메뉴
HN
Hacker News • 25일 전

AI는 당신이 더 빨리 망하는 것도 도와준다

IMP
6/10
핵심 요약

한 컨설턴트 개발자가 AI 코딩 도구의 '10배 생산성' 주장을 실험으로 검증하고 반박하는 글입니다. 오픈소스 LLM이 등장한 지 4년이 지났지만 에어비앤비·스트라이프·드롭박스급 혁신 기업은 나타나지 않았으며, 코딩은 소프트웨어 개발에서 가장 시간을 많이 잡아먹는 병목이 아니라는 것이 핵심论点입니다.

번역된 본문

AI는 당신이 더 빨리 망하는 것도 도와준다

Jordan Andersen, 2025년 8월 17일

얼마 전 Disesdi Shoshana Cox가 쓴 글을 공유한 적이 있다. 그 글에서 저자는 AI 덕분에 지금쯤 몇 개의 혁신적인 테크 기업이 나왔어야 하는지 계산해봤다. 개발 속도가 10배 향상되었다면, 앱을 출시하는 기술적 측면은 더 이상 목표 시장에 도달하는 데 장벽이 아니다. 마치 AI가 드디어 프로그래머가 너무 느리다는 문제를 해결하는 만능 해결책을 제공한 것처럼 보인다.

Disesdi의 산술 논리(반박하기 어려운)를 빌리자면, 오픈소스 LLM이 등장한 지 4년이 지난 지금, AI의 힘 덕분에 에어비앤비 세 개, 스트라이프 두 개, 드롭박스 세 개가 존재해야 한다.

그래서. 대체 그 기업들은 어디 있는가?

GenAI가 등장한 이후 세계에서 가장 큰 새로운 테크 기업은 다름 아닌 GenAI 기업들이다. OpenAI, Anthropic, High-Flyer(DeepSeek 개발사)는 사상 최대 규모의 금액으로 평가받고 있다. 하지만 이들이 일으키고 있는 가장 큰 혼란은 글로벌 경제에 대한 것이다. GenAI가 사회에 기여한 가장 눈에 띄는 성과는 어설프게 작성된 이메일을 일상으로 만들고, 소셜 미디어를 쓰레기로 범람시키고, 인터넷 검색 엔진의 품질을 망가뜨린 것이다.

아니면 인정해야 한다. 이건 사실 당신이 '우주에서 가장 특별한 프로그래머™'처럼 느끼게 만드는 도파민 게임이며, 실제로는 게임화된 대규모 지적 의존일 뿐이라는 것을.

Disesdi의 글이 나에게 와닿은 이유는, 나도 이 AI 열풍 속에서 같은 실망스러운 소프트웨어 산출물을 직접 목격했기 때문이다. 기술적 배경 없이 너드적인 것들을 배우는 데 시간과 노력을 들이지 않은 사람들 손에 이 도구들은 지속 가능하고 일반적인 공격에 대해 보안이 강화된, 작동하는 앱을 만들어내지 못한다. 그리고 경험 많은 소프트웨어 엔지니어가 사용할 때도 그 효과는 힘 없는 악수 같다.

여기서 가장 큰 맹점은 코드를 작성하는 것이 소프트웨어 개발에서 가장 많은 시간을 소모하는 부분이 아니라는 것이다. 이게 '지금쯤 에어비앤비나 드롭박스가 몇 개는 나왔어야 한다'는 내 주장과 모순처럼 보일 수 있다. 하지만 모순이 아닌 것이, 기업과 스타트업의 리더들은 LLM이 제품 출시의 병목을 해결했다고 진심으로 믿고 있기 때문이다. '그냥 Claude한테 시키면 되니까'라는 이유로 소프트웨어 개발의 모든 기술적 문제가 더 이상 존재하지 않는다는 인식이다.

최근 나는 기술 컨퍼런스에 참석했는데, 메인 스테이지에서 '코드 없음, 문제없음(No code, No problem)' 같은 제목의 패널 토론이 있었다. 그 패널은 바이브 코딩(vibe coding)으로 MVP를 만들었고, 아마 고객에게 돈을 받고 있는 네 명의 스타트업 창업자로 구성되어 있었다. 내가 그 세션을 나오게 만든 순간은 창업자 중 한 명이 CTO가 없고 앞으로도 절대 필요하지 않을 것이라고 말했을 때였다. Claude로 기술 문제를 해결하면 되기 때문이라는 것이다.

상상해보라. 소프트웨어 회사에 소프트웨어 전문가가 필요 없다. 사무용 건물을 설계하는 데 구조 엔지니어가 필요 없다. 수술을 하는 데 외과의사가 필요 없다.

I. 실험

나보다 훨씬 뛰어난 개발자들이 Claude와 다른 코딩 에이전트를 사용해 소프트웨어 작성 생산성을 높였다고 말하는 것을 들었다. 그래서 생각했다, 한번 해볼까? 컨설팅 회사를 운영하면 이런 종류의 개발 속도를 실제로 돈으로 바꿀 수 있다!

나는 이미 몇 가지 AI 에이전트를 사용해서 지루한 작업들 — 보일러플레이트 코드, 반복적인 SQL 등 — 을 작성하는 데 도움을 받고 있었다. 하지만 이 10배 매직에 동참할 수 있는지 확인해보고 싶었다. 말에 책임을 다하기로 하고, 진행 중이던 프로젝트에 사용할 DeepSeek 크레딧을 10달러어치 샀다.

그 과정은 극도로 짜증났다. 챗봇은 상상할 수 있는 가장 멍청한 짓들을 추천했다. 분명히 하자: DeepSeek이 작성한 코드는 작동은 했다. 하지만 그건 덕트 테이프로 바퀴를 붙여놓고 굴러다니는 광대차였다.

이쯤 되면 이렇게 생각할 수 있다. '이 사람은 프롬프트를 어떻게 쓰는지 모르는구나. 만약…'

원문 보기
원문 보기 (영어)
AI Can Make You Suck Faster Too By Jordan Andersen on 17/08/26 Not too long ago, I shared a post that I really liked by Disesdi Shoshana Cox . They went through the math of how many revolutionary tech companies we should have by now thanks to AI. With a 10x improvement on development speed, the technical aspects of launching an app are no longer a limit to reaching a target market. It seems like AI has finally provided the silver bullet that solves the problem of programmers moving so slowly. Borrowing Disesdi's logic in their arithmetic (which is hard to argue against), after four years of open source LLMs, we should have three AirBnBs, 1 two Stripes, 2 and three Dropboxes 3 thanks to the power of AI. So. Where the fuck are they? The biggest new tech companies in the world since the advent of GenAI are, well, GenAI companies. OpenAI, Anthropic, and High-Flyer (developer of DeepSeek) are being valued at record-breaking sums of money. But the biggest disruptions they're causing is to the global economy. The most noticeable contribution to society from GenAI is making poorly worded emails the norm, flooding social media with garbage, and ruining the quality of internet search engines . Or else admit this is a dopamine game that makes you feel like The Universe's Most Special Programmer™️ when it's really just gamified mass-scale intellectual dependency. Disesdi Shoshana Cox Disesdi's post struck a chord with me because I've witness the same underwhelming output of software during this AI rage. These tools, in the hands of non-technical people who have not put in the time and effort to learn about the nerdy stuff, are simply not able to generate working apps that will persist and be security-hardened against run-of-the-mill attacks. And when experienced software engineers use them, the impact feels like a limp handshake. The massive blindspot here is that writing lines of code isn't the part of software development that drains the most time. This may seem like I'm contradicting myself on the argument that we should have at least a handful of AirBnBs or Dropboxes by now. But there isn't a contradiction because leaders of companies and startups truly believe that LLMs have solved the bottlenecks in product delivery. It's the perception that all technical issues in software development no longer exist because you can "just get Claude to do it". 4 I recently attended a tech conference where there was a panel discussion on the main stage that was titled something like "No code, No problem". The panel comprised four startup founders who vibe coded their way through MVP and, assumedly, are taking money from customers. 5 The moment that made me walk out of the session was when one of the founders said he didn't have, and won't ever need, a CTO because he can just use Claude to solve his technical problems. Imagine that. You don't need software expertise in your software company. You don't need a structural engineer to design the structure of an office building. You don't need a surgeon to perform surgery. I. An Experiment I've been told by people who are much, much better developers than I am that they've sped up their productivity in writing software by using Claude and other coding agents. So I thought, what the heck? When you run a consultancy, you can actually convert that type of development speed into money! I had already used a few different AI agents to help write the boring stuff - boilerplate code, repetitive SQL, etc. But I wanted to see if I can get in on this 10x magic. I decided to put my money where my mouth was and I bought $10 worth of DeepSeek credits to use with a project I was working on. 6 The process was incredibly infuriating. The chatbot recommended some of the dumbest shit you could possibly do. And let's be clear here: the code DeepSeek wrote would run , but it was a clown car rolling around with wheels held on by duct tape. Now, you might be thinking, "Well this fucking guy doesn't know how to prompt . If he was just better at prompting , he wouldn't suck so much." And you might be right. But I've also shipped a product before , which is infinitely more than most people accomplish in their entire engineering career. So maybe I do suck at prompting. Or maybe something else sucks. II. Today's Sad Reality Until very, very recently, the way you would find the answer to something you didn't know was by Googling it, reading a bunch of different opinions, and filtering and combining those thoughts to generate your own position. And before Googling was the go-to method for finding the answer to something, you'd have to travel to a monstrous building full of dusty humans so you can search, aisle by aisle, for the right collection of tomes that each contained a portion of the thing you wanted to understand better (I like to call it " doing it in the stacks" whenever I visit a library). That's all been replaced with a text war between you and a robot that spews whatever shit some asshole on Reddit posts . Just take a second to digest this stat!!!: Reddit outranks financial experts 176% of the time when ChatGPT answers finance questions, despite YMYL guidelines prioritizing authoritative sources. Carlos Silva, writer for Semrush Too many people don't think for themselves enough in this post-GenAI world. The curiosity hasn't left us - which is very encouraging - but instead of thinking critically about a problem or question, there's an app for that. The issue is that the "app for that" can't reason. Its source of information is a range of differing opinions mostly from non-experts that aren't verified for legitimacy. There used to be a barrier to entry into skilled domains - which was a good thing. For instance, would you ever do the electrical wiring for your own home? Well, why not? You can get a step-by-step guide on your phone while staring at your switch board (that is, if ChatGPT recommended starting at the switch board). The chatbot can probably look up ways to connect your lights to one circuit and your oven to another. You may even learn how to spread parts of your home across different breakers to balance load. But what if you make a mistake? What are the consequences? Maybe an electrical fire in the middle of the night, or maybe you electrocute yourself before that happens. The obvious, and potentially catastrophic, consequences are probably enough for most of us non-electricians to hire an expert. So why isn't the same logic used when building software? Most apps these days gather credit card details and enough personal identifiable information to ruin someone's life if it gets leaked to the wrong person. All it takes is some bored 16-year-old with an internet connection somewhere on this planet to infiltrate a poorly guarded production database. The outcome won't be as cinematic as a house fire, but your life can be completely fucked by careless software design. On top of that, people who are new to software development aren't the only ones who have access to these chatbots. If you really believe in this 10x effect of AI on productivity, imagine how it's going for hackers. It's the same idea when it comes to vibe coding an app, charging people for it, and collecting sensitive data. Except you're not gambling whether your amateur electrical work will burn down your house. Instead, you're gambling with other people's money (from VCs or your customers), privacy (anyone who's trusted you with their data by using your app), and livelihood (every single person you've hired in your company). III. The Wizard of Oz I think the path to how much trust we put into GenAI was paved with "good enough" information that these AI chatbots provided. These LLMs curate a vast amount of information in seconds about topics the user has no idea about. The chat box and conversational feel has replaced the painstaking process of doing it yourself. (As an aside, it's incredibly hilarious that Googling something has become a pain point. A microsecond-in-human-