소프트웨어 개발에서 AI의 역할에 대한 비판

### AI 생산성의 환상: LLM과 코딩에 대한 비판적 관점
* 🤖 저자는 AI 도구를 둘러싼 과대 광고, 특히 **와이콤비네이터의 과대 광고**가 기술의 기능에 대한 비현실적인 기대를 조장한다고 생각하며 불만을 표합니다.
* 💬 이들은 **게리 탄의 트윗을 인용**하며 기술 리더들이 AI의 이점을 선전하는 동시에 투자를 홍보하는 경향이 있어 그 동기에 대한 회의론이 제기되고 있다고 강조합니다.
* 🛠 저자는 클로드와 같이 AI의 잠재력을 인정하면서도 **"사용해 본 적도 없고 앞으로도 그럴 것"**이라며 그 우수성에 대해 확신을 갖지 못하고 있으며, **생산성을 진정으로 향상시키는지 직접 테스트해 볼 계획**이라고 밝혔습니다.
* 📊 이들은 **AI가 간단한 작업의 속도를 높일 수는 있지만, 복잡한 코딩 작업에는 여전히 인간의 독창성과 문제 해결 능력이 필요**하며, 현재로서는 AI가 효과적으로 모방할 수 없다고 주장합니다.
* ⚡ 저자는 **LLM이 효율성이라는 착각**을 불러일으켜 숙련도가 낮은 개발자가 이해 없이 빠르게 코드를 생성할 수 있게 하여 궁극적으로 품질이 떨어지고 코딩 작업이 지나치게 단순화될 수 있다고 주장합니다.
* 🧠 이들은 프로그래밍에서 기초 지식의 중요성을 강조하며, AI 도구에 의존하면 개발자가 기술을 연마하는 대신 이러한 도구에 의존하게 되는 **'학습된 무력감'**을 조장할 수 있다고 주장합니다.
* 📉 저자는 AI가 생성한 코드의 확산이 숙련된 개발자의 깊이와 이해가 부족한 부실하게 구축된 솔루션으로 업계가 포화 상태에 이르러 **코딩 품질 저하**로 이어질 수 있다고 경고합니다.
* 💡 결론적으로, AI 도구는 특정 영역에서 도움을 줄 수 있지만 **진정한 코딩 숙달은 자동화로 대체할 수 없는 원리와 관행에 대한 확실한 이해**가 필요하다고 주장합니다.

AI 도구에 대한 Y Combinator의 과대 광고는 창업자들이 코딩 기술 없이도 성공할 수 있다고 착각하게 만듭니다.

This is just so... this is just I'm trying not to be triggered by it, and we're not even... we're not even into it, believe this, but a couple days ago I made the rookie mistake of opening up Twitter and what I saw shocked me. There's this tweet from

Gary Tan, who runs Y Combinator right now. And it's always Y Combinator, I swear it's always Y Combinator, and it's not that I'm just trying to dunk on Y Combinator because Paul Graham uh blocked me. I really wish I had a better

opinion of Y Combinator, but I don't. I just don't. I have no good opinion about it. I think it's all nonsense. I just think that they're there to sell you hype tools constantly. And so of course, absolutely, of course, Mr. Y Combinator is

going to talk about how great it is to use Cloud Sonnet 3.5 and then right afterwards post about how great AI is, and then right afterwards sell you on a company that they've been invested into. It's always this, every single time. This so he's reacting by the way.

Look at this. I'm streaming. I'm streaming. Yeah, yeah, yeah, yeah. I'm streaming. I'm in the video. Did I make an appearance? Am I in? Oh, I'm in. Oh, I'm in. Let's go. Let's go. And code to a post that was made on Reddit.

Feeling very powerful as a technical founder with Cloud Sonnet 3.5. It's mindblowing how quickly I... wait, did he just... is it not Sonnet? Wait, what? How did he say it? Am I saying it wrong? I could very well be.

saying it wrong because I I'm actually stupid founder with Claude at 3.5 it's mind-blowing how quickly I can move and they're not saying LLMs in general but specifically Claude like I'm pretty sure

I could implement copies uh quick quick quick question is Claude that much better I'm saying it right okay let's go let's go uh so I thought it was on it everyone keeps telling me that Claude is actually good like actually good it's definitely it's definitely so it okay

Claude is good a bit better not just good just better okay no not good but better uh never used it never will yeah I'm going to use it I'm going to use it and show you if it's good or not and I'm going to fully believe and try to commit

to it right that's like the upcoming thing is that we're going to I have a feature that I want to build in for those that don't know we have the Tower Defense of course which we're almost done with which by the way is looking

pretty good these days which has a player one on one side and has player two on the other side this is going to be ChatGPT um and this is going to be Twitch and maybe I'm going to make this also become say uh some other program

maybe this could be that Claude uh or whatever it is just like KET um it's going to become that but what I wanted to do is I wanted to add a feature I wanted to add a second tower called an ice tower and an ice tower is simply

going to make enemies slow down wherever they're walking now how I did pathing in this game is I did pathing based on a boolean right and so I have a list of bools and so anywhere that's a false creeps cannot walk anywhere that's a

true creeps can walk so in my head how to make the ice tower the best test would be to turn this from a boolean array into a float array and the float array allows me to be able to say what speed they can run at a zero makes them not

able to run at all where A1 is full speed that means we should be able to add in an ice tower because an ice tower will simply slow it down it'll be a 0.5 right and so this is kind of Big Brain Change so I already know how to

make the change or how I would implement it so what I want to do is actually have Cloud come in here cloud cloud not cloud cloud I'm sorry I I'm thinking of cloud flare uh the NSA backed company but uh instead Cloud I would like to see how

does cloud work those p uh those are weights will those affect pathing yes they will affect pathing I'll have to make it so that even pathing then works so will the creeps avoid your ice tower potentially if it actually makes it makes them slower or I won't let it

affect uh pathing maybe we'll do both I don't know anyways it's kind of an interesting if it's a float it acts as a multiplier it would be parallel with any pathfinding uses movement cost well yeah so it's kind of a fun thing it's

kind of a fun idea right so there you go I think that's a really great use case for an AI can I do a bigger project because remember if we look in here and we go uh we go here and go like this uh find this grep Zig uh not that one there

we go there's about 22 there's about 4,200 lines of code of Zig I don't make a lot of comments I'm not not a big commenter so it's mostly code and some white space so that's pretty dang good

that's a pretty good amount of code in fact we could probably even do something that's like uh what is it grep uh grep -v uh something that starts with a new line or starts ends with a new line right okay apparently I'm terribly apparently I'm really bad at the apparently I'm

really bad at the old uh the old uh whatever it's called Uh I tried to use slock but slock doesn't work with uh Zig so I just said forget about it all right anyways let's keep on going these are the technical parts of most

popular apps in the App Store 10 times as fast as I could before LLMs I still need to make architectural and infrastructure decisions but stuff like programming the functionality of a UI component is 10 times faster now literally 10 times faster so either this

guy is using literally as figuratively or literally as literally and so that's a pretty big ass claim can we just pause for a second to actually say you are 10x faster like remember what this means that means you should be able to do your job in one day and not work

for the rest of the week and nobody would even know you're not doing things I actually do agree with this I forget who said it first I really actually want to say Luke Pagetti Luke P Luke Pagetti Luke Luke Luke P Luke hey cross 200,000

let's go right there let's go thank you very much it just happened crazy um all right hold on I I think it was Luke Pagetti that actually first made this uh observation uh let's go let's go right here I'm just going to say it people who

say this uh Tech is making them 10x or 0.1x Engineers to begin with OR they're working on 0.1x problems so Luke I mean Luke genuinely made a pretty pretty based statement and you know I I am I'm

willing to be wrong because you know I have this perception of AI but I also want to be faster at writing code my entire life my entire system every single time I want to change anything I change exactly where I want to go I can change between projects I do everything

everything has been aligned such that when I want to do something I try to do it as fast as possible and so if I can become 2x faster that's a huge amount of benefit for me so if Cloud really is that much better then yes I will

actually use it from here on out if I can be 2x faster I will hands down 100% use it that way speed an obsession absolutely it's an obsession I'm willing to wrong this results in such fast iteration speed and they kind of outline

the steps to do that I think they're completely right when it comes to like code snippets and like small pieces of code LLMs have been pretty good and it seems like they're improving. Claude is even better now and right now if you

look at the replies a lot of them are pretty reasonable but when I first saw this post I can tell you everybody in the comments was acting like they lived in a different universe. Like I was thinking either I'm crazy or all these people are crazy but now we've

got that. That is such a good way to say it. I do feel this exact same thing cuz I I don't understand this obsession with like say co-pilot. I don't understand the obsession with chat GPT when people keep telling me they use chat GPT all the

time I sit here and think I'm like what am I doing wrong? Am I that bad at programming that I like am I really that bad? I have to be that bad because I am not getting the same result at all. I feel pissed off constantly and I feel like, you know, have

you ever watched someone pair program and all you want to do is go get out of the way, right? That's how I feel. It's like I just feel like the guy that's just like or Nick Burns, right? Nick Burns, your company computer guy. I sit there and

watch and be like just okay, could you just do okay? Could you just move right? Like I just want to do the thing. I feel like Nick Burns constantly. You've got some more reasonable people that are actually kind of questioning this because what exactly are you guys coding

or is anybody coding where LLMs are making the work 10 times as fast? It's what are you I again I I completely agree with that statement. That's why I think Luke is right, you're working on a problem that can be so concisely

described and the code so isolated that it can be produced faster. Like that's the only way I can think of it. Baffling to me. I've been using them the last couple years just like everybody else and I feel like I'm pretty familiar with

their limitations and I made a response to this because I genuinely meant this sincerely. I just don't understand, am I not doing it right? Anytime I need to accomplish something of medium difficulty, LLMs waste my time. Hey, I have a CSV full with 10

different columns here's the columns and I need to be able to graph these three columns in Python using matplotlib. K thanks bye. Because for complex tasks they just don't know what they're doing. I've tried like this is not something I'm making up. It's from experience using

them for snippets is fine but who's writing their entire app or even designing their entire app or even like the dude think about how Devon is feeling right now. Devon is in literal shambles, like actual shambles right now. Devon is dying, he's actually dying.

complex parts of it with the help of LLMs. If you're building something so simple that you don't need to write a quick doc or list out edge cases yourself or even design like the major tradeoffs to be fair, I don't do that uh

Typically what I do is that I use the power of my brain, think about the problem for a moment, design the entire thing in my head in a matter of seconds, and then just start typing until I cannot stop typing. Are people out there paying for glorified snippets? They are.

I I don't I don't design it like I literally even uh the the this the sweet um our our delicious little Tower Defense, which is fantastic, goes through here and it just it it all just works via via via the old brain. I think that most people I

think most people undersell themselves in how fast they can actually realize stuff, right? Like I think they by the way this is just me typing, printing out a bunch of stuff. Don't worry, it's a simulation, but you get the idea that I think you're I think a lot of people

forget how good they actually are at solving problems, and so they see some something new and fancy and they feel like oh I just need to use that now, and they use it not realizing that they actually are just smart enough, but this feels like it's supposed to be easier.

And by it feeling like it's supposed to be easier, you just like lean on it, and when you lean on it, you it's like you learned, yeah Jeremy, exactly. It's like learned helplessness. It's solved one of your problems really quickly, and so you

just keep on leaning on it over and over and over again. There are times in which it's like really useful, and I will say this, like uh I don't remember what the antsy code is for coloring backgrounds. I know it's it's it's it's it's it's

Escape Open Bracket 48 semicolon 2 semicolon uh your your true color M right I'm pretty sure that's what it is uh but that's like a background color so for me to learn that what I did is or for me to know that I just asked chat

GPT to give me a couple different an code stuff like hey explain an an code for me it did a really good job because there's like a there's a definite answer for that it's not a hard lookup but I just wanted two sentences so I found

documentation search it's actually pretty good at like if you want to do good stuff with documentation I feel like uh an LLM is actually really really good summarize this and give me like the the three sentence answer it's very great at but everything else I just I've

yet to find something that it's good at that I actually could use it in principle for for programming then maybe sure like that was my hypothesis and it's from experience and I said I would love to watch a non-technical person create neat code and this was

actually more in response to another tweet I saw that was saying that like non-technical people can now be Founders and build tech companies which which by the way that oh you know I humbly would disagree with and I me I nonh humbly

disagree with I whole ass wholeheartedly disagree that if you're not a technical person and you think you can found a technical company because you can use cloud Sonet like I hate to tell you this but you're you have been sold such a

just such a lie right they can but they should I don't even think they can right like what are you going to do Beyond like problem number one it's like peak delusion like you do have to learn something this is this is like Agent Orange level snake oil mention my site

is actually pretty simple I would love to watch somebody create it and so many people saw this as a challenge and some people disagreed with me have you tried Claude it absolutely shits on GPT 4.0 yeah but you know I had a response to

that which I'll probably link in the comments but this person yes Swix I'd like this because he is most certainly top shiller of them all uh he you know he is really good friends with the Devon team and he is talking about how great

Devon is and you can't even just you just wait it's going to be the greatest thing in the universe and so I would love to see actually I would actually like to see him do this right because you know what I felt like the Devon stuff I felt like

Devon's alive I still feel like Devon's alive I feel like people that that promoted Devon should get kicked in the shin you know know not I don't want to hurt them but they need to be like oh ow oh gosh owie

oh and like you know go on the ground like Peter Griffin for a moment and just like have just have a second for you know just have like a good three you know a good two-minute like ouch session I'm not going to insult them I have

nothing against them, I don't even know who they are to be honest, but apparently they're a Founder. I mean, let's be honest nowadays, like everybody's a Founder. You just start a company and you can call yourself a Founder. Found... I'm a Founder, found his nuts in

your mouth. Founder is more of a vanity title I think at this point, but whatever, but I respect this person because they took time out of their day to actually try it out. And so the first screenshot they sent though was not promising. They're using Devon AI. Whoa, guys, now I'm

nervous. Devon AI, it's definitely going to put me out of business pretty soon. And I said, "Looking forward to it, especially the code sandbox functionality, which I honestly thought he was going to get closer than he did." The result will... um, again, I'm not making

fun of the person, I'm just talking about the results. So they said, "Okay, I got pretty far after some prompting." And they did this in a day, like a full day, it looks like. I don't know how many hours they spent on it, a decent amount, I'm

sure. And so with a day of prompting, this is how far they got. Let me show you. Sure, it's not pretty. There's no Dark theme, it's not pretty, but like the main functionality looks like it's there, right? I mean, I don't know why there's a

progress bar on the homepage myself. The LLM doesn't know why there's a progress bar either, but this kind of looks... I mean, I'm not going to lie to you, this looks like it's some sort of progression. So obviously, there's a

progress bar I mean let's be real here if there's something that looks like a progression there obviously must be a progress bar I mean it did get the pro button let's go let's go upsell hold on we got we someone someone Someone got

something in here uh non-technical Founders thinking they can start a tech company using AI are the exact type of an numpty that Google the best free programming tool heavy on the free because they're not they're going to cuss C cuss everywhere possible or every

way possible gosh I cannot read uh sorry I have dyslexia uh saw that Visual Studio code comes up as number one option and decided that's what's going to be used for the development team oh what does their company do build iOS apps imagine having to do that with iOS

apps site did not have that but you know hallucinations are a thing and so there's a bunch of pages and and it looks like the content is kind of there did we just see and it looks like the cont CSS is hard do we just see

ourselves a little bit of CSS as hard yo dog I mean that's a that's a pretty cool Flex box you got I like how none of it's even like aligned it's like this cascading upsetness that's happening join Discord I wonder what the

Discord link is what's the dis what do you think the Discord link is content is kind of there but first of all it's all skeleton like nothing is implemented not a single thing not even the problem list from n code iio which I mean I'm not

trying to plug it I'm just trying to show you what it looks like like by the way you should plug it in he did a great job with what he's doing he built like a useful version of leak code uh you should go out and check n code I

okay he did a great job he's very very smart he's actually achieved real things in his life I I fully think that this is good there's a list of problems that you can mark complete and that's basic CRUD functionality I would not consider that

a medium difficulty task so that is not done at this point so this is just an ugly UI skeleton and one major thing I noticed is I can't even get to the homepage anymore there's no button to get back to the homepage nice wow okay

this roadmap feature right wow like AI did this that is kind of impressive I mean it doesn't really do anything right now but look at mine it's not about the functionality the feature parody but this is very different from this one you're probably thinking well it got you

like 50% of the way there however you want to calculate it got you pretty close but pretty close in I mean close counts in horseshoes and hand grenades close does not count with this the two things that you just saw are so vastly

different that making one into the other is rewriting it like it's not even the thing is is that close in these kind of situations is completely redoing it it's just simply not the same categorically by the way priority queue rating at the very least AI was able to

save time, but the problem is I don't think this component is even compatible. Like, no matter which human or how hard they try or how smart they are, I don't think a human would be able to fix this and make it like mine, or even close to

mine with a progress bar with actual functionality. Yeah, look at, look at that. There's like little progress parts like they took the least hard part of building it, which is a bunch of CSS lines and boxes, and said it's the same thing. It's obviously

not the same thing because like look at this one, these are like a coalescing node point. The other one does not have any coalescing, it only simply has shortcuts. Like that's it, there's no actual, there's no, there's no

actual intelligence behind it. It gives the appearance of some. This, I mean, this again, this is exactly AI in a nutshell. It gives the appearance of something that is smart, but it itself is not smart at all. It's not actually smart, it looks smart, right? Like look at this, notice

that it's there's no coalescent here, there's no coalescing because you can go here to here to graphs, right? You can go from tries to graphs. That's not even an option in what he did. It it's a reflection of something

that's made by humans, but it's not anything. It's just a really bad simulation, but on first glance, this looks actually really impressive, right? When I first saw this I was like, 'Wow, that's actually wow. Okay, that, that's a pretty cool little UI. Okay, like you

could even highlight each one of the like lines that you've already closed that you've already done you could make them green the ones that you skipped you could make red right you could do all these things that feel really good but then I realized it's the yeah exactly

it's the Uncanny Valley but for programs it looks like it should be good but when you apply any level of critical reasoning you go oh this doesn't make any sense why would I go from arrays and hashing to two pointers to binary search to tries to graphs to

greedy to bit manipulation right like it like the the path doesn't even make any sense like nothing about it makes sense it's just dare I say the Kamala Harris of UI this and make it like mine or even close to mine with a progress bar sorry I've

just been on the Kamala Harris cuz she she just gives me the Uncanny Valley vibes she's been looming me for years at this point I can't even help it I I'm just so unburdened by what has been I

just can't even help it with actual functionality making it movable so the problem with that is this will have to be rewritten from scratch right it's not a bug fix it will have to be reimplemented from scratch exactly none even the links make sense absolutely so

it looks like AI in this particular case didn't save time it wasted time who who could have possibly predicted that I told them the exact same thing it's not compatible it's going to have to be rewritten from scratch and it's been a

day and I think they gave up. I don't think they're going to keep trying. If they do, I'll update you guys because again I don't have anything against this person, but again it's Prime. What if it gets better? How much better, like how

much better does it have to get? It has to get a lot better. We're talking a lot, a lot, a lot, a lot, a lot better, by at least seven better units, right? Like it has to get incredibly better. Now here's the thing is you're making an assumption

and I've said this a lot of times and I always, I don't know why people get so angry about this, but you got to remember that in the in the in the 60s through the 70s, uh, AI was like, like there's a

lot of improvements in AI, then effectively from the 70s until about the early 2000s AI did nothing, and then from the early 2000s to about the 2010s AI did something and then nothing for a long time, and then LLMs kind of hit the scene

and that, you know, behind the scenes that was obviously happening in research, you just didn't see it. And so there feels like there's this huge pop in AI ability, right there, right? It feels like a holy new object, but then what happens right here? Does it flatten out? Does it

just keep on popping? Does it pop, but just not at as fast of a rate? Does it actually go down because we are poisoning the well? We are effectively producing content for it to be trained on that's been produced by itself, thus it gets worse and worse and worse and worse?

like think about how much GitHub code that has been produced in the last few months that actually has more bugs, more problems, and it's all being produced by co-pilot. Like there's this really interesting study that shows that the average line modification in GitHub has

effectively gone from 6 months to two weeks. Like that that's that's going to cause inbreeding, right? This is inbreeding at its finest. And so that's the problem is that where what is going to happen? So when someone says it's going to get better, I

say why the perception of better is really just human skills getting worse, possibly. You might be right or it might it could also be that we've become reliant on, say, ChatGPT 3.5, and so when 4 came out it felt better. It I mean

it could it could just simply feel better, but maybe your base productivity was already better to begin with. You just simply have been reinforced to think in a certain way. Inbreeding AI is crazy. It is it's just I I just don't understand like why are people making

NeetCode's Hot Take Is SO Good

This is a place for all the things that are awesome on stream. 

ThePrimeTime

와이콤비네이터의 AI 도구 홍보에 대한 비판적 시각

화자는 Cloud Sonnet 3.5에 대한 Gary Tan의 트윗에 대해 과대 광고라고 생각하며 불만을 표출합니다.

연사는 와이 컴비네이터가 투자한 AI 도구와 회사를 지속적으로 홍보하는 것에 대해 비판합니다.

연사는 와이 컴비네이터가 진정한 혁신보다는 과대 광고 도구 판매에 초점을 맞추고 있다고 생각합니다.

연사는 와이콤비네이터가 AI를 홍보한 후 곧바로 관련 기업을 홍보하는 패턴을 지적합니다.

연사는 Claude 3.5의 효과, 특히 코딩 속도 향상에 대한 주장에 의문을 제기합니다.

화자는 Claude의 실제 기능을 확인하기 위해 테스트하기로 결정합니다.

이 발표자는 Claude를 사용하여 타워 디펜스 게임의 새로운 기능을 개발할 계획입니다.

화자가 게임과 계획 중인 '얼음 탑' 기능에 대해 설명합니다.

연사는 게임 내 경로 탐색에 대한 접근 방식과 Claude가 이를 최적화하는 데 어떻게 도움을 줄 수 있는지 설명합니다.

LLM의 10배 생산성 주장에 대한 토론

발표자는 LLM을 사용하면 코딩 속도가 10배 빨라진다는 주장에 대해 분석하며 그 타당성에 의문을 제기합니다.

발표자는 특정 문제가 있을 때만 10배의 개선 효과가 나타난다는 Luke Pagetti의 관찰을 언급합니다.

발표자는 코딩 프로세스에서 속도를 우선시하며, 진정으로 속도를 향상시킬 수 있다면 Claude를 사용할 의향이 있습니다.

발표자는 작은 코드 스니펫에는 LLM의 유용성을 인정하지만 대규모 프로젝트에서는 그 효과에 의문을 제기합니다.

LLM 과대 광고에 대한 개인적인 경험과 회의론

발표자는 Copilot 및 ChatGPT와 같은 LLM을 둘러싼 과대광고에 대한 불만을 표출합니다.

발표자는 LLM에 대한 자신의 경험을 비효율적인 쌍 프로그래밍을 보는 느낌과 비교합니다.

발표자는 LLM이 소규모의 고립된 코딩 작업에 더 효과적이라는 데 동의합니다.

발표자는 LLM이 효과적이지 않았던 작업의 예를 들며 그 한계를 강조합니다.

기술 비전문가인 창업자의 Devon AI 사용 경험 분석

이 연사는 기술 비전문가인 창업자가 Devon AI를 사용하여 앱을 개발하려는 시도에 대해 설명합니다.

발표자는 결과를 분석하면서 일부 기능이 만들어졌지만 상당한 결함이 있다고 지적합니다.

연사는 열악한 UI와 누락된 기능 등 앱의 문제점을 강조합니다.

스피커는 AI가 생성한 앱과 자신의 앱을 비교하여 품질과 기능의 엄청난 차이를 강조합니다.

AI에 대한 장기적인 전망과 한계

연사는 향후 AI의 발전 가능성에 대해 논의하지만, 상당한 발전이 필요하다고 강조합니다.

연사는 AI의 역사적 발전 과정을 언급하며 급속한 발전과 정체기를 강조합니다.

연사는 AI 근친 교배로 인한 품질 저하로 이어질 수 있는 잠재적인 부정적인 결과에 대한 우려를 제기합니다.

연사는 AI의 개선이 인간의 기술력 감소로 인한 것일 수 있다고 제안합니다.

연사는 인공지능이 주니어 엔지니어의 업무를 자동화할 수 있다는 주장에 의문을 제기합니다.