YT YouTube2Text

YouTube2Text → Transcripts → [편집본] 미국 빅테크 일자리가 한국으로 왔다!

[편집본] 미국 빅테크 일자리가 한국으로 왔다! — Transcript

by 홍정모 · 5,269 words · 816 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Since the talent couldn't go there, the
0:02jobs came here. Yes. Big Tech jobs have
0:04been flooding into Korea. When I
0:07analyzed the job postings, it turns out
0:09these are roles they couldn't even fill
0:11in the U.S., so they're coming over.
0:13They are all high-level positions.
0:15They're really hiring. Yes. I'm telling
0:17you, they are really hiring. The jobs
0:18are coming to us. New Korean recruits
0:20are outstanding. Since Koreans in the
0:22U.S. perform well, they obviously know
0:24that. A somewhat absurd situation has
0:25occurred where the jobs are coming to
0:27us. I'm a backend developer with 7
0:28years of experience. Changing fields
0:30isn't easy, but web backend roles
0:32building systems with AI—that’s
0:34where the highest pay is right now. But
0:38the burden of accumulating knowledge
0:41one by one or memorizing things has
0:43been greatly reduced. You must never
0:46compromise on sharpening your brain's
0:48thinking circuits. You just need the
0:50circuits in your head. Actually, it’s
0:53because Korea has a firm grip on the
0:55manufacturing industry. But we have to
0:57protect that. That is the mission of
0:59your generation. You must protect it no
1:01matter what. Now, today's topic.
1:03Today's topic is that U.S. Big Tech
1:04jobs are coming to Korea. I will talk
1:06about how you should prepare for that.
1:09I investigated about 500 job postings.
1:13As I looked into it, I found there are
1:14patterns. There are commonalities as
1:16well. I'll start by talking about the
1:19areas where I see opportunities. Big
1:23Tech companies are actually hiring in
1:25Korea. Honestly, at first, I thought
1:30they were just trying to push sales or
1:33marketing roles here. But when I looked
1:38into it, they were hiring engineers and
1:40researchers. It wasn't that. They are
1:43really high-quality jobs, senior
1:45engineering roles based in Korea—not
1:48even remote work—so the jobs have
1:50come to us. So, Korea is overflowing
1:54with talent. Overflowing with talent.
1:57We have tons of talent, which is why
1:58we've been exporting them. To the U.S.
2:00Yes. Since the talent couldn't go there
2:03, the jobs came here. Yes. Big Tech
2:05jobs have been flooding into Korea. So
2:08I’ve been tracking and investigating
2:10this for months. It’s not just a
2:12temporary thing, either. And there is a
2:14reason for this. So I don't think it's
2:16a short-term trend. I expect it to last
2:19for quite a long time. Geopolitical
2:21factors don't change that quickly,
2:23after all. It’s an opportunity. An
2:25opportunity. Big Tech is coming here to
2:28treat you well. The reason they are
2:31doing this is, in fact, because Korea
2:33has a firm grip on the manufacturing
2:34industry. The infrastructure is
2:36stronger than you think—it's very
2:37powerful. But we have to keep it that
2:39way. That is the mission of your
2:41generation. You must uphold that. You
2:43have to do that. The reason I emphasize
2:45C ++, graphics, physics, and the
2:47fundamentals of LLMs and RAM is because
2:50that is your mission. You must uphold
2:52that. You must uphold it no matter what
2:54. Without this chaos, we would just
2:57have to stand in line. This means
3:00people with a score of 100 get into
3:02good places first, followed by those
3:04with 90, forcing us to live a life of
3:07standing in a single line. Unless you
3:11possess exceptional abilities to
3:13guarantee a score of 100, these changes
3:16in the world and technological
3:18advancements are actually a form of
3:20positive chaos. Therefore, when we go
3:23through that process and things start
3:25to settle, it is actually a fantastic
3:27opportunity. These are just big tech
3:29companies. I’ve gathered big tech
3:31companies here. Looking here, Riot
3:33Games is also hiring in Korea. Wow,
3:36there are so many. After analyzing
3:37these job postings, it’s clear they
3:39aren't outsourcing just to save money.
3:41They are genuinely looking because they
3:42need people. These are positions they
3:44couldn't even fill in the U.S., so they
3:46are looking abroad. The jobs I have
3:48gathered here are all high-level
3:50positions. NVIDIA autonomous driving
3:52system software. C ++, Python, Linux,
3:54and CPU/GPU architecture—this means
3:57you have to use both the CPU and GPU
3:59together. So, things like heterogeneous
4:01parallel processing are fundamental.
4:03Surprisingly, there aren't actually
4:05that many people with a solid
4:07foundation. Amazingly, that is how the
4:09world works. In some sense, once you
4:11cross that barrier to entry, a new
4:13world opens up unexpectedly. That’s
4:16the feeling I get. Linux kernel
4:18engineer. It requires strong C
4:20programming skills. Yes. You can be a
4:22Tesla employee with C language skills,
4:24and you can even work in Seoul. Living
4:26in the U.S. might suit some, but
4:29honestly, Seoul is better for leisure.
4:33And with your family and relatives here
4:36, you can work in Korea with pretty
4:39decent pay. And then, you get "Tesla"
4:42stamped on your resume. Over here is a
4:45Field Deployment Engineer. There was
4:48some talk that there would be a lot of
4:49jobs in this area. Looking at the
4:51languages, they mention Python and
4:53JavaScript, so think about it. In the
4:57past, there were very few cases that
4:59required both JavaScript and Python.
5:02But now, they require both. It’s
5:03written in small print, but it includes
5:05mentions of LLMs and generative models.
5:08To handle neural networks, deep
5:10learning, or things like LLMs, you also
5:12need to know Python. Then, Teradyne is
5:16a semiconductor test equipment company.
5:20But from my perspective, looking at
5:21these job postings, I really hope that
5:23semiconductor engineers, especially
5:25those working in software, get the
5:26recognition they deserve. That's what I
5:30think. It's Qualcomm, and Qualcomm is
5:32famous, right? Qualcomm is famous, and
5:34I’ve seen many people move to even
5:36better companies after working there.
5:38Qualcomm new grad. Qualcomm new grad.
5:41What could this mean? Python and C/C ++
5:44. While other companies are poaching
5:46experienced workers, Qualcomm has been
5:48looking for experienced hires too.
5:51Nevertheless, new grad positions have
5:52appeared. What does that tell you?
5:54There’s a real shortage. Jobs are
5:56coming back. But the same jobs won't
5:58return exactly as they were. I’ve
6:00been saying that they come back in a
6:01different form. So, please, study
6:03according to global standards. I didn't
6:06expect Big Tech to start hiring in
6:08Korea like this. New grad positions are
6:10starting to appear. New grad. Once jobs
6:13start to spread, they spread quickly.
6:16As the hype around AI begins to subside
6:19, I believe companies will go back to
6:22hiring again. The reason I’m
6:25analyzing these job postings is that
6:27they are the clearest evidence for you
6:30to predict where the opportunities will
6:32emerge when they come flooding back.
6:35Entry-level embedded software engineer,
6:37new grad, commercial—it seems anyone
6:40can apply. They really are hiring. Yes,
6:42I’m telling you, they really are
6:43hiring. Jobs are coming to you. New
6:45Korean graduates are exceptional. Since
6:48Koreans in the US are doing well, they
6:49would naturally know that. An
6:50incredible situation has occurred where
6:52jobs are coming to us. Lam Research is
6:54one of the world's top 5 semiconductor
6:56equipment companies, a US firm that
6:57supplies etching and deposition
6:59equipment to Samsung Electronics and SK
7:01Hynix factories. And they are now
7:03requiring more advanced skills, like
7:05multi-threading. You are all studying,
7:07right? So, I compiled statistics on the
7:09languages. Python is overwhelmingly
7:11high. But the Python mentioned here
7:14refers to Python as a professional
7:16programming language. And interestingly
7:19, C and C ++ are even more common.
7:23It’s not that Java is bad, but rather
7:25that there is already an abundance of
7:26that talent, which is why they come
7:28abroad to find the talent they lack. US
7:31Big Tech is coming here specifically to
7:33hire talent that is in short supply. C,
7:37C ++, and professional-level Python are
7:40in extremely short supply. I think this
7:44clearly demonstrates that. Then there's
7:46CUDA, and I think that’s how the
7:48order goes. The Ministry of Employment
7:51and Labor had a survey on why companies
7:53couldn't find the people they needed.
7:56Uh, 25.8%said they couldn't hire
7:58because there were no applicants with
8:00the required experience, and
8:02specifically in IT R&D and engineering
8:04technical jobs, the vacancy rate for
8:06roles they wanted to fill but couldn't
8:08was 16.2%, which is a whopping 2.5
8:10times the average of other occupations.
8:13So, the hiring side is desperately
8:15looking. But if the applicants 'skills
8:18don't match, they just can't be hired.
8:22Actually, I’m generally known as
8:24someone who teaches the basics or
8:26difficult subjects, and you might think
8:29of me that way, but what I pursue is
8:31actually avoiding competition. If you
8:35find a good niche market point and
8:37strike hard, a vacancy temporarily
8:39opens up like a vacuum. That’s how
8:42you should push forward. Areas with
8:44barriers to entry are advantageous. Uh,
8:47barriers to entry—it's especially
8:48easy to study nowadays. It’s very
8:50different from the past. So, when I say
8:52now is an opportunity, it was a time of
8:54chaos before. It's heading toward a
8:56stable phase, so the future is
8:57predictable now. Because we are in such
9:00an era, it’s honestly a really great
9:02time for you all to live smart. So, you
9:06just need some proof that your
9:07foundation is solid, and then you need
9:09to be ahead of the curve. Things like
9:12CUDA. And accelerated computing. I see
9:14this as a very good opportunity right
9:16now. The Korean government's investment
9:19direction, and then US trends arrive in
9:22Korea with a slight delay. So, in the
9:25current situation, big tech companies
9:27are hiring in Korea, and they want to
9:29know if you understand the
9:31characteristics of LLMs well. Many
9:33people think that C ++ is difficult. Uh
9:36, C ++ proficiency—C ++ is just a
9:39symbol. I’ve said this many, many
9:42times. C ++ proficiency is not about
9:45memorizing syntax and knowing where to
9:47use which grammar. It’s not about
9:49understanding inheritance or things
9:51like that either. It’s about whether
9:53you know data structures and algorithms
9:56well. I mean, we use C ++ because
9:58it’s the language that can implement
9:59data structures and algorithms most
10:01efficiently. If you want to be good at
10:03C ++, you have to do data structures
10:05and algorithms. Next, NVIDIA physics
10:07simulation engineer. Uh, I guess they
10:09really are short on people. Well, there
10:11are a lot of interesting things. Right,
10:13AMD is also hiring in Korea. But AMD
10:15seems to be hiring for gaming roles.
10:17But something very interesting is that
10:19Korean is required. Wow, Korean—are
10:22you all good at Korean? Yes. If you are
10:25good at Korean, you can work at AMD in
10:27Korea. Rendering engineers' market
10:29value seems likely to go up. Next, in
10:32terms of the study order, I recommend
10:35you quickly learn DirectX 3D 11, then
10:38study Vulkan, and then move on to
10:40DirectX 3D 12. A Deep Learning
10:44Applications Engineer is a very
10:45high-level engineer. So, it is
10:48important to study LLMs themselves.
10:52Even if you are just using an LLM as a
10:54tool, understanding the nature of the
10:56LLM itself is extremely important.
10:58After all, an LLM is the culmination of
11:00all the machine learning technology
11:02that humanity currently possesses. Top
11:05talent is focused on it, and all the
11:08electricity and GPUs in the world are
11:10being used to develop these LLMs. From
11:12a study perspective, what is the best
11:14software technology humanity has? It is
11:16currently the LLM. Therefore, studying
11:18this LLM is a very meaningful endeavor
11:21for you all. Furthermore, it is
11:23directly linked to jobs. It seemed like
11:26AI would do everything, but when you
11:27actually use it, there are many cases
11:29where it makes things harder for me.
11:32It’s good at rambling on and on about
11:35things, but when you ask, "So what?",
11:39it can’t answer. But on the flip side
11:41, there are things that have become
11:42easier. So, if you definitely delegate
11:44the things AI is good at to the AI,
11:46productivity does indeed go up.
11:49Therefore, I think it is important to
11:50distinguish these things well. The
11:53easiest way to distinguish them is by
11:54studying the principles of LLMs; then
11:56you will see what it will be good at.
11:58And what it won't be good at. It just
11:59becomes clearly visible. So there is an
12:01aspect of needing to study LLMs, and
12:03then there is what we call AX. In Korea
12:06, those doing things like AX need to
12:08have a precise understanding of the
12:10nature of AI and the nature of these
12:12LLMs. I think there is an aspect of
12:13needing to understand how to handle
12:15them. Uh, quantitative tasks, simple
12:17repetitive tasks, and tasks that are
12:19just agonizing to do—those have
12:21decreased a lot. They have decreased a
12:24lot, but what you must do instead is
12:26never compromise on your fundamental
12:28intelligence. The era of piling up
12:31knowledge in your head is over. But you
12:34must never compromise on clearing the
12:35circuits in your brain for thinking. In
12:37the past, if you said you were good at
12:39C ++, or good at Java, you first had to
12:41memorize the syntax. But now, you don't
12:44even need to do much typing. Physically
12:47, at least, it has become much more
12:49comfortable. Once you actually start
12:51working, you only need to evaluate
12:53whether it was made well. Because the
12:57difficulty of creating a good program
12:59and the difficulty of evaluating if it
13:02was made well are completely different.
13:05You know those 1,000-piece or 10,000-
13:07piece puzzles? Putting those puzzles
13:10together is extremely difficult.
13:13However, regardless of whether a puzzle
13:14has ten thousand, a hundred thousand,
13:16or a million pieces, checking if it’s
13:18put together correctly happens in an
13:20instant. Now, does this story come from
13:22somewhere extraordinary? Not at all. It
13:24comes from algorithms. This is an
13:26example used when teaching algorithms.
13:28So, if you don't know this story or
13:30this example, you haven't actually
13:31studied algorithms. Therefore,
13:32there’s no need to increase your
13:34proficiency in areas that AI can handle
13:36for you. You just need to have the
13:37right mental framework. If you get a
13:40feel for the easy things and expand
13:41little by little, you can study very
13:43efficiently. It’s true that it’s
13:46become easier to study. It’s also
13:48true that it’s become easier because
13:49of AI. If you set your direction well,
13:52you can achieve high results very
13:55comfortably. It even drew out how the
13:59500 job postings I analyzed today
14:03connect to what I teach. My approach to
14:07teaching LLMs has an aspect of studying
14:10them as a case study for the latest
14:11machine learning and deep learning
14:13technologies. Three years after I
14:17shouted to study based on global
14:19standards, Big Tech companies have come
14:21to Korea and fulfilled my prediction.
14:24The fact that they are operating
14:26engineer and researcher positions in
14:28Korea is, from your perspective,
14:30absolutely a good thing. Since some
14:33people ask about the order of building
14:35an LLM from scratch, I'll organize it
14:37for you. In Part 1, Principles of
14:39Statistical Language Models, we start
14:41with what a basic language model is. In
14:44Part 2, we move on to examples of
14:45entering the language model field using
14:47neural networks. Part 3 finally moves
14:49on to Transformers, exploring how
14:51modern LLMs began. And I have to say,
14:53even by my own standards, I taught the
14:54principles really well. You couldn’t
14:56visualize this any better. The core
14:58point is that that’s what I’m good
15:00at. Teaching difficult things easily. I
15:01really do teach difficult things in a
15:03simple way. So, I pick out the most
15:05important aspects of Transformers and
15:07teach them to you precisely. After that
15:09, there’s theoretical content on
15:11in-context learning, and that is very
15:14important. So, please make sure to
15:16study that. Next, what I’m covering
15:18in Part 4 now is training
15:20conversational LLMs through fine-tuning
15:22. Doing fine-tuning. It’s excellent
15:24for studying the basics. Then Part 5 is
15:26building a RAG system, which is
15:28directly connected to what is referred
15:30to as AX in Korea. Yes. And for the
15:33final part, there is building agents.
15:38My lectures are structured with fun
15:39examples, but in the latter part, I
15:41cover high-efficiency computing like
15:43multi-threading and multi-processing. I
15:47incorporated things like rendering,
15:49thinking it would eventually become
15:51such a hot technology. I had faith.
15:54That technology would eventually head
15:55in this direction. Yes, I had that
15:57faith. So that’s what's included in
15:58there. What I recommend is to watch the
16:00Python crash course, check out the C ++
16:02core summary lecture, and if you feel,
16:04"Oh, I think I can do this much using
16:06AI," then you should go straight to
16:08data structures and algorithms. But
16:10here, you absolutely have to use your
16:12brain. Focusing on intellectual effort.
16:14If you study the basics of data
16:15structures and algorithms, especially
16:17the algorithm lectures taught at
16:18universities. As for my lectures, I
16:22took the Stanford curriculum and broke
16:24it down to explain it more easily. Then
16:27there is the algorithms part, and Part
16:292 includes some more practical elements
16:32. My "First Steps in Deep Learning"
16:34course is currently free. So, just try
16:36skimming through it once. If you look
16:38through it and think, "Deep learning is
16:40doable," then you can move on to
16:41building an LLM from scratch. The "
16:43Building an LLM from Scratch" course
16:44has a lot of videos. I've put a lot of
16:47effort into summarizing them into
16:49videos, so even if you just watch them
16:51for fun, they contain good talking
16:53points for interviews and content that
16:56can help you sound like an LLM expert
16:58anywhere. There are quizzes, too, and
17:01it’s quite fun. Then there is my CUDA
17:03course, where the focus is on
17:04heterogeneous computing, which means
17:06parallel computing across different
17:08types of processors. And that is
17:10exactly what the market requires right
17:12now. Suppose you’re working with
17:14robots. You need to boost performance
17:15in those robots. Maximizing hardware
17:18capability ultimately comes down to
17:20keeping both the CPU and GPU busy and
17:22configuring them to maintain an optimal
17:24state, which is what this part covers.
17:27Next, in terms of graphics, algorithms
17:29that handle 3D space are inevitably
17:32best developed in the graphics field.
17:34It’s here that you realize, "Ah, this
17:35is what programming with math is like."
17:37And "This is what it means to handle 3D
17:39space." For studying this area,
17:41graphics is, and will always be, the
17:43absolute best. For those who want to go
17:48deeper, you can look into Vulkan and
17:50Gaussian Splatting. Especially for
17:53those who want to study in professional
17:55depth, these are topics that are hard
17:56to find elsewhere. If you've tried SPH
17:58simulation, you’ve probably gotten a
18:00good grasp of a few things. Like what
18:02it means to perform simulations on a
18:03computer. And how to handle 3D space.
18:05Then, how it connects to graphics, and
18:08since we extract polygons to render SPH
18:11, right? That technology is actually
18:14called "reconstruction" when building
18:16things like digital twins. It is
18:19directly related to that 3D
18:21reconstruction technology. So, studying
18:23that is very meaningful. I actually saw
18:26a job posting from Naver Labs. The
18:29first thing they were hiring for was
18:31mesh-based 3D reconstruction. Postings
18:34for that keep appearing. So I realized,
18:36the openings are definitely there. But
18:38right now in Korea, while LLMs are
18:41being applied and released as services
18:44that any company can use, the real
18:48battleground is Physical AI. Physical
18:52AI needs to operate in a 3D space, and
18:54what is the most developed field for
18:56dealing with 3D space? It’s computer
18:59graphics. So, approaching it from that
19:01angle is really great. The technology
19:03used in game engines is speed-oriented.
19:05Surprisingly, though, we’re in a
19:07situation where we have no choice but
19:09to use the physics engine technology
19:11found in game engines. Why? Because of
19:13the speed. So there are some
19:15limitations we hit. But new
19:17technologies are emerging now, even for
19:18simulation precision. To do things like
19:21reinforcement learning, speed is what
19:23matters. Reinforcement learning really
19:25needs to be fast. It means you have to
19:27do it a lot. Surprisingly, you can't
19:28really use expensive simulation
19:30technologies. So, we end up borrowing
19:33game technology for now, but studying
19:35physics simulation for games first, and
19:37then moving on to technologies used in
19:39robotics and similar fields is very
19:41advantageous. You mentioned two things,
19:45and both of them are areas with plenty
19:49of job openings. Honestly, job
19:51prospects in other areas aren't very
19:53good. You have to choose between making
19:57robots or building systems for a
19:59company using LLMs; the two are very
20:02different. Considering the hardware
20:06side of things, I don't think it'll be
20:08easy for a data engineer to suddenly
20:10jump into Physical AI. However, for
20:13those who usually do graphics, I highly
20:15recommend heading into the Physical AI
20:17field. Graphics experts are usually
20:20interested in simulations and things
20:22like that, so I think the AI engineer
20:24path would be better than Physical AI
20:26for them. For now, I have a lecture
20:29series uploaded for free on my YouTube
20:31channel. "First Steps in Deep Learning.
20:33" Just take a look at that casually.
20:35Yes. You should watch that first and
20:37then make your decision. If you are
20:39proficient in Python, I think it would
20:41be better to look into the LLM side
20:43from the perspective of a data engineer
20:45. I'm a job seeker preparing for the
20:47gaming industry. Should I start
20:48studying AI now as well? Well, in the
20:50gaming industry, art and tech are
20:52separated, aren't they? If you are
20:54headed toward the tech side—meaning
20:56if you aren't purely in art—I
20:57strongly recommend studying AI. Game
21:00companies these days are even
21:01struggling desperately to transform
21:02into AI companies. What I recommend is
21:04this. What I recommend is that you
21:07build a robot simulator. Assuming
21:10you've already finished the basics. But
21:13looking at you now, it seems like you
21:14haven't studied data structures and
21:16algorithms. You must do it, no matter
21:17what. Yes. If you're twenty, you must
21:19do it, no matter what. If you're in
21:21your twenties, I always tell you to
21:22just do it. You will never lose out.
21:24Yes. If you're doing physics and things
21:26like that, you have to be good at
21:28coding. Even if you go to graduate
21:29school, being good at coding gives you
21:31an automatic advantage. It will never,
21:32ever mess up your life. My graphics
21:34course is not a course for memorizing
21:36APIs. It's a course that teaches the
21:38principles of graphics, but you have to
21:40use an API after all. So I teach it
21:42using DirectX 11. It makes it easier to
21:46connect to DirectX 12, allows you to
21:48study what an API actually is, and
21:50fundamentally, it's a graphics course.
21:54That math program is really good; even
21:57I think I made it well. Yes. I'm a
22:01backend developer with 7 years of
22:02experience. They say changing
22:04industries isn't easy, but web backend
22:06developers using AI to build systems
22:08are the most highly valued right now.
22:11Practitioners are talking among
22:13themselves, and they're having a hard
22:15time because there's no one good at it.
22:17Among themselves. That talk will start
22:19to break out now. It doesn't work as
22:20well as you think. But they didn't
22:21study the underlying principles. So
22:23they start by building the system first
22:24. They buy GPUs, build a system, set up
22:27the backend, and think it'll work, but
22:29the "how" is missing. They thought it
22:32would work, but the "how" doesn't work.
22:35Those who fill that gap by studying
22:37will be treated very well. Why, if
22:39you're a backend developer with 7 years
22:40of experience, you've done well in your
22:41company, right? And you're likely good
22:42at backend. But if you know a bit about
22:44AI. That's an opportunity. This is a
22:46really great opportunity. There are far
22:48more developers on the East Coast now.
22:51They say there are more developers on
22:53the East Coast than in Silicon Valley.
22:56The reason is that places like Wall
22:57Street and banks need to use AI for
22:59their web backends. They need to adopt
23:02AI, but there are no people. There
23:04weren't many engineers on the East
23:05Coast before. That's why I always say
23:07this. Korea just waits and watches, and
23:09then once the US starts doing it in
23:11earnest, they chase after it like crazy
23:13. So if you study, there will
23:14definitely be opportunities. The reason
23:17is that companies need to adopt AI, but
23:20there are no people, so if you think
23:22that LLM-based services, which we call
23:25AI, will be used in every company—
23:27finance, law, accounting, etc.—then
23:30it's the right choice to study. Next,
23:33for those who worked in back-end and
23:35handled large services, if you have
23:37that experience and also know a bit of
23:39AI. And have built a service that
23:41incorporates AI. I think you’d be
23:43compensated quite well, as businesses
23:45have now decided that AI is essential.
23:47There is no need to obsess over
23:49knowledge. You need to be better at
23:51handling knowledge that is complexly
23:53interconnected. And algorithms are what
23:56do that. I thought people would
23:58naturally study data structures and
24:00algorithms. But surprisingly, they
24:03don't. I was actually quite shocked by
24:05that. Grammar exists to implement
24:08something. It’s a convenience feature
24:10meant to make things easier. But if you
24:12don't know the intent behind those
24:13convenience features, it just looks
24:15complex. That is why I tell you to
24:16study data structures and algorithms.
24:17Paradoxically, since you don't need to
24:19memorize that grammar anymore, the
24:21generation that maintained their
24:23competitiveness by memorizing grammar
24:25is disappearing. There is no need for
24:26that anymore. From the perspective of a
24:28student, things have gotten much better
24:29. Conversely, if that was your core
24:31identity, you will be pushed out. When
24:33I created my Python course, I was very
24:34careful to ensure that the difficulty
24:36of the introductory practice problems
24:37was something you truly had to solve
24:39with your own brain. Exercises that
24:42purely cultivate your own mental
24:43faculties are truly important. Whether
24:46or not you have that will determine the
24:49course of your life. It's certain. As
24:52AI becomes more powerful, your
24:54fundamental cognitive abilities will
24:57need to work even harder. It will be
25:01even more intense than in the past.
25:03Things that you could get by with
25:04through memorization are, in fact,
25:06being swept away. So, things like being
25:08a "human encyclopedia" are no longer
25:09needed. If you are incredibly good at
25:12data structures, algorithms, and using
25:14AI for coding, you will be successful.
25:16Yes. There are barriers to entry, but
25:19once you cross them, you will do well.
25:23Yes. And those barriers to entry have
25:26changed. It has shifted from
25:29memorization to understanding and
25:31building neural circuits in your brain,
25:33so once those circuits open up, you
25:35instantly cross that barrier. Then,
25:38it’s about how well you can control
25:41AI. Also, can you avoid being swayed
25:43when the AI talks nonsense? Can you
25:45judge what the AI has created quickly
25:48and accurately? Everything is shifting
25:50toward this. For instance, if you look
25:53at snowboarding, experts glide down
25:56smoothly, effortlessly, and handle it
25:58with such ease. But beginners struggle,
26:02asking why they keep falling, and it's
26:04extremely painful if you don't grasp
26:06the technique. So, I truly hope you all
26:08study in a way that leads you to
26:09success and ease. If you are in a STEM
26:11field, I absolutely recommend studying
26:14Python programming. Without exception,
26:17regardless of the field, Python is a
26:20must for all STEM areas; people in
26:23group chats get confused whenever a new
26:26model is released, wondering what it is
26:28. It sometimes feels like chasing after
26:31girl groups. When you look at the flood
26:35of messages in group chats, it feels
26:37like people talking about which girl
26:40group released a new song, what the
26:42style is, or what dance they're doing.
26:46But if you chase after things that way,
26:48there is no end to it. Just feeling
26:49burdened by "technical debt" and
26:51struggling. Getting all stressed out.
26:52Thinking, "I feel like I need to know
26:54this," there is no need to be like that
26:55. There's no need, just study. You just
26:57need to study the basics. A significant
26:59portion gets resolved easily. For
27:01example, many people suffer because of
27:04RAG, right? But the technologies used
27:07in RAG are actually just traditional
27:10search technologies with an LLM
27:12attached. For those who worked on
27:16traditional search, they had no
27:17interest in it until the LLM came out,
27:20and now they ask, "Why does the LLM
27:22talk such nonsense during searches?""
27:25Why is the accuracy so low?" If you
27:27approach it like that, you start
27:28wondering, "Should I use this library
27:30or LangChain?""Should I use LangGraph
27:32or something else that just came out?"
27:34—that’s when you get confused. Yes.
27:36That is why studying the basics is so
27:37important. You said it's not easy for a
27:40newcomer to get into such positions,
27:42and you need to judge that coldly right
27:45now. There are currently a lot of
27:47experienced workers. So, if there is no
27:50change, you have to get in line. If you
27:53want to enter backend development as a
27:54newcomer, you have to be better than
27:56those with experience. It’s not easy
27:58to be better than experienced backend
27:59developers. Because, for example, if
28:01it’s Java Spring, versions are
28:03tangled and complex, and you need to
28:04have that knowledge. To be able to do
28:06that better than someone with
28:07experience. That is impossible. Even if
28:10you're smart and good at algorithms, in
28:13fields where the know-how itself builds
28:16a barrier to entry, a newcomer cannot
28:19just break in like that. If both the
28:24experienced person and the newcomer are
28:26backend, would the experienced dev say,
28:28"I’ll do AI, so you take over the
28:29backend"? Would that happen? That’s
28:32not easy. It might be good in theory,
28:34but it doesn’t work that way. It is
28:36human nature to continue doing what
28:38they were already doing. Even within a
28:40company, if I want to do other work, I
28:42have to move to a different company.
28:44Moving companies is easier. Than
28:45staying within the same company. "I was
28:46originally a backend dev, but I’ll
28:48focus on AI now." If you tell me to
28:49hire a junior and put them in charge of
28:50the backend, do you think I'd hire them
28:51? No, that’s not it; I want someone
28:55who can complement me, someone whose
28:57skill set can balance mine. Oh, if I
29:01feel like I can work with them, then
29:02I’ll hire them. Korean society is
29:05quite dynamic. The new generation wants
29:08to take on new roles. That’s why the
29:11government is pouring money into
29:12projects. They’re saying, "Hey, stop
29:13doing what you were doing and try this
29:14instead.""Do high-tech.""Do something
29:16difficult.""Do something that uses GPUs
29:17.""Do AI.""Do something that can become
29:20a key export industry." They are
29:22driving us in that direction. They’re
29:23pushing us intentionally, but some
29:24think, "Oh, I don't think I can do that
29:26.""I’ll just stick to what I was
29:28doing." Is that even possible? Again, I
29:30want to emphasize that this is only
29:32because of the current situation. In a
29:34normal situation, once you’re settled
29:36in a company, you’d hold your
29:38position, hire some juniors, train them
29:39to grow the team, and get promoted—
29:41that’s how it’s supposed to be. Yes
29:43. I think it would be wise for you all
29:46to cope with this smartly. It would be
29:49a good time to buckle down and study
29:50during the holidays. My lectures are
29:53well-structured for just browsing
29:55through the videos, so it would be good
29:57to watch them during the holidays, and
29:59I hope you spend your time in a way
30:01that helps your growth. Well, let's
30:03wrap up here for today, and I look
30:05forward to seeing you all again next
30:06time. Thank you.

About this transcript

This page contains the full transcript of [편집본] 미국 빅테크 일자리가 한국으로 왔다! by 홍정모, generated from the public captions YouTube serves with the video. The transcript has 5,269 words across 816 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.