will continue to emerge.20:37 The world creates change. But is change20:39 bad? Absolutely not. Actually, if20:43 you’re in a position where you have20:45 to make a living solely on your own20:46 abilities, constant change is a great20:48 opportunity. In a world that keeps20:51 changing, having a solid foundation20:53 allows you to learn quickly and live20:55 stably. You never know what might20:56 happen. Another technology like AI20:58 could emerge. Some next-generation21:00 technology might come along. If you21:01 learn it quickly and get ahead, you can21:03 gain significant advantages. In that21:06 sense, I believe you should build a21:08 proper foundation whenever you have the21:10 chance. The idea that you don't need a21:14 foundation because AI does everything21:16 is a twisted prophecy. That is why21:17 studying LLMs is extremely important.21:19 Studying AI properly is essential. I21:21 mean, knowing how to use Claude is one21:23 thing. Beyond that, being able to code21:25 with ChatGPT is another. But for those21:27 who understand the core and have21:29 studied the basics to build AI systems,21:31 it’s a real opportunity. No matter21:32 how I look at it, I strongly feel that21:34 this is an opportunity. You should21:36 quickly delegate tasks that AI can21:38 handle. And you must rapidly expand21:39 your focus on the work that only humans21:41 can do. So, while it is powerful as an21:43 assistant, it is weak as a leader. AI21:45 clearly has its limitations. There is21:48 work that only humans can do, and21:50 becoming stronger in that direction is21:52 the opportunity. I think it would be21:54 good to view it that way. There’s a21:56 post written by a security engineer at21:58 Anthropic. The solution Claude provided22:00 was very smart, but in a dangerous way.22:03 It was a solution a talented junior22:04 engineer might offer, but the issue was22:06 that only someone with judgment and22:08 experience could spot the problem. So,22:11 it didn't provide a senior-level22:13 solution. A senior-level professional22:16 is someone who can manage risks. It's22:18 not about just coding well. Or coding a22:20 lot. That’s not it. It means you can22:21 feel at ease when you entrust the task22:23 to that person. Can you feel at ease22:25 leaving things to AI? You can’t.22:27 Right. How do you know how it might22:29 ruin your life? If it ruins it and says22:31, "Oh, I made a mistake.""I messed up."22:33 That would be a really difficult22:35 situation, wouldn't it? There was a22:37 case where someone commented on YouTube22:39 that they received an offer from22:40 NVIDIA's recruiting team earlier this22:42 year. They got the offer, but they were22:44 rejected during the document screening.22:47 The next thing they said was, "I'll22:49 study CUDA hard from the basics." What22:51 does that mean? It means they hadn't22:53 studied CUDA. You're applying to NVIDIA22:55, yet you don't know CUDA. It was quite22:58 unfortunate. Many people seem to think23:01 that being a newcomer means only doing23:02 easy work. But in the future, there23:06 will only be a place for newcomers who23:08 take on new challenges. That is why you23:12 must study the basic principles of LLMs23:14 to be able to build AI systems. This is23:17 the foundation. That's what I wanted to23:19 tell you. Breaking a twisted prophecy23:22 is simple. In the current situation,23:25 anyone who keeps learning at a steady23:27 pace will never fail. Increasing your23:31 learning speed is easier now than ever23:33 before. Since we have to go this way23:37 anyway, I'd say the best time to start23:39 is right now. So today, I wanted to23:44 share some positive words that you can23:46 also survive in the AI era. You can23:49 definitely survive. I focused on the23:54 fundamental limitations of LLMs or AI23:56 services, and the parts that humans23:58 will still need to handle in the future24:00. That is all.
4,071 words · 627 lines
![[선데이 라이브] 당신도 AI 시대에 살아남을 수 있다!](https://i.ytimg.com/vi/D95UC9U5ShA/hqdefault.jpg)
