AI가 코딩과 개발에 미치는 영향에 대한 라이너스 토발즈의 이야기

### 라이너스 토발즈가 말하는 AI와 코딩의 미래
* 👨 Linux 커널과 Git의 창시자 라이너스 토발즈(Linus Torvalds)가 인터뷰에서 프로그래밍에서 인공지능의 역할에 대한 인사이트를 공유하며 **이 주제를 논의할 수 있는 자신만의 특별한 자격**를 강조합니다.
* 🤖 그는 대규모 언어 모델이 진정한 지능을 발휘하기보다는 주로 다음 단어를 예측한다고 주장하면서 AI를 **"스테로이드에 의한 자동 수정"**이라고 설명하지만, 코딩 관행에 AI가 미치는 영향이 크다는 점은 인정합니다.
* 💻 토발즈는 AI가 코드 작성에 점점 더 많은 도움을 주겠지만, 이는 **기계 코드에서 파이썬이나 러스트 같은 고급 언어로 발전**해 온 프로그래밍 자동화의 오랜 역사에서 다음 단계에 불과하다고 생각합니다.
* 🔍 그는 현재 도구가 더 미묘한 오류로 어려움을 겪을 수 있지만 **명백한 버그를 잡아내어 코드 리뷰를 개선할 수 있는 AI의 잠재력**에 대해 낙관적인 견해를 밝히며, 코딩 과정에서 사람의 감독이 중요하다고 강조합니다.
* 📉 토발즈는 AI가 프로그래밍의 특정 측면을 혁신할 수는 있지만 **인간 개발자의 필요성을 완전히 없애지는 못할 것**이라고 말하며 AI를 둘러싼 과대 광고에 대해 경고합니다.
* 🛠 ️ 그는 사용자가 점점 더 자연어 쿼리에 의존하여 AI로부터 직접 기능을 얻음에 따라 **애플리케이션 개발이 줄어들고 기존 애플리케이션의 필요성이 줄어들 수 있는 미래**를 상상합니다.
* 🌐 마지막으로 토발즈는 오픈 소스 알고리즘도 중요하지만 다양하고 풍부한 데이터에 대한 접근이 AI 기술 발전에 매우 중요하다고 주장하며 **AI 개발에서 오픈 데이터의 중요성**을 강조합니다.
* 💬 이 토론은 코딩에서 AI의 의미에 대한 시청자의 생각을 공유하도록 초대하여 **기술 커뮤니티에서 진행 중인 논쟁**을 강조하며 마무리됩니다.

라이너스 토발즈는 AI가 코딩을 향상시킬 것이라고 믿지만, 그 기능을 과대 포장하는 것은 경계해야 한다고 말합니다.

Linus Torvalds is one of the most famous and influential computer scientists in the entire world, and he gave his thoughts on using artificial intelligence to code, and it is fascinating. So, we're going to watch the interview with him together in this video. So, before we get into the video, let me just show you a little bit about Linus. So, first of all, he's the creator and lead developer of the Linux kernel, no big deal, and also the creator of Git, the versioning software that basically every single software company uses. He created Linux as a free and open-source alternative and is still actively involved to this day. So, if anybody is qualified to give their opinions on artificial intelligence and coding, it is him. Let's watch this video. So, this guy on the left is Linus Torvalds. I typically say artificial intelligence is autocorrect on steroids because all a large language model does is it predicts what's the most likely next word that you're going to use, and then it extrapolates from there. So, not really

very intelligent, but all right. Obviously, this guy has a very strong opinion of AI, and this has been debated over and over again. Is all AI doing predicting the next token, or does it actually have a deeper intelligence? And right now, it's probably leaning more towards just predicting the next token, which in itself is an intelligence, but fine. Soon, we will have other technologies, other techniques that go way beyond just predicting the next token. Let's keep watching. Obviously, the impact that it has on our lives, on the reality we live in, is significant. Do you think we will see LLM-written code that is submitted to you as a request? I'm convinced it's going to happen. Yes, I, and it may well be happening already, maybe on a smaller scale, where people really use it more as a help in writing code. But automation has always helped people write code. I mean, this is not anything new at all. And by the way, this video came out at least 7 months ago, so it's a little bit older.

but he already says, yeah, it's probably happening already. Fast forward 7 months, it's definitely happening, and especially with the explosion of cursor usage, the IDE that is AI native. Everybody seemingly is building code, even if you don't know how to code. But then he continues and says, automation has been helping, and yeah, Automation and programming go hand in hand, and this is just the next level of automation. We don't write machine code anymore, we don't even write assembler, and now we're moving on from C to rust. So, I don't see this as something as revolutionary as all the news talk is every day about AI. So, what he's explaining is how abstraction works. In the very beginning, we were writing ones and zeros, then we had machine code, and then we continued to abstract layer after layer on top of that to make it easier for humans to write code that machines can understand. And today, we don't write machine code, we write these high-level languages: C, rust, Ruby, Python, all of these languages that then get translated down into binary in

which a computer can actually understand and run. And that's what he's explaining. This is the next level and potentially the last level of abstraction. It is natural language. No longer do you have to be precise and exact with your explanation of what you want out of the functionality of code. You just explain it in fuzzy terms, and the large language model will then translate that into a coding language that will eventually be translated all the way down to machine code and binary. Now, here's the interesting thing that I've been thinking about for a while. Eventually, large language models and AI, in general, will write more and more of the code that we use every day. And at a certain point, the large language model probably will write code in a language that doesn't look familiar to us at all because they don't need to. The reason why we've had these layers of abstraction on top of machine code and binary is because humans are really bad at writing code. We can't read binary. We can't even read machine code very well. It takes us a long time to

understand what's going on, but large language models don't have that problem. So eventually, we might see a coding language that is completely foreign to us. Maybe it's even symbols. Who knows what it's going to look like? Because the only ones that are really going to have to understand it is the AI. And yeah, there's a whole set of implications that come from that. We're not going to touch on that in this video. Let's keep watching. But so you say you expect this can help people write code. This can help people get started. But then if we look at the previous conversation and the challenges around code reviews and maintaining. So do you think that large language models will get to the point that they can help us review code? That they can help maintain subst. I hope. I hope. Because that's certainly one of the areas where which I see them really being able to shine. To find the obvious stupid bugs. Because I mean, how many people here are actually programmers in this room? A lot. So a fair number. A lot.

of the bugs I write, write a lot of the bugs I see other people write. They're not like subtle bugs; a lot of them are just the stupid bugs that you did not think about, and you don't need any kind of higher intelligence to find them. All right, so he is 100% true. I was just talking to a buddy of mine who works at Microsoft, and they're using AI to do code reviews. He said for the most part, it's not super helpful. I think it'll get there, but this is his opinion. It wasn't super helpful right now, but the point is the bugs that it does catch that are helpful are the ones where maybe it's a misspelling or it's like an off-by-one error or something like that. Very obvious bugs that could and should have been caught by a human, but for whatever reason, it didn't. But it might not currently be able to catch the subtle bugs, the issues with the functionality, the issues with the business logic, but maybe we'll see. But having tools that warn, I mean, we

have compilers that warn about the obvious, really obvious ones, but having LLMs that warn about slightly more subtle cases where it may just say, This pattern does not look like the regular pattern. Are you sure this is what you mean? So, we already have something called linters, which checks your code for syntactical errors and other obvious things that can be programmed to check against. But now what he's explaining is large language models can maybe check more sophisticated errors, which is looking at a pattern that you used in your code and saying, Hey, I see you've used this other pattern a bunch of times, and you're maybe trying to replicate that pattern, but it's not exactly right. That would be really awesome. The answer may be no, That was not at all what I meant. That you found an obvious bug. Thank you very much. So, I do think that LLMs are going to be a big, you call them disparaging generally, like autocorrects on steroids, and I actually think that they're way more than that. And how most people work is we

All are autocorrects on steroids to some degree, and I see that as a tool that can help us be better at what we do. Yeah, so look, he called it autocorrect on steroids, and Lus said, Well, then we're all autocorrect on steroids. Humans learn from the environment. We ingest all of this data through our eyes, our ears, our different senses. We learn from it, and we develop pattern recognition, which is essentially how large language models work. And when we're talking, when we're responding, maybe we are just trying to predict the next token in our own sentence. We could just be very sophisticated large language models. But I've always been optimistic, the whole hopeful, hopeful, the word helpful, hopeful and humble, hopeful and humble, that's my middle name. But on the other hand, I mean, I have been so optimistic that 32 years ago, I was stupid enough to think that you can write a better Colonel than anybody else. So you have to kind of be a bit too optimistic at times to make a

difference. So, my approach to Yens really has been that, Hey, this is wonderful. I love seeing the optimism. I don't necessarily share it. Now, a lot of people disagree with one of the things that I worry about in all this is we see the hallucinations. We see, and that's a technical term for LLMs. They do hallucinate, and they do make up stuff. And so, the more they are being put into the position where they will automatically do things without an actual human being there to catch them, the more this becomes scary. Not scar as in they will rule the world and not in the Sci-Fi sense, but in the so many bugs that will happen and that will affect our lives or our code. So, he's talking about hallucinations, and if you watch this channel, you know what that is. It's basically when large language models get something wrong but say it with such confidence that it looks like they think it's right. Now, here's the thing. Humans make mistakes. Large language models being absolutely perfect is probably never going to

happen. We're probably going to continuously reduce the number of hallucinations that large language models generate through techniques like QAR and strawberry long-term thinking multi-agent systems, other ways to verify and validate the outputs, but it probably won't ever be perfect. But that's why you write tests and you do other things to protect the code against bugs that humans would write. We basically have to treat the large language models the same way. Well, I see the bugs that happen without them every day, so that's that may be why I'm not so worried. I think we're doing those just fine on our own. All right, so he has the same sentiment as me basically. Humans make mistakes, large language models make mistakes. We have to have systems in place that hopefully prevent those mistakes from making it into production and also minimizing the resulting impact of those mistakes, especially for mission-critical code. Now, here's another interview with the same guy, Lonus Tals, and again he talks about AI with coding, and this is a more recent interview, so

let's watch it together. Let's go for something that's fun and entirely non-controversial. So let's talk a little bit about AI and large language models. And so obviously this is the current hype kit on the blog. And if you want to double your salary, add AI to your title. And it's hilarious to watch the self-description of companies. And every company on Earth has an AI angle to their story as of this year. It is, yeah, it's very true. Every single company is now powered by AI. Pottery Barn powered by AI. And it basically 10x is your valuation overnight. But what I find so interesting is this idea that geni is going to be the end of, you know, insert whatever, end of programmers, the end of authors, the end of movie creators, the end of so many jobs. So you are going to be replaced by an AI model finally. Yeah, no, I don't know. I hate the hype. At the same time, I think AI is very interesting. I, when I went to

University, we were still talking about rule-based expert systems and Bayesian logic and all of this kind of thing, and it was called AI, but it was not in any way intelligent. Um, and I do think that the last few years have... two days, yeah, but uh, I don't want to be part of the hype, and uh, I will say that the AI revolution has had, even on the kernel level, a few positives. For example, uh, a company like Nvidia, who is not exactly famous for being great at interacting with the kernel community, has actually been much more active and been involved in the Linux memory management code because suddenly they start caring about Linux when they are selling a lot of AR hardware. I mean, it used to be crypto, and it's still obviously GPUs, and it's being used in big servers and running Linux, so it has actually had a positive impact. But my personal op and is, let's wait 10 years and see where it actually goes before we make all these

crazy announcements of your job will be gone in 5 years. I kind of disagree, especially for programmers. I really think 10 years from now, there's going to be very few application developers. What is the point of building an application when you can get the exact information that you want from a large language model directly? You're simply going to say what you want, and it will give it to you. Let's use a really basic example, a to-do app. Rather than having a to-do app, you simply tell the large language model, Remind me of this thing at this time, and it will know to store it. It will know when to remind you, and there's no need for the actual to-do application because the large language model is the application. And that can apply to any amount of sophistication in the application that you want. You simply tell the large language model what you want, and it will give it to you. That's why I truly believe maybe 10 years is a bit aggressive, but the application is going away. There's just no reason for it.

It's going to be natural language to AI direct to compute. I think we already see lasting and almost irreversible change happening through many of these AI tools, not on the hype genie front, but just in the tools that make our lives better and easier. I keep calling it autocorrect on steroids, and people are mad at me, but if you use a generative AI tool, hopefully an SLM tool, to help you with basically code completion in your editor, isn't that a huge opportunity? So, I'm actually, I mean, I'm one of those people who are very optimistic about AI, and I'm looking forward to the tools to actually find bugs. We have a lot of tooling around the kernel and around any software projects, obviously, and we use them religiously, but making the tools smarter is not a bad thing. I to some degree compare it to writing things in assembly, which literally I started doing with the initial kernel, was I think about 50% Assembly Language and using a compiler.

and using Smarter Tools is just the next inevitable step. So, that's going to happen, but I don't think it's necessarily the gloom and doom that some people say it is, and I definitely don't think it's the promised world that the people who are having their hand out for cash say it is. So, you need to be a bit cynical about this whole hype cycle in the tech industry. I hope you all realize that before AI, it was crypto; before crypto, it was whatever. It's a cloud-native. Cloud-native, um, so there's a loud voice in the front that's not hype, okay? I mean, the hype, there's always like a grain of reality behind it, but you need to be care. Yeah, these things that they're mentioning that are hype, uh, these are real trillion-dollar companies and products, so I don't know how much hype it really is. Cloud-native, I mean, the cloud infrastructure powers everything on the internet nowadays, so it's certainly not hype. I look at these tools, and I look at you said assembler.

and then we talk about compilers. Then we talk about things like sparse a compiler that tries to find certain kinds of bugs. We talk about Julia to do code refactoring. We have had for the last 30 plus years a sequence of tools that helped make development better and more robust. And in that lineage of things, I hope there is really cool stuff coming down. Yeah, I mean, um, we have tools that do kernel rewriting with very complicated scripts and pattern recognition and things like that. And that is actually literally why I think um, AI can be a huge help because some of these tools are very hard to use because you have to specify things at a low enough level that the natural reaction would be to hey, can we make this a bit easier and automate more of it. So what he's describing is deterministic scripts code that could improve other code. And that's how it's been done a lot of essentially decision trees if this then that. And that can only go so far because those get incredibly complex really quickly. And so

if we have this non-deterministic way through large language models to essentially apply the pattern recognition without having to specify every single rule, all of a sudden we have this enormously powerful and valuable tool that we can use to improve our code, even write our code that wasn't previously available. One of the interesting angles that this whole large language model and the training data brings up is the role that data plays in our modern world, where we all talk about open source, about the source code, the algorithms being available, but open data really is kind of the almost the more interesting question. No, it's not. Well, I mean, it's not to me, and actually, I'd like to clarify that. I mean, the LF obviously has open data projects, and to other people, it's more interesting, and that is, I think, the whole, for me, the point of Open Source is that different people are interested in different things, and I was always interested in the low-level nitty-gritty of how the... All right, so let me just stop.

him for a second. I actually think the guy on the right is correct. He said open data is actually more interesting than open algorithms. Look, both of them are important, but the data is seemingly harder to come by. Algorithms are out there, and they're being used. Every time a research paper comes out, the entire community can use it, understand it, and try to implement it in the way that they best see fit. But the data seems to be behind closed walls for the most part, and all of the data that is open has already been used to train models. So, they need new data, they need more data, and where do you get that from? That's why open data is very, very interesting. So, that is what Linus Torvalds, one of the most famous programmers in the history of programming, has to say about AI and programming. And what do you think? I have pretty strong thoughts about this, and I've talked about it at length in previous videos, but I want to know what you think. Drop your comments below and

If you enjoyed this video, please consider giving a like and subscribe, and I'll see you in the next one.

Linux Creator Reveals the Future Of Programming with AI

Artificial Intelligence (AI), Open Source, Generative Art, AI Art, Futurism, ChatGPT, Large Language Models (LLM), Machine Learning, Technology, Coding, Tutorials, AI News, and more

** Exclusive Pinecone Partner **

Matthew Berman

라이너스 토발즈의 AI와 코딩에 대한 관점

Linux 커널과 Git의 창시자 라이너스 토발즈가 코딩에서 인공지능의 역할에 대한 자신의 생각을 공유합니다.

그는 현재 인공지능의 기능은 주로 다음 단어를 예측하는 것이라고 생각하지만, 코드 작성에도 상당한 영향을 미친다고 인정합니다.

토발즈는 코딩 자동화를 자연스러운 발전으로 보고 있으며, AI가 결국 익숙하지 않은 언어로 코드를 작성하여 인간 중심의 추상화 계층에 대한 필요성에 도전할 것이라고 믿습니다.

코딩에서 대규모 언어 모델의 잠재력

대규모 언어 모델은 코드를 이해하고 생성하는 것이 주된 목표이기 때문에 결국 인간에게는 낯선 코딩 언어를 만들 수 있습니다.

이러한 모델은 인간 프로그래머가 종종 놓치는 명백한 버그를 식별함으로써 코드 검토 및 유지 관리에 중요한 역할을 할 수 있습니다.

현재로서는 미묘한 버그나 복잡한 문제를 잡아내지는 못하지만 코드에서 잠재적으로 문제가 될 수 있는 패턴에 대해 경고하는 기능은 개발자에게 유용한 도구가 될 수 있습니다.

AI에서 낙관주의와 환각의 역할

연사는 인간과 대규모 언어 모델을 비교하며 둘 다 환경으로부터 학습하고 패턴 인식을 발전시킨다고 제안합니다.

연사는 AI의 잠재력에 대해 낙관적이지만 착각과 실수의 위험성을 인정하며 이러한 오류의 영향을 최소화할 수 있는 시스템의 필요성을 강조합니다.

연사는 인간의 실수로부터 코드를 보호하는 방법과 유사하게 대규모 언어 모델을 다룰 것을 주장하며 검증 및 유효성 검사 기술의 중요성을 강조합니다.

현재 AI와 대규모 언어 모델은 모든 기업이 AI로 구동된다고 주장하면서 큰 과대광고를 일으키고 있습니다.

연사는 특히 프로그래밍과 애플리케이션 개발 분야에서 AI가 일자리를 대체할 것이라는 이야기에 대해 회의적인 입장을 표명합니다.

이들은 미래에는 대규모 언어 모델이 사용자가 원하는 정보와 서비스를 직접 제공할 수 있기 때문에 기존 애플리케이션의 필요성이 줄어들 수 있다고 생각합니다.

AI 도구는 소프트웨어 개발에서 지속적이고 중요한 변화를 주도하며 코드 완성 및 버그 탐지의 기회를 제공하고 있습니다.

어셈블리 언어에서 더 스마트한 AI 기반 솔루션에 이르기까지 개발 도구의 진화는 자연스러운 과정으로 여겨지고 있습니다.

AI를 둘러싼 과대 광고에 회의적인 시각도 있지만, 복잡한 작업을 자동화하고 단순화할 수 있는 이러한 도구의 잠재력은 개발 프로세스에 중요한 기여를 하는 것으로 인정받고 있습니다.

AI 및 프로그래밍에서 오픈 데이터의 중요성

대규모 언어 모델은 코드 개선과 작성을 위한 강력한 도구를 제공하므로 오픈 데이터의 중요성에 대한 의문이 제기되고 있습니다.

오픈 소스 알고리즘은 널리 사용 가능하지만, 오픈 데이터는 구하기 어렵고 모델 학습에 매우 중요합니다.

라이너스 토발즈는 오픈 데이터는 종종 폐쇄된 벽 뒤에 숨어 있고 AI 개발에 대한 수요가 높기 때문에 오픈 데이터의 중요성을 강조합니다.