Transcript of 백엔드가 LLM 공부하면 AI 엔지니어
홍정모
0:01In my last live stream, I focused on0:04 jobs where major companies in the U.S.0:07 West Coast come directly to Korea to0:10 hire talent. Now, looking at job trends0:15 within the U.S. itself, there's a very0:17 interesting shift developing. The0:20 industrial landscape in the U.S. is0:22 quite different between the West and0:23 the East. The East is a bit more0:25 traditional. While represented by Wall0:27 Street, it is actually the hub for law0:29 and politics as well. And in cities0:32 like New York, traditional industries0:35 are aggressively absorbing AI0:36 technology. This trend has actually0:40 been underway for about two years now.0:42 New York's IT workforce has now0:44 surpassed that of Silicon Valley. From0:46 the perspective of 2025, New York0:48 already has more, while the traditional0:50 tech sector has actually seen a0:51 decrease. It's said that the financial0:54 sector has increased its IT workforce0:56 the most. Then there are consulting and0:58 accounting—as those of you who have1:00 used AI know. Using AI for research and1:04 drafting in consulting is great, but1:07 because sending sensitive company1:10 information to external APIs can cause1:13 problems, it must be kept internal. So,1:17 this means there is a need for in-house1:18 capabilities. The idea that companies1:20 need to hire people for this is now a1:22 common narrative. Accounting is similar1:24 as well. So, for sectors like finance,1:26 insurance, and real estate, building1:27 their own AI systems is essential. The1:29 necessity to build these systems is1:31 absolutely critical. And because these1:34 industries must adopt tech, they are1:36 also adopting the human talent to go1:39 with it. Now, let's look at this slide1:41 a little more closely. There is also1:42 information related to shipping,1:43 warehousing, and wholesale. I was1:45 surprised to see even trucking1:47 logistics companies hiring AI engineers1:49. The proportion of AI-related roles in1:51 U.S. IT job postings has risen sharply.1:54 But back in mid-2022, it was actually1:56 too early. Looking at the statistics2:00 from around June 2026, about one-third2:04 already involve AI. So, I analyzed 7202:07 job postings related to AI on the U.S.2:10 East Coast. You can think of this as2:12 cases where AI is included in the2:14 search keywords. You know, like just2:17 using AI? Things like "I know how to2:19 use OpenAI" or "I know how to use2:20 Claude." You can basically consider2:22 those as non-existent. The way it works2:24 is that they just provide retraining2:26 for things like Claude usage. Because2:29 it's just too easy. Whether it's a2:31 financial firm, Goldman Sachs, or any2:33 famous bank like Chase or Morgan2:34 Stanley, those companies just go ahead2:36 and train their own employees. The AI2:39 job postings we're talking about here2:41 are for engineering roles. If you look2:44 at the job titles, most of them are AI2:46 Engineer. They make up more than half,2:49 and then there's the FDE role. FDE2:51 stands for Field Deployed Engineer.2:53 They go to client companies to explain2:56 things and handle negotiations. People3:00 in Korea often compare this to an3:02 upgraded version of SI, but for SI to3:04 be upgraded, one must have foundational3:07 knowledge of LLMs. For those who have3:11 been doing backend or web development,3:13 it's a great time to upgrade your3:15 skills by studying AI technology—or3:17 more precisely, LLM-related tech—to3:19 find a job. There are many3:21 opportunities. I believe this trend3:22 will continue for quite a while. As you3:24 can see here, it says Software Engineer3:27 3, Java, Generative AI, LLM, and so on.3:29 So, what does Java represent? Java is3:31 the symbol of backend development. It's3:33 traditionally the symbol for backend,3:35 but that alone isn't enough anymore.3:37 Generative AI means they want an3:40 engineer who can integrate AI systems3:42 into that company. Then you see things3:45 like Java, AWS, AI/ML keep getting3:47 attached. They don't just want a simple3:51 Java technician; from a leadership3:53 perspective, they want to bring in3:56 someone who can handle AI to build a3:58 team. You can look at it from this4:01 perspective. So, if you look at various4:04 jobs, they all fundamentally include4:06 things like AI, GenAI services, RAG,4:08 agents, tool orchestration, evaluation4:10 frameworks, and guardrails—all the4:13 components needed to build internal AI4:15 systems. Now, if you look at the graph,4:18 From around 2025, job postings related4:21 to AI agents are shooting up. Based on4:24 my research, they want to build4:27 internal AI systems. It doesn't seem4:29 like they are trying to replace people4:31 with AI. It's not about replacement;4:33 developing tools for internal use is4:35 the priority. Yes. Well, the technology4:38 doesn't actually work as well as you4:39 might think. This is the point I've4:41 been making all along. In the past, the4:45 technology was so unstable that we4:46 needed researchers who could study it4:49 before actual engineers could work with4:51 it. Now, LLM technology itself,4:53 including various models and4:54 open-source options, has reached a4:56 fairly high, standardized level. Since4:59 the technology has become much more5:00 stable, you can now do this work even5:02 if your ML skills aren't that strong.5:05 So, the difficulty level of jobs has5:06 decreased significantly. Well, now if5:08 you study a little, do some Python, and5:10 learn the basics, and since there are5:13 books and courses available, you can do5:16 it if you put in the effort. Also, in5:19 the current situation, there isn't one5:21 specific framework that dominates AI5:24 system development. If you look at job5:26 postings, you'll see two strangely5:27 different things together. There are5:29 cases where both Java and Python appear5:30 in the same job posting. It's a very5:32 unusual situation. It means we cannot5:34 abandon the ML ecosystem built on5:36 Python. So, here, let's look at the AI5:39 engineer, AI agent engineer, and the FD5:42 roles, just these three for now. First,5:45 the AI engineer basically builds the5:48 company's internal system. They build5:50 internal systems, but they were5:52 actually just search tools. But with5:54 the advent of LLM technology,5:55 performance has jumped significantly.5:58 It has become much easier for people to6:00 work, and even those with less6:02 knowledge or those who haven't worked6:04 at the company for long can get more6:06 accurate information from a good AI6:08 system even if they explain things6:11 vaguely. Moving on, for AI engineers,6:14 they build these types of internal AI6:16 systems, and I want to emphasize that6:19 it's crucial for them to have a solid6:21 grasp of LLM technology. The next step6:24 is the AI agent engineer. You can think6:27 of the AI agent engineer as someone who6:29 includes all the tasks an AI engineer6:31 can do. Performance has improved by6:34 incorporating agent patterns, such as6:36 tool usage and ReAct patterns, where it6:39 acts, evaluates the results itself, and6:42 then retries or tries a better approach6:44. In terms of building those agent6:47 systems, some companies are now hiring6:49 AI agent engineers separately. On the6:52 flip side, FD (Field Developer/Engineer6:55) roles involve being sent to client6:58 companies to handle communication.7:00 Anyway, for the FD role, if you have a7:02 sufficient understanding of AI, there7:04 will likely be many opportunities.7:06 Looking at LinkedIn job postings, there7:09 are actually more AI engineer roles7:10 than ML engineer roles. The time it7:12 takes for jobs to transition from the7:15 US to Korea has shortened dramatically.7:17 For MLOps roles, it used to take two to7:19 three years. It has become much faster7:21 now. In the current situation, everyone7:23 is scrambling to integrate generative7:25 AI into their company systems as7:27 quickly as possible. So, when these7:30 companies adopt AI—and by AI, I don't7:33 mean just using a Claude API, but7:36 rather building their own internal7:38 systems. To build those systems, they7:41 need to hire people. The companies that7:44 are hiring are all, well, they're great7:46 companies. So, on the left are the7:48 companies on the US East Coast. On the7:49 right are the corresponding Korean7:51 companies. I’ve analyzed over 2,0007:54 job postings in Korea related to AI and7:57 LLMs. It turns out that about one-third8:00 are for AI, LLM, or ML engineers and8:02 researchers, while the others—like8:04 planning, management, consulting, and8:06 sales—aren't necessarily hiring for8:08 those roles, but rather roles that8:10 utilize LLMs and AI. Basically, every8:13 role here is a job that uses AI. I’m8:15 just looking at the ratios among them.8:17 It’s not that there aren't jobs that8:20 prioritize backend experience. They8:22 exist, but in terms of proportions, the8:24 new types of jobs have a much higher8:26 ratio. On the left, I’ve organized AI8:29 engineer job postings from Wall Street8:31 in New York. And on the right, we have8:34 AI engineer openings at Hyundai Card.8:37 As you can see, it mentions things like8:39 "experience with RAG or agent8:41 frameworks." Yes. RAG here refers to8:45 taking traditional search technology—8:47like the internal database search8:49 systems used within a company—and8:51 layering an LLM on top to find relevant8:54 data even in ambiguous cases. To put it8:57 simply, that’s about as much as I can8:59 say for now. Yes. Korean banks are also9:01 building AI agents. You can think of9:04 the people building these AI agents as9:07 AI engineers or agent engineers. For9:10 instance, in September alone, four9:12 banks posted openings for LLM9:14 developers, including KB Kookmin Bank,9:16 NH Nonghyup, and Hana Bank, while9:18 companies like Kia are looking for9:20 backend developers, including for GPU9:22 serving systems, which requires a fair9:24 amount of skill. That's closer to MLOps9:27. But there is also service backend9:29 platform development. Jobs are9:31 continuing to emerge in various forms,9:33 so those who understand AI can apply to9:35 roles involving Java, Kotlin, and9:37 Spring-based AI service platform9:39 development or core business logic and9:41 common platform functions. Next, look9:44 at Danggeun, which has a community9:45 backend role. It's a software engineer9:47 backend role, clearly a backend9:48 position, yet it mentions AI. Here,9:51 they talk about Spring AI, the OpenAI9:54 SDK, and LLM-integrated AI agents. Jobs9:58 in this field are clearly a major trend10:00. Now, there is actually a bit of a10:01 hopeful side to this. Surprisingly,10:04 surprisingly, there are job postings10:05 where you don't have to be an expert;10:07 you can apply just by having a proper10:09 understanding of LLM principles. Yes.10:11 If you look at the job postings, they10:13 require a foundation in operational10:15 principles and about a year of relevant10:17 experience, so the requirements have10:19 eased significantly. I have statistics10:22 showing that out of 132 job postings in10:24 Korea, quite a few list "understanding"10:26 as the essential requirement for LLMs.10:29 So, there are various job postings that10:31 list "understanding" as the essential10:32 requirement for LLMs. For example, you10:34 need to be able to structurally explain10:35 the operational principles, failure10:36 modes, hallucinations, and grounding10:37 departures of LLMs. Next, for AI agent10:40 developers. People who possess an10:43 understanding of this. Yes. It doesn't10:46 ask for a specific number of years of10:47 experience, a flashy portfolio, or a10:49 background at a big tech company, does10:50 it? People who possess an understanding10:53. Next, those with an understanding of10:55 LLM application technologies like NC,10:56 MCP systems, RAG, embedding, and vector10:58 search—if you just have this level of11:00 understanding, it's much easier.11:02 Besides that, there are many other11:04 companies like this, yes. So, in my11:07 view, the area most advantageous for11:09 you right now in terms of volume is11:11 definitely backend.11:13>> So, they are hiring for AI. If they are11:16 hiring for AI, backend experience is11:17 often considered a fundamental11:19 requirement. This is just a slight11:20 shift in perspective. Does this mean11:22 you must do backend if you want to do11:24 AI? That’s not the case. Yes.11:27 Analyzing the statistics, a significant11:29 portion of AI job postings require11:32 backend skills, which means if you are11:34 a backend developer, learning LLMs will11:37 drastically increase your hiring11:39 opportunities. But doing AI itself is11:43 not necessarily the same as being a11:45 backend developer. I hope there's no11:47 misunderstanding about that. Conversely11:50, if you have some backend experience11:52 but aren't finding the right jobs. And11:55 you're trying to apply here and there11:56 but not having much luck. For those11:59 people, studying and moving toward jobs12:01 that involve AI is a much more12:03 advantageous situation. Yes. That is my12:06 thought. Yes. Also, the technologies12:08 listed in AI agent job postings are12:09 surprisingly similar between Korea and12:11 the US. So, for things like RAG, MCP,12:13 or LangGraph, they aren't always used,12:15 but you should be able to use them if12:16 needed. Yes. Even for vector databases,12:19 there's talk about whether they're12:20 becoming obsolete with new technology,12:21 but you should still be aware of them.12:22 And using AI coding tools. Yes. This is12:26 a different matter from being an AI12:28 engineer. So, using AI coding tools is12:30 just a basic requirement. Yes. It's a12:31 basic. So, for example, wanting to gain12:33 a competitive edge just by using Claude12:36 Code—That doesn't make any sense. You12:38 should assume that everyone is already12:40 using them. So, as I said again, the12:42 jobs appearing now are for building AI12:45 systems. In some ways, I see basic LLM12:47 knowledge not just as a preference, but12:49 as a necessity moving forward. Now,12:51 looking at the statistics, there were12:53 also many job postings open to12:54 entry-level applicants. Even in those12:57 job postings, having a sufficient13:00 understanding of AI and LLMs is, yes,13:03 an unavoidable requirement. So, while I13:06 don't have expertise in backend13:08 development, from the perspective of13:10 someone who has taught you for a long13:12 time and watched you struggle with the13:14 shortage of jobs, I’ve shared this13:16 because it seems like a great13:17 opportunity. I believe you absolutely13:21 must study LLMs. I wouldn't expect you13:25 to study at the level of an ML engineer13:27 or researcher, and people at that level13:29 wouldn't even be applying for these13:31 roles anyway. Can I become an AI13:34 engineer too? Yes, you can. The LLM13:36 course I am currently creating is more13:38 of a conceptual course. In the current13:41 situation, it's about going into an13:43 interview and giving off the vibe—13:44That you've done your homework. Yes.13:46 It’s designed as a top-down course13:48 that instills those concepts in you. I13:51 understand that you may feel like13:53 it’s lacking. I have intentionally13:56 compressed this course heavily,13:57 focusing only on the core essentials so13:59 you can study quickly. But there is14:01 this advantage. Once you watch this14:04 course and grasp the concepts, studying14:06 on your own becomes much easier. Yes.14:08 So I think it would be good to approach14:10 your studies from that perspective. The14:12 easiest way to study is to just build14:14 things from the ground up. In the past,14:16 building from scratch meant typing14:18 everything out yourself, so you might14:20 have some resistance to the idea, but14:22 if you look at my course, I use AI14:24 starting from Part 1. I teach you how14:26 to work by instructing Claude. So, your14:28 shoulders won't even hurt. We’ve14:30 entered an era where you can build from14:32 scratch just by "mouth-coding." It's a14:34 fantastic time to study. Oh, there’s14:35 a misunderstanding here. Companies14:37 aren't expecting you to have poured14:40 tons of money into preparing a14:42 portfolio. And even if you’ve built14:44 an LLM, it’s about having tried it14:46 with minimal cost, like with an Llama14:48 model. That should be enough. If you14:51 want to keep up with the latest machine14:53 learning trends, looking into LLMs is14:55 the most efficient way. Because that's14:57 where all the resources are being14:59 poured into. When you study, don't draw15:01 strict lines between fields, thinking,15:03 "I'm in computer vision, I work with15:05 OpenCV, so I don't need to study LLMs."15:07 That's not how it works. Fields are15:09 just blending together now. If you're15:11 curious about something and feel like15:14 giving it a try, I believe it's the15:16 right path to trust your instincts and15:18 grow by satisfying that curiosity. If15:20 you're confused about where to start,15:22 there's a free course called "Deep15:24 Learning First Steps." Yes. You should15:26 watch that first. Being an engineer15:28 just means you never stop studying.15:32 Even if you make a lot of money and15:33 reach retirement, or even if you reach15:35 a position where you just manage others15:36, you still need to stay on top of the15:38 technology. So, while the level of15:40 study might change, Yes. that's the way15:42 you should look at it. But now,15:43 studying has become much easier. There15:47 are more resources, and if you use AI15:48 as a tool, you can make it much more15:50 fun and engaging, so you just have to15:52 study until you die. Yes. That is the15:55 right mindset. Yes. There was a time15:58 when I thought I could learn a skill16:00 once and live off it for the rest of my16:02 life. That era is over. The idea that16:05 you just need to get your work done is16:05 wrong. It doesn't make sense because,16:07 these days, people often struggle at16:09 work not because they can't handle the16:11 tasks, but because they can't adapt to16:13 rapidly changing technology. Especially16:15 since the rise of AI, if you understand16:17 the fundamentals and control AI well,16:19 you can excel, so those with weak16:21 foundations take a bigger hit. A16:23 massive hit. That's why you need to16:25 clearly distinguish who said what and16:28 in what context. You shouldn't just16:30 blindly accept everything you hear.16:33 Alright, that's it for today, and I'll16:34 see you all next time.
2,858 words · 446 lines
![[선데이 라이브] 당신도 AI 시대에 살아남을 수 있다!](https://i.ytimg.com/vi/D95UC9U5ShA/hqdefault.jpg)
