아모데이 인터뷰

https://youtu.be/n1E9IZfvGMA?si=ZF3xuWmhsCFRW8Yt

Dario Amodei — “We are near the end of the exponential”

So we talked three years ago. I’m curious, in your view, what has been the biggest update of the last three years?
3년 전에 이야기했었죠. 당신 관점에서 지난 3년 동안 가장 큰 변화나 업데이트는 무엇이었나요?
What has been the biggest difference between what it felt like last three years versus now?
3년 전과 지금의 느낌 사이에 가장 큰 차이는 무엇인가요?
Yeah, I would say, actually, the underlying technology, like the exponential of the technology, has gone… broadly speaking, I would say, about as I expected it to go.
네, 기술의 기반, 즉 기술의 기하급수적 성장은… 대체로 제가 예상했던 대로 진행되었다고 말할 수 있습니다.
I mean, there’s like plus or minus, you know, a couple, there’s plus or minus a year or two here, there’s plus or minus a year or two there.
물론 여기저기 ±1~2년 정도의 오차는 있습니다.
I don’t know that I would have predicted the specific direction of code, but actually when I look at the exponential, it is roughly what I expected in terms of the march of the models from like, you know, smart high school student to smart college student to like beginning to do PhD and professional stuff, and in the case of code, reaching beyond that.
코드의 구체적인 방향은 예측하지 못했을 수도 있지만, 모델들의 진행 과정을 보면 스마트 고등학생 수준에서 스마트 대학생, 박사/전문가 수준으로 나아가고 코드에서는 그 이상으로 나아가는 기하급수적 성장은 제가 예상했던 것과 대체로 일치합니다.
So the frontier is a little bit uneven. It’s roughly what I expected.
그래서 프론티어는 약간 불균등하지만, 대체로 예상했던 대로입니다.
I will tell you, though, what the most surprising thing has been.
하지만 가장 놀라운 점은 이것입니다.
The most surprising thing has been the lack of public recognition of how close we are to the end of the exponential.
가장 놀라운 것은 우리가 기하급수적 성장의 끝에 얼마나 가까이 있는지에 대한 대중의 인식 부족입니다.
To me, it is absolutely wild that you have… You know, within the bubble and outside the bubble, you know, but you have people talking about these, you know, just the same tired old hot button political issues and like, you know, around us were like near the end of the exponential.
저에게는 버블 안팎에서 사람들이 여전히 똑같은 진부한 정치적 이슈들을 이야기하고 있는 것이 정말 놀랍습니다. 우리 주변에서 기하급수적 성장의 끝이 다가오고 있는데 말입니다.
I want to understand what that exponential looks like right now.
지금 그 기하급수적 성장이 어떤 모습인지 이해하고 싶습니다.
The first question I asked you when we recorded three years ago was, “what’s up with scaling and why does it work?” I have a similar question now, but it feels more complicated.
3년 전 녹음할 때 처음 물었던 질문은 “스케일링은 어떻게 되고 왜 작동하나요?”였습니다. 지금도 비슷한 질문을 하지만 더 복잡해진 느낌입니다.
At least from the public’s point of view, three years ago there were well-known public trends across many orders of magnitude of compute where you could see how the loss improves.
대중 관점에서 3년 전에는 컴퓨트 규모의 여러 차수에 걸친 잘 알려진 추세가 있었고, loss가 어떻게 개선되는지 볼 수 있었습니다.
Now we have RL scaling and there’s no publicly known scaling law for it.
이제 RL 스케일링이 있고, 이에 대한 공개된 스케일링 법칙은 없습니다.
It’s not even clear what the story is. Is this supposed to be teaching the model skills? Is it supposed to be teaching meta-learning? What is the scaling hypothesis at this point?
무엇을 가르치는 것인지조차 명확하지 않습니다. 모델에게 기술을 가르치는 건가요? 메타러닝을 가르치는 건가요? 지금 스케일링 가설은 무엇인가요?
I actually have the same hypothesis I had even all the way back in 2017.
사실 2017년부터 가지고 있던 같은 가설을 지금도 가지고 있습니다.
I think I talked about it last time, but I wrote a doc called “The Big Blob of Compute Hypothesis”.
지난번에도 이야기했지만, “Big Blob of Compute Hypothesis”라는 문서를 썼습니다.
It wasn’t about the scaling of language models in particular. When I wrote it GPT-1 had just come out.
당시 언어 모델 스케일링에만 국한된 것은 아니었습니다. GPT-1이 막 나왔을 때 썼으니까요.
That was one among many things. Back in those days there was robotics. People tried to work on reasoning as a separate thing from language models, and there was scaling of the kind of RL that happened in AlphaGo and in Dota at OpenAI.
그때는 로보틱스 등 많은 것들이 있었습니다. 사람들은 언어 모델과 별도로 추론을 연구하려 했고, AlphaGo나 OpenAI의 Dota RL 스케일링 등이 있었습니다.
People remember StarCraft at DeepMind, AlphaStar.
DeepMind의 StarCraft AlphaStar도 기억할 겁니다.
It was written as a more general document.
더 일반적인 문서로 썼습니다.
Rich Sutton put out “The Bitter Lesson” a couple years later. The hypothesis is basically the same.
Rich Sutton이 몇 년 후 “The Bitter Lesson”을 발표했죠. 가설은 기본적으로 같습니다.
What it says is that all the cleverness, all the techniques, all the “we need a new method to do something”, that doesn’t matter very much.
모든 영리함, 모든 기술, “새로운 방법이 필요하다”는 것들은 그다지 중요하지 않다는 것입니다.
There are only a few things that matter. I think I listed seven of them.
중요한 것은 몇 가지뿐입니다. 일곱 가지를 나열했던 것 같아요.
One is how much raw compute you have. The second is the quantity of data. The third is the quality and distribution of data. It needs to be a broad distribution.
첫째는 raw compute 양, 둘째는 데이터 양, 셋째는 데이터의 질과 분포(넓은 분포여야 함)입니다.
The fourth is how long you train for. The fifth is that you need an objective function that can scale to the moon.
넷째는 훈련 시간, 다섯째는 달까지 스케일할 수 있는 objective function입니다.
The pre-training objective function is one such objective function. Another is the RL objective function that says you have a goal, you’re going to go out and reach the goal.
Pre-training objective가 그런 것 중 하나이고, RL objective(목표를 설정하고 달성하는 것)도 그렇습니다.
Within that, there’s objective rewards like you see in math and coding, and there’s more subjective rewards like you see in RLHF or higher-order versions of that.
그 안에는 수학과 코딩 같은 objective reward와 RLHF 같은 subjective reward가 있습니다.
Then the sixth and seventh were things around normalization or conditioning, just getting the numerical stability so that the big blob of compute flows in this laminar way instead of running into problems.
여섯째와 일곱째는 normalization이나 conditioning으로, 큰 compute blob이 문제없이 laminar flow처럼 흘러가게 하는 것입니다.
That was the hypothesis, and it’s a hypothesis I still hold. I don’t think I’ve seen very much that is not in line with it.
그게 제 가설이고, 지금도 유지하고 있습니다. 이에 어긋나는 것은 거의 보지 못했습니다.
The pre-training scaling laws were one example of what we see there. Those have continued going.
Pre-training scaling laws가 그 예시였고, 계속 진행되고 있습니다.
Now it’s been widely reported, we feel good about pre-training. It’s continuing to give us gains.
Pre-training은 계속 이득을 주고 있다고 널리 알려져 있습니다.
What has changed is that now we’re also seeing the same thing for RL. We’re seeing a pre-training phase and then an RL phase on top of that.
변화된 점은 RL에서도 같은 현상을 보고 있다는 것입니다. Pre-training 단계 후 RL 단계를 추가합니다.
With RL, it’s actually just the same. Even other companies have published things in some of their releases that say, “We train the model on math contests — AIME or other things — and how well the model does is log-linear in how long we’ve trained it.”
RL도 마찬가지입니다. 다른 회사들도 math contest(AIME 등) 훈련에서 모델 성능이 훈련 시간에 log-linear하다는 것을 발표했습니다.
We see that as well, and it’s not just math contests. It’s a wide variety of RL tasks. We’re seeing the same scaling in RL that we saw for pre-training.
수학 콘테스트뿐 아니라 다양한 RL task에서 같은 스케일링을 보고 있습니다.
You mentioned Rich Sutton and “The Bitter Lesson”. I interviewed him last year, and he’s actually very non-LLM-pilled.
Rich Sutton과 Bitter Lesson을 언급하셨죠. 작년에 그를 인터뷰했는데, 그는 LLM에 그다지 열광하지 않습니다.
I don’t know if this is his perspective, but one way to paraphrase his objection is: Something which possesses the true core of human learning would not require all these billions of dollars of data and compute and these bespoke environments, to learn how to use Excel, how to use PowerPoint, how to navigate a web browser.
그의 반대 의견을 paraphrasing하면, 인간 학습의 진정한 핵심을 가진 것은 Excel, PowerPoint, 웹 브라우저 사용법 등을 배우기 위해 수십억 달러의 데이터와 compute, 맞춤 환경이 필요하지 않을 것이라는 것입니다.
The fact that we have to build in these skills using these RL environments hints that we are actually lacking a core human learning algorithm. So we’re scaling the wrong thing.
이런 RL 환경으로 기술을 주입해야 한다는 사실은 핵심 인간 학습 알고리즘이 부족하다는 신호이며, 잘못된 것을 스케일링하고 있다는 뜻입니다.
That does raise the question. Why are we doing all this RL scaling if we think there’s something that’s going to be human-like in its ability to learn on the fly?
그렇다면 인간처럼 즉석 학습 능력을 가질 것이라 믿는다면 왜 이렇게 RL 스케일링을 하는 건가요?
I think this puts together several things that should be thought of differently. There is a genuine puzzle here, but it may not matter. In fact, I would guess it probably doesn’t matter.
이것은 여러 가지를 다르게 생각해야 할 부분입니다. 진짜 퍼즐이 있지만, 그다지 중요하지 않을 수 있습니다.
There is an interesting thing. Let me take the RL out of it for a second, because I actually think it’s a red herring to say that RL is any different from pre-training in this matter.
RL을 잠시 빼고 보죠. RL이 pre-training과 다르다고 보는 것은 red herring이라고 생각합니다.
If we look at pre-training scaling, it was very interesting back in 2017 when Alec Radford was doing GPT-1. The models before GPT-1 were trained on datasets that didn’t represent a wide distribution of text.
2017년 Alec Radford가 GPT-1을 할 때 pre-training scaling은 흥미로웠습니다. 그 전 모델들은 넓은 텍스트 분포를 대표하지 않는 데이터셋으로 훈련되었습니다.
You had very standard language modeling benchmarks. GPT-1 itself was trained on a bunch of fanfiction, I think actually.
표준 언어 모델링 벤치마크였죠. GPT-1은 fanfiction 등으로 훈련되었습니다.
It was literary text, which is a very small fraction of the text you can get. In those days it was like a billion words or something, so small datasets representing a pretty narrow distribution of what you can see in the world. It didn’t generalize well.
문학 텍스트는 전체 텍스트의 작은 부분이었고, 데이터셋이 작아 일반화가 잘 안 됐습니다.
If you did better on some fanfiction corpus, it wouldn’t generalize that well to other tasks.
Fanfiction에서 잘해도 다른 task로 일반화되지 않았습니다.
We had all these measures. It was only when you trained over all the tasks on the internet — when you did a general internet scrape from something like Common Crawl or scraping links in Reddit, which is what we did for GPT-2 — that you started to get generalization.
인터넷 전체 task로 훈련(Common Crawl, Reddit 등)했을 때 일반화가 시작되었습니다. (GPT-2 때 했던 방식)
I think we’re seeing the same thing on RL. We’re starting first with simple RL tasks like training on math competitions, then moving to broader training that involves things like code. Now we’re moving to many other tasks. I think then we’re going to increasingly get generalization.
RL에서도 같은 현상을 보고 있습니다. 수학 콘테스트 같은 단순 task부터 시작해 코드 등 broader training으로, 이제 여러 task로 확장하며 일반화가 증가할 것입니다.
So that kind of takes out the RL vs. pre-training side of it.
RL vs pre-training 구분을 없애는 부분입니다.
But there is a puzzle either way, which is that in pre-training we use trillions of tokens. Humans don’t see trillions of words. So there is an actual sample efficiency difference here. There is actually something different here.
어쨌든 퍼즐은 pre-training에서 trillions of tokens을 사용하는데 인간은 그렇지 않다는 sample efficiency 차이입니다.
The models start from scratch and they need much more training. But we also see that once they’re trained, if we give them a long context length of a million — the only thing blocking long context is inference — they’re very good at learning and adapting within that context.
모델은 처음부터 시작해 더 많은 훈련이 필요하지만, 훈련 후 million context length를 주면 (inference만 문제) 그 안에서 학습과 적응이 매우 뛰어납니다.
So I don’t know the full answer to this.
이 부분에 대한 완전한 답은 모르겠습니다.
I think there’s something going on where pre-training is not like the process of humans learning, but it’s somewhere between the process of humans learning and the process of human evolution.
Pre-training은 인간 학습 과정과 정확히 같지 않고, 인간 학습과 인간 진화 과정 사이의 중간 지점에 있다고 생각합니다.
We get many of our priors from evolution. Our brain isn’t just a blank slate.
우리는 진화로부터 많은 prior를 얻습니다. 인간 뇌는 blank slate가 아닙니다.
The language models are much more like blank slates. They literally start as random weights, whereas the human brain starts with all these regions connected to all these inputs and outputs.
언어 모델은 blank slate에 훨씬 가깝습니다. random weights로 시작하지만, 인간 뇌는 이미 다양한 영역이 input/output에 연결되어 시작합니다.
Maybe we should think of pre-training — and for that matter, RL as well — as something that exists in the middle space between human evolution and human on-the-spot learning.
Pre-training과 RL을 인간 진화와 즉석 학습 사이의 중간 공간으로 생각하는 게 좋을 것 같습니다.
And we should think of the in-context learning that the models do as something between long-term human learning and short-term human learning.
모델의 in-context learning은 인간의 장기 학습과 단기 학습 사이로 보아야 합니다.
So there’s this hierarchy. There’s evolution, there’s long-term learning, there’s short-term learning, and there’s just human reaction.
이런 계층 구조가 있습니다. 진화, 장기 학습, 단기 학습, 인간의 즉각 반응.
The LLM phases exist along this spectrum, but not necessarily at exactly the same points.
LLM 단계들은 이 스펙트럼을 따라 존재하지만, 정확히 같은 지점은 아닙니다.
There’s no analog to some of the human modes of learning the LLMs are falling in between the points. Does that make sense?
인간 학습 방식 중 일부에는 직접적인 아날로그가 없고 LLM은 그 사이 지점에 있습니다. 이해 되시나요?
Yes, although some things are still a bit confusing.
네, 하지만 여전히 약간 혼란스러운 부분이 있습니다.
For example, if the analogy is that this is like evolution so it’s fine that it’s not sample efficient, then if we’re going to get super sample-efficient agent from in-context learning, why are we bothering to build all these RL environments?
예를 들어 진화와 비슷해서 sample efficiency가 낮은 게 괜찮다면, in-context learning으로 super sample-efficient agent가 나올 텐데 왜 RL 환경을 그렇게 많이 만드는 건가요?
There are companies whose work seems to be teaching models how to use this API, how to use Slack, how to use whatever.
어떤 회사들은 모델에게 API 사용법, Slack 사용법 등을 가르치는 일을 합니다.
It’s confusing to me why there’s so much emphasis on that if the kind of agent that can just learn on the fly is emerging or has already emerged.
즉석 학습 가능한 agent가 나오고 있거나 이미 나왔다면 왜 그렇게 강조하는지 혼란스럽습니다.
I can’t speak for the emphasis of anyone else. I can only talk about how we think about it.
다른 사람들의 강조점에 대해 말할 수는 없고, 우리 생각만 말하겠습니다.
The goal is not to teach the model every possible skill within RL, just as we don’t do that within pre-training.
목표는 RL 안에서 모든 가능한 skill을 가르치는 것이 아닙니다. Pre-training에서도 그렇지 않죠.
Within pre-training, we’re not trying to expose the model to every possible way that words could be put together.
Pre-training에서 모든 가능한 단어 조합을 노출시키려 하지 않습니다.
Rather, the model trains on a lot of things and then reaches generalization across pre-training.
대신 많은 것을 훈련시켜 pre-training 전반에 걸친 generalization에 도달합니다.
That was the transition from GPT-1 to GPT-2 that I saw up close.
GPT-1에서 GPT-2로 넘어가는 전환을 가까이서 봤습니다.
The model reaches a point. I had these moments where I was like, “Oh yeah, you just give the model a list of numbers — this is the cost of the house, this is the square feet of the house — and the model completes the pattern and does linear regression.”
모델이 어느 지점에 도달하면, 집 가격과 평방피트 숫자 리스트를 주면 패턴을 완성해 linear regression을 하는 순간들이 있었습니다.
Not great, but it does it, and it’s never seen that exact thing before.
완벽하지는 않지만, 정확히 본 적 없는 것을 해냅니다.
So to the extent that we are building these RL environments, the goal is very similar to what was done five or ten years ago with pre-training.
RL 환경을 만드는 목적은 5~10년 전 pre-training에서 했던 것과 매우 비슷합니다.
We’re trying to get a whole bunch of data, not because we want to cover a specific document or a specific skill, but because we want to generalize.
특정 문서나 skill을 커버하려는 것이 아니라 generalization을 위해 많은 데이터를 얻으려 합니다.
I think the framework you’re laying down obviously makes sense. We’re making progress toward AGI.
당신이 제시하는 프레임워크는 분명히 타당합니다. AGI를 향해 진전하고 있습니다.
Nobody at this point disagrees we’re going to achieve AGI this century.
올해 안에 AGI를 달성할 것이라는 데 이견은 없습니다.
The crux is you say we’re hitting the end of the exponential.
핵심은 당신이 기하급수적 성장의 끝에 도달하고 있다고 말하는 점입니다.
Somebody else looks at this and says, “We’ve been making progress since 2012, and by 2035 we’ll have a human-like agent.”
다른 사람은 2012년부터 진전해왔고 2035년까지 human-like agent를 가질 것이라고 봅니다.
Obviously we’re seeing in these models the kinds of things that evolution did, or that learning within a human lifetime does.
모델에서 진화나 인간 평생 학습이 한 것 같은 현상을 보고 있습니다.
I want to understand what you’re seeing that makes you think it’s one year away and not ten years away.
1년 후가 아니라 10년 후라고 보는 이유가 무엇인지 이해하고 싶습니다.
There are two claims you could make here, one stronger and one weaker.
여기에는 강한 주장과 약한 주장이 있습니다.
Starting with the weaker claim, when I first saw the scaling back in 2019, I wasn’t sure. This was a 50/50 thing.
약한 주장부터 하면, 2019년에 scaling을 처음 봤을 때 확신하지 못했습니다. 50/50이었습니다.
I thought I saw something. My claim was that this was much more likely than anyone thinks. Maybe there’s a 50% chance this happens.
무언가를 봤다고 생각했고, 누구보다 가능성이 높다고 주장했습니다. 50% 정도일 수 있습니다.
On the basic hypothesis of, as you put it, within ten years we’ll get to what I call a “country of geniuses in a data center”, I’m at 90% on that.
10년 안에 “data center 속 천재 국가”에 도달한다는 기본 가설에는 90% 확신합니다.
It’s hard to go much higher than 90% because the world is so unpredictable.
세상이 워낙 예측 불가능해서 90% 이상은 어렵습니다.
Maybe the irreducible uncertainty puts us at 95%, where you get to things like multiple companies having internal turmoil, Taiwan gets invaded, all the fabs get blown up by missiles.
불가피한 불확실성 때문에 95% 정도일 수 있고, 회사 내부 혼란, 대만 침공, fab 폭파 등의 사건이 있을 수 있습니다.
Now you’ve jinxed us, Dario.
이제 우리를 저주하셨네요, Dario.
You could construct a 5% world where things get delayed for ten years.
10년 지연되는 5% 세계를 상상할 수 있습니다.
There’s another 5% which is that I’m very confident on tasks that can be verified.
검증 가능한 task에 대해서는 매우 확신합니다.
With coding, except for that irreducible uncertainty, I think we’ll be there in one or two years.
코딩에서는 불가피한 불확실성을 제외하면 1~2년 안에 도달할 것 같습니다.
There’s no way we will not be there in ten years in terms of being able to do end-to-end coding.
10년 안에 end-to-end coding은 반드시 가능할 것입니다.
My one little bit of fundamental uncertainty, even on long timescales, is about tasks that aren’t verifiable: planning a mission to Mars; doing some fundamental scientific discovery like CRISPR; writing a novel.
장기적으로도 근본적 불확실성은 verifiable하지 않은 task(화성 임무 계획, CRISPR 같은 과학 발견, 소설 쓰기)입니다.
It’s hard to verify those tasks.
그런 task는 검증하기 어렵습니다.
I am almost certain we have a reliable path to get there, but if there’s a little bit of uncertainty it’s there.
신뢰할 수 있는 경로가 있다고 거의 확신하지만, 약간의 불확실성은 있습니다.
On the ten-year timeline I’m at 90%, which is about as certain as you can be.
10년 timeline에서는 90%로, 확신할 수 있는 정도입니다.
I think it’s crazy to say that this won’t happen by 2035. In some sane world, it would be outside the mainstream.
2035년까지 안 일어난다고 하는 건 미친 소리입니다. 정상적인 세상에서는 mainstream 밖일 것입니다.
But the emphasis on verification hints to me a lack of belief that these models are generalized.
verification 강조는 모델이 generalized되지 않았다는 믿음 부족으로 보입니다.
If you think about humans, we’re both good at things for which we get verifiable reward and things for which we don’t.
인간은 verifiable reward가 있는 것과 없는 것 모두 잘합니다.
No, this is why I’m almost sure. We already see substantial generalization from things that verify to things that don’t. We’re already seeing that.
아니요, 그래서 거의 확신합니다. verifiable한 것에서 non-verifiable한 것으로 상당한 generalization을 이미 보고 있습니다.
But it seems like you were emphasizing this as a spectrum which will split apart which domains in which we see more progress.
스펙트럼으로 강조하신 것은 진전이 더 많은 domain을 구분하는 것처럼 보입니다.
That doesn’t seem like how humans get better.
인간이 발전하는 방식은 아닌 것 같습니다.
The world in which we don’t get there is the world in which we do all the verifiable things.
도달하지 못하는 세계는 verifiable한 것들만 하는 세계입니다.
Many of them generalize, but we don’t fully get there. We don’t fully color in the other side of the box. It’s not a binary thing.
많은 것이 generalize되지만 완전히 도달하지 못합니다. binary가 아닙니다.
Even if generalization is weak and you can only do verifiable domains, it’s not clear to me you could automate software engineering in such a world.
generalization이 약해 verifiable domain만 해도 software engineering 자동화는 명확하지 않습니다.
You are “a software engineer” in some sense, but part of being a software engineer for you involves writing long memos about your grand vision.
소프트웨어 엔지니어이지만, grand vision에 대한 긴 메모 쓰기 같은 부분이 포함됩니다.
I don’t think that’s part of the job of SWE. That’s part of the job of the company, not SWE specifically.
그건 SWE의 일이 아니라 회사 전체의 일이라고 생각합니다.
But SWE does involve design documents and other things like that. The models are already pretty good at writing comments.
하지만 SWE는 design document 등도 포함되고, 모델은 comment 쓰기에 이미 꽤 좋습니다.
Again, I’m making much weaker claims here than I believe, to distinguish between two things.
여기서는 믿는 것보다 약한 주장을 해서 두 가지를 구분합니다.
We’re already almost there for software engineering.
software engineering에서는 이미 거의 도달했습니다.
By what metric? There’s one metric which is how many lines of code are written by AI.
어떤 metric으로요? AI가 쓴 코드 라인 수 같은 metric이 있습니다.
If you consider other productivity improvements in the history of software engineering, compilers write all the lines of software.
과거 software engineering 생산성 향상(컴파일러 등)을 보면 모든 라인을 쓰는 것과 생산성 향상은 다릅니다.
There’s a difference between how many lines are written and how big the productivity improvement is.
라인 수와 생산성 향상 규모는 다릅니다.
“We’re almost there” meaning… How big is the productivity improvement, not just how many lines are written by AI?
“거의 도달”이란 AI가 쓴 라인 수가 아니라 생산성 향상 규모를 의미하나요?
I actually agree with you on this. I’ve made a series of predictions on code and software engineering.
그 점에 동의합니다. 코드와 software engineering에 대해 여러 예측을 했습니다.
I think people have repeatedly misunderstood them. Let me lay out the spectrum.
사람들이 자주 오해했다고 생각합니다. 스펙트럼을 설명하겠습니다.
About eight or nine months ago, I said the AI model will be writing 90% of the lines of code in three to six months. That happened, at least at some places.
8~9개월 전, 3~6개월 안에 AI가 90% 라인을 쓸 것이라고 했고, 일부 곳에서는 실제로 일어났습니다.
It happened at Anthropic, happened with many people downstream using our models.
Anthropic 내부와 다운스트림 사용자들에게서 일어났습니다.
But that’s actually a very weak criterion. People thought I was saying that we won’t need 90% of the software engineers.
하지만 매우 약한 기준입니다. 사람들이 90% 엔지니어가 필요 없어진다고 오해했습니다.
Those things are worlds apart. The spectrum is: 90% of code is written by the model, 100% of code is written by the model. That’s a big difference in productivity.
그것들은 완전히 다릅니다. 90% 코드 vs 100% 코드 — 생산성 차이가 큽니다.
90% of the end-to-end SWE tasks — including things like compiling, setting up clusters and environments, testing features, writing memos — are done by the models.
end-to-end SWE task 90% (컴파일, 클러스터 설정, 테스트, 메모 작성 등)이 모델에 의해 됩니다.
100% of today’s SWE tasks are done by the models.
오늘날 SWE task 100%가 모델에 의해 됩니다.
Even when that happens, it doesn’t mean software engineers are out of a job. There are new higher-level things they can do, where they can manage.
그렇다고 엔지니어가 일자리를 잃는 것은 아닙니다. 더 높은 수준의 관리 업무를 할 수 있습니다.
Then further down the spectrum, there’s 90% less demand for SWEs, which I think will happen but this is a spectrum.
더 나아가 SWE 수요 90% 감소 — 일어날 거지만 스펙트럼입니다.
I wrote about it in “The Adolescence of Technology” where I went through this kind of spectrum with farming.
“The Adolescence of Technology”에서 farming 스펙트럼으로 설명했습니다.
I actually totally agree with you on that. These are very different benchmarks from each other, but we’re proceeding through them super fast.
완전히 동의합니다. 서로 다른 benchmark지만 매우 빠르게 진행 중입니다.
Part of your vision is that going from 90 to 100 is going to happen fast, and that it leads to huge productivity improvements.
90에서 100으로 가는 것이 빠르게 일어나고 엄청난 생산성 향상을 가져온다는 비전입니다.
But what I notice is that even in greenfield projects people start with Claude Code or something, people report starting a lot of projects… Do we see in the world out there a renaissance of software, all these new features that wouldn’t exist otherwise?
하지만 greenfield 프로젝트에서도 Claude Code로 시작하고, 많은 프로젝트를 시작한다고 하지만, 세상에 새로운 소프트웨어 renaissance나 기존에 없던 새로운 기능이 보이나요?
At least so far, it doesn’t seem like we see that.
지금까지는 그렇게 보이지 않습니다.
So that does make me wonder. Even if I never had to intervene with Claude Code, the world is complicated. Jobs are complicated.
그래서 궁금합니다. Claude Code 개입 없이도 세상은 복잡하고 job도 복잡합니다.
Closing the loop on self-contained systems, whether it’s just writing software or something, how much broader gains would we see just from that?
self-contained system loop를 닫는 것만으로 얼마나 넓은 이득을 볼까요?
Maybe that should dilute our estimation of the “country of geniuses”.
그것이 “country of geniuses” 추정을 희석시킬 수 있습니다.
I simultaneously agree with you that it’s a reason why these things don’t happen instantly, but at the same time, I think the effect is gonna be very fast.
즉시 일어나지 않는 이유에는 동의하지만, 효과는 매우 빠를 것이라고 생각합니다.
You could have these two poles. One is that AI is not going to make progress. It’s slow. It’s going to take forever to diffuse within the economy.
두 극단이 있습니다. 하나는 AI가 진전하지 않고 경제에 천천히 diffusion된다는 것.
Economic diffusion has become one of these buzzwords that’s a reason why we’re not going to make AI progress, or why AI progress doesn’t matter.
Economic diffusion은 AI 진전이 없거나 중요하지 않다는 buzzword가 되었습니다.
The other axis is that we’ll get recursive self-improvement, the whole thing. Can’t you just draw an exponential line on the curve? We’re going to have Dyson spheres around the sun so many nanoseconds after we get recursive.
다른 극단은 recursive self-improvement로 Dyson sphere까지 즉시 간다는 것입니다.
I’m completely caricaturing the view here, but there are these two extremes.
과장해서 말하지만 두 극단이 있습니다.
But what we’ve seen from the beginning, at least if you look within Anthropic, there’s this bizarre 10x per year growth in revenue that we’ve seen.
Anthropic 내부에서 본 것은 연간 10배 revenue 성장입니다.
So in 2023, it was zero to $100 million. In 2024, it was $100 million to $1 billion. In 2025, it was $1 billion to $9-10 billion.
2023: 0 → 1억, 2024: 1억 → 10억, 2025: 10억 → 90~100억.
You guys should have just bought a billion dollars of your own products so you could just…
자사 제품을 10억 달러 사서…
And the first month of this year, that exponential is… You would think it would slow down, but we added another few billion to revenue in January.
올해 첫 달에도 exponential이 계속되어 1월에 수십억 추가.
Obviously that curve can’t go on forever. The GDP is only so large.
물론 영원히 갈 수는 없고 GDP 한계가 있습니다.
I would even guess that it bends somewhat this year, but that is a fast curve. That’s a really fast curve.
올해 약간 bend될 수 있지만 매우 빠른 곡선입니다.
I would bet it stays pretty fast even as the scale goes to the entire economy.
전체 경제 규모로 가도 꽤 빠르게 유지될 것 같습니다.
So I think we should be thinking about this middle world where things are extremely fast, but not instant, where they take time because of economic diffusion, because of the need to close the loop.
극단이 아닌 중간 세계 — extremely fast but not instant, economic diffusion과 loop closing 때문에 시간이 걸리는.
Because it’s fiddly: “I have to do change management within my enterprise… I set this up, but I have to change the security permissions on this in order to make it actually work… I had this old piece of software that checks the model before it’s compiled and released and I have to rewrite it. Yes, the model can do that, but I have to tell the model to do that. It has to take time to do that.”
enterprise change management, security permission, legacy software rewrite 등 fiddly한 부분 때문에 시간이 걸립니다. 모델이 할 수 있지만 지시하고 실행하는 데 시간이 필요합니다.
Because it’s fiddly: “I have to do change management within my enterprise… I set this up, but I have to change the security permissions on this in order to make it actually work… I had this old piece of software that checks the model before it’s compiled and released and I have to rewrite it. Yes, the model can do that, but I have to tell the model to do that. It has to take time to do that.”
그런 fiddly한 일들 때문입니다. “enterprise에서 change management를 해야 하고, 설정은 했지만 security permission을 바꿔야 제대로 작동하고, compile/release 전에 모델을 체크하는 old software를 rewrite해야 합니다. 모델이 할 수 있지만, 모델에게 지시하고 실행하는 데 시간이 걸립니다.”
I actually do think that the API model is more durable than many people think.
API 모델은 많은 사람들이 생각하는 것보다 더 durable하다고 생각합니다.
One way I think about it is if the technology is advancing quickly, if it’s advancing exponentially, what that means is there’s always a surface area of new use cases that have been developed in the last three months.
기술이 빠르게, exponentially 발전하면 지난 3개월 동안 새로운 use case surface area가 항상 생깁니다.
Any kind of product surface you put in place is always at risk of sort of becoming irrelevant.
어떤 product surface를 만들더라도 relevance를 잃을 위험이 있습니다.
Any given product surface probably makes sense for a range of capabilities of the model.
특정 product surface는 모델 capability 범위에 맞을 뿐입니다.
There’s always going to be this front of new startups and new ideas that weren’t possible a few months ago and are possible because the model is advancing.
몇 달 전에는 불가능했지만 모델 발전으로 가능한 새로운 startup과 아이디어가 항상 앞서 나갑니다.
I actually predict that it’s going to exist alongside other models, but we’re always going to have the API business model because there’s always going to be a need for a thousand different people to try experimenting with the model in a different way.
API business model은 다른 모델들과 함께 존재할 것이고, 수천 명이 모델을 다양한 방식으로 실험할 필요가 항상 있기 때문입니다.
I think the way to think about this is there’s a frontier of new use cases that are enabled by the new capabilities and that frontier is always moving forward.
새 capability로 enabled되는 new use cases frontier가 항상 앞으로 움직인다고 생각해야 합니다.
And so the API is kind of the way you access that frontier.
API는 그 frontier에 접근하는 방식입니다.
And then there will be various products that are built on top of the frontier at any given time.
그리고 특정 시점에 frontier 위에 다양한 제품이 만들어집니다.
But because the frontier is moving so quickly, those products are always at risk of becoming outdated.
하지만 frontier가 너무 빠르게 움직여서 제품들은 항상 outdated될 위험이 있습니다.
So the API remains this durable thing that lets people access whatever the current frontier is.
그래서 API는 현재 frontier에 접근하게 하는 durable한 것입니다.
Now you could imagine a world in which the exponential slows down or stops, and then in that world the API becomes less important and products become more important.
Exponential이 느려지거나 멈추는 세계에서는 API 중요도가 떨어지고 제품이 더 중요해질 수 있습니다.
But in a world where the exponential continues, which is the world that I think we’re in, the API remains this very important and durable thing.
하지만 우리가 있다고 믿는 exponential이 계속되는 세계에서는 API가 매우 중요하고 durable합니다.
How will frontier labs actually make money?
Frontier lab들은 실제로 어떻게 돈을 벌까요?
I think the API business is a good business and it will continue to be a good business.
API business는 좋은 사업이고 계속 좋을 것입니다.
But I also think that there will be various other businesses that frontier labs do.
Frontier lab들은 다른 사업도 할 것입니다.
Some of them will be things like selling models to enterprises or governments or whatever.
Enterprise, 정부 등에 모델 판매 같은 것들.
Some of them might be sort of doing joint ventures with big companies where you have a big company that has a lot of data or a lot of distribution and you partner with them.
Big company와 joint venture (데이터나 distribution 보유 회사와 파트너십).
I think there will be a lot of heterogeneity in how the business models work.
Business model은 매우 heterogeneous할 것입니다.
I think the thing that people sometimes miss is that frontier labs are not the only ones who will make money from AI.
사람들이 놓치는 점은 frontier lab만 AI로 돈 버는 게 아니라는 것입니다.
There will be a whole ecosystem around them.
주변에 전체 ecosystem이 있을 것입니다.
There will be many applications built on top of the models.
모델 위에 많은 application이 만들어질 것입니다.
There will be many infrastructure companies.
많은 infrastructure 회사도 있을 것입니다.
So the pie is going to be very large and there will be many winners.
파이가 매우 크고 많은 winner가 나올 것입니다.
I think frontier labs will capture a reasonable fraction of that pie, but not all of it.
Frontier lab들은 그 파이의 상당 부분을 차지하겠지만 전부는 아닙니다.
If AGI is imminent, why not buy more compute?
AGI가 임박했다면 왜 더 많은 compute를 사지 않나요?
I think we are buying a lot of compute.
많이 사고 있습니다.
We’re one of the largest buyers of compute in the world.
세계에서 가장 큰 compute buyer 중 하나입니다.
We’re investing billions of dollars in compute.
수십억 달러를 compute에 투자합니다.
But there are constraints on how quickly you can actually deploy that compute.
하지만 실제로 deploy할 수 있는 속도에는 제약이 있습니다.
There are supply chain constraints, there are power constraints, there are data center build-out constraints.
Supply chain, power, data center build-out 제약이 있습니다.
So we’re moving as fast as we can, but it’s not instantaneous.
최대한 빠르게 움직이지만 즉각적이지는 않습니다.
We’re also not the only lab doing this.
우리만 하는 것도 아닙니다.
All the frontier labs are scaling aggressively.
모든 frontier lab이 aggressively scaling 중입니다.
I think the amount of compute that’s being brought online is growing very rapidly.
온라인으로 들어오는 compute 양이 매우 빠르게 증가하고 있습니다.
But there are real physical constraints in the world.
현실적인 물리적 제약이 있습니다.
It’s not like we can just snap our fingers and have 10x more compute tomorrow.
내일 당장 10배 compute를 snap fingers로 만들 수는 없습니다.
We’re pushing as hard as we can on all of those fronts.
모든 front에서 최대한 push하고 있습니다.
Will regulations destroy the boons of AGI?
Regulation이 AGI의 boons를 파괴할까요?
I think regulation is a real risk.
Regulation은 진짜 위험입니다.
There are ways that regulation could go wrong and slow things down or make things worse.
Regulation이 잘못되어 속도를 늦추거나 악화시킬 수 있습니다.
But I also think that there are smart ways to regulate that could mitigate risks without destroying the benefits.
하지만 risk를 완화하면서 benefit을 파괴하지 않는 smart regulation도 있습니다.
I think the key is to have regulation that is focused on the highest risk applications, like biological weapons or cyber weapons, while allowing the beneficial uses to proceed.
가장 높은 risk application (biological weapons, cyber weapons)에 집중하면서 beneficial use는 진행하게 하는 regulation이 핵심입니다.
I think we need to have a balanced approach.
Balanced approach가 필요합니다.
Why can’t China and America both have a country of geniuses in a datacenter?
중국과 미국이 둘 다 data center에 country of geniuses를 가질 수 없는 이유는?
I think they can both have very powerful AI systems.
둘 다 매우 강력한 AI system을 가질 수 있습니다.
But there are differences in the ecosystems.
하지만 ecosystem에 차이가 있습니다.
The US has advantages in terms of talent, in terms of certain kinds of innovation, in terms of alliances with other countries.
미국은 talent, innovation, 동맹국 측면에서 장점이 있습니다.
China has advantages in terms of scale, in terms of manufacturing, in terms of government coordination.
중국은 scale, manufacturing, government coordination에서 장점이 있습니다.
So it’s not going to be one winner takes all.
One winner takes all은 아닙니다.
But I do think the US has a significant lead right now and we should work to maintain it.
하지만 미국이 현재 상당한 lead를 가지고 있고 유지해야 한다고 생각합니다.
I think the competition is good in some ways because it pushes everyone forward.
Competition은 모두를 앞으로 밀어붙이는 면에서 좋습니다.
But we also need to be careful about the risks of proliferation and arms races.
Proliferation와 arms race risk에는 주의해야 합니다
I think the competition is good in some ways because it pushes everyone forward.
경쟁은 모두를 앞으로 밀어붙이는 면에서 좋습니다.
But we also need to be careful about the risks of proliferation and arms races.
하지만 proliferation과 arms race risk에는 주의해야 합니다.
There’s a lot of discussion about whether we should have export controls on AI technology.
AI 기술 export control에 대한 논의가 많습니다.
I think some level of control is necessary to prevent the most dangerous capabilities from spreading too quickly.
가장 dangerous capability가 너무 빨리 퍼지지 않게 어느 정도 control은 필요하다고 생각합니다.
But we have to be careful not to stifle innovation or international collaboration on beneficial uses.
Beneficial use에서의 innovation이나 국제 협력을 stifle하지 않도록 조심해야 합니다.
Overall, I remain optimistic that we can navigate these challenges.
전반적으로 이러한 challenge를 잘 헤쳐나갈 수 있을 것이라고 낙관합니다.
The benefits of this technology are so enormous that it’s worth the effort to get it right.
이 기술의 benefit이 워낙 거대해서 제대로 하는 노력이 가치 있습니다.
That’s why I’m sending this message of urgency.
그래서 urgency 메시지를 보내는 것입니다.
We are near the end of the exponential, and the decisions we make in the next few years will shape the future for a very long time.
우리는 exponential의 끝에 가까워졌고, 앞으로 몇 년 동안의 결정이 오랫동안 미래를 결정할 것입니다.
I want to make sure that we’re all paying attention and acting with the seriousness that this moment deserves.
모두가 주의를 기울이고 이 순간의 seriousness에 맞게 행동하기를 바랍니다.
The conversation continues with deeper dives into specific topics:
I actually think that the scaling hypothesis still holds very strongly.
scaling hypothesis는 여전히 매우 강력하게 유지된다고 생각합니다.
We’re seeing gains from both pre-training and post-training with RL.
pre-training과 RL post-training 모두에서 gain을 보고 있습니다.
The models are getting better at reasoning, at coding, at general capabilities.
모델은 reasoning, coding, general capabilities에서 더 좋아지고 있습니다.
The progress is steady and it’s compounding.
진전은 steady하고 compounding됩니다.
One of the things that’s surprising is how quickly the models are absorbing new skills once we give them the right training signals.
놀라운 점은 right training signal을 주면 모델이 새로운 skill을 얼마나 빨리 absorb하는지입니다.
We’re not hitting walls in the way some people predicted.
일부 사람들이 예측한 wall을 만나지 않고 있습니다.
Instead, we keep finding that more compute and better data lead to better performance across the board.
대신 more compute와 better data가 전반적으로 better performance로 이어집니다.
This is why I keep coming back to the big blob of compute idea.
그래서 big blob of compute 아이디어로 계속 돌아갑니다.
It’s not that algorithms don’t matter, but the dominant driver is scale.
algorithm이 중요하지 않다는 게 아니라 dominant driver는 scale입니다.
When you combine scale with good engineering, you get these leaps.
scale과 good engineering을 결합하면 leap가 생깁니다.
We’re seeing that in real time.
실시간으로 보고 있습니다.
Regarding the economic impact, I think AI will diffuse through the economy faster than people expect.
경제적 impact에 대해, AI는 사람들이 예상하는 것보다 빠르게 경제에 diffusion될 것입니다.
There will be bottlenecks, but the pressure from the capability improvements will be immense.
bottleneck이 있겠지만 capability improvement의 pressure는 엄청날 것입니다.
Companies that don’t adopt it will fall behind very quickly.
채택하지 않는 회사는 매우 빠르게 뒤처질 것입니다.
This is going to be disruptive, but in a net positive way if we manage it well.
disruptive하지만 잘 관리하면 net positive입니다.
On the question of safety and alignment, we’re working hard on that at Anthropic.
safety와 alignment 문제에 대해 Anthropic에서 열심히 작업 중입니다.
We believe in constitutional AI and other techniques to make models more helpful, honest, and harmless.
constitutional AI 등으로 모델을 helpful, honest, harmless하게 만들려고 합니다.
But we also recognize that as capabilities increase, the stakes get higher.
capability가 증가할수록 stakes가 높아진다는 것도 인지합니다.
That’s part of why I’m emphasizing the urgency.
urgency를 강조하는 이유 중 하나입니다.
We need to get the governance and safety right alongside the technical progress.
technical progress와 함께 governance와 safety를 제대로 해야 합니다.
I don’t think pausing is the answer, but thoughtful acceleration with safeguards is.
pausing은 답이 아니지만, safeguard와 함께 thoughtful acceleration이 답입니다.
Looking ahead, I’m excited about what’s coming.
앞으로 올 것에 기대됩니다.
A country of geniuses in a data center could solve many of humanity’s biggest problems.
data center 속 천재 국가가 인류 최대 문제들을 해결할 수 있습니다.
From biology to physics to climate, the potential is enormous.
biology, physics, climate 등 잠재력이 엄청납니다.
But we have to navigate the risks carefully.
risk는 신중히 navigate해야 합니다.
I believe we can do both.
둘 다 할 수 있다고 믿습니다.
That’s the message I want to leave people with.
사람들에게 남기고 싶은 메시지입니다.
Pay attention, act with urgency, but also with wisdom.
주의를 기울이고 urgency 있게 행동하되 wisdom도 가져야 합니다.
The future is bright if we make the right choices.
right choice를 하면 미래는 밝습니다.
We also talked about how AI will affect different industries.
AI가 다양한 산업에 미칠 영향에 대해서도 이야기했습니다.
In software engineering, the changes are already visible and will accelerate.
software engineering에서는 변화가 이미 보이고 가속될 것입니다.
Models are writing more code, debugging, and even designing systems.
모델이 더 많은 코드 작성, debugging, system design을 합니다.
The role of human engineers will shift towards higher-level architecture, creativity, and oversight.
인간 엔지니어 역할은 higher-level architecture, creativity, oversight로 이동합니다.
In science, particularly biology and medicine, AI could dramatically speed up discovery.
science, 특히 biology와 medicine에서는 AI가 discovery를 dramatically 가속할 수 있습니다.
We could see new treatments, better understanding of diseases, and personalized medicine at scale.
새로운 치료법, disease 이해 향상, 대규모 personalized medicine을 볼 수 있습니다.
Climate modeling and materials science are other areas with huge potential.
climate modeling과 materials science도 huge potential이 있습니다.
Of course, there are risks like misuse for bioweapons or cyber attacks.
물론 bioweapons이나 cyber attacks 같은 misuse risk가 있습니다.
That’s why responsible development and international cooperation are crucial.
그래서 responsible development과 international cooperation이 중요합니다.
Anthropic’s approach is to prioritize safety research alongside capability development.
Anthropic의 접근은 capability development와 함께 safety research를 우선합니다.
We want to ensure that as models get more powerful, they remain aligned with human values.
모델이 더 powerful해질수록 human values와 aligned되게 하려 합니다.
The conversation also touched on public perception and policy.
대화는 public perception과 policy에도 닿았습니다.
Many people outside the tech bubble don’t fully grasp how close we are to transformative AI.
tech bubble 밖 많은 사람들이 transformative AI가 얼마나 가까운지 fully grasp하지 못합니다.
That lack of awareness is concerning because the decisions made now will have long-lasting impacts.
그 awareness 부족은 우려스럽습니다. 지금 결정이 long-lasting impact를 미치기 때문입니다.
Policymakers need to engage seriously with these issues.
정책 입안자들은 이 issue에 seriously engage해야 합니다.
Regulation should be smart — protecting against catastrophic risks without stifling innovation.
Regulation은 catastrophic risk 보호하면서 innovation을 stifling하지 않게 smart해야 합니다.
On a personal note, Dario emphasized the importance of urgency without panic.
개인적으로 Dario는 urgency without panic의 중요성을 강조했습니다.
We should be motivated to act responsibly and proactively.
responsibly하고 proactively 행동하도록 motivated되어야 합니다.
The next few years will be defining for humanity.
앞으로 몇 년은 인류에게 defining할 것입니다.
I encourage everyone to stay informed and contribute positively.
모두가 informed하게 있고 positive하게 contribute하기를 권합니다.
One more thing on the economic side: the diffusion will not be uniform.
경제 측면에서 한 가지 더: diffusion은 uniform하지 않을 것입니다.
Some sectors will adopt AI very rapidly, others more slowly due to regulatory, infrastructural, or cultural reasons.
일부 sector는 매우 빠르게 adopt하고, 다른 곳은 regulatory, infrastructural, cultural reason으로 천천히.
But the overall effect will still be massive productivity gains across the board.
하지만 전체 effect는 여전히 massive productivity gains입니다.
Frontier labs like Anthropic will play a key role not just in building models but in shaping the ecosystem responsibly.
Anthropic 같은 frontier lab은 모델 구축뿐 아니라 ecosystem을 responsibly shaping하는 데 key role을 합니다.
We’re investing in interpretability, alignment techniques, and red-teaming to stay ahead of risks.
interpretability, alignment technique, red-teaming에 투자해 risk를 앞서갑니다.
To the audience: don’t get lost in day-to-day noise.
청중에게: day-to-day noise에 빠지지 마세요.
The big picture is that we’re approaching a historic inflection point.
big picture는 historic inflection point에 접근하고 있다는 것입니다.
Stay engaged, support good policies, and think about how you can contribute.
engaged하게 있고, good policy를 support하며, 어떻게 contribute할지 생각하세요.
Whether through research, business, or public discourse — every bit helps.
research, business, public discourse를 통해 — 모든 것이 도움이 됩니다.
Thanks for listening, and I look forward to continuing the conversation as things evolve.
들어주셔서 감사하고, 상황이 evolve함에 따라 대화를 이어가길 기대합니다.
I actually do think that the API model is more durable than many people think.
API 모델은 많은 사람들이 생각하는 것보다 더 durable하다고 생각합니다.
One way I think about it is if the technology is advancing quickly, if it’s advancing exponentially, what that means is there’s always a surface area of new use cases that have been developed in the last three months.
기술이 빠르게 exponentially 발전하면 지난 3개월 동안 새로운 use case surface area가 항상 생깁니다.
Any kind of product surface you put in place is always at risk of sort of becoming irrelevant.
어떤 product surface를 만들더라도 relevance를 잃을 위험이 있습니다.
Any given product surface probably makes sense for a range of capabilities of the model.
특정 product surface는 모델 capability 범위에 맞을 뿐입니다.
There’s always going to be this front of new startups and new ideas that weren’t possible a few months ago and are possible because the model is advancing.
몇 달 전에는 불가능했지만 모델 발전으로 가능한 새로운 startup과 아이디어가 항상 앞서 나갑니다.
I actually predict that it’s going to exist alongside other models, but we’re always going to have the API business model because there’s always going to be a need for a thousand different people to try experimenting with the model in a different way.
API business model은 다른 모델들과 함께 존재할 것이고, 수천 명이 모델을 다양한 방식으로 실험할 필요가 항상 있기 때문입니다.
I think the way to think about this is there’s a frontier of new use cases that are enabled by the new capabilities and that frontier is always moving forward.
새 capability로 enabled되는 new use cases frontier가 항상 앞으로 움직인다고 생각해야 합니다.
And so the API is kind of the way you access that frontier.
API는 그 frontier에 접근하는 방식입니다.
And then there will be various products that are built on top of the frontier at any given time.
그리고 특정 시점에 frontier 위에 다양한 제품이 만들어집니다.
But because the frontier is moving so quickly, those products are always at risk of becoming outdated.
하지만 frontier가 너무 빠르게 움직여서 제품들은 항상 outdated될 위험이 있습니다.
So the API remains this durable thing that lets people access whatever the current frontier is.
그래서 API는 현재 frontier에 접근하게 하는 durable한 것입니다.
Now you could imagine a world in which the exponential slows down or stops, and then in that world the API becomes less important and products become more important.
Exponential이 느려지거나 멈추는 세계에서는 API 중요도가 떨어지고 제품이 더 중요해질 수 있습니다.
But in a world where the exponential continues, which is the world that I think we’re in, the API remains this very important and durable thing.
하지만 우리가 있다고 믿는 exponential이 계속되는 세계에서는 API가 매우 중요하고 durable합니다.
How will frontier labs actually make money?
Frontier lab들은 실제로 어떻게 돈을 벌까요?
I think the API business is a good business and it will continue to be a good business.
API business는 좋은 사업이고 계속 좋을 것입니다.
But I also think that there will be various other businesses that frontier labs do.
Frontier lab들은 다른 사업도 할 것입니다.
Some of them will be things like selling models to enterprises or governments or whatever.
Enterprise, 정부 등에 모델 판매 같은 것들.
Some of them might be sort of doing joint ventures with big companies where you have a big company that has a lot of data or a lot of distribution and you partner with them.
Big company와 joint venture (데이터나 distribution 보유 회사와 파트너십).
I think there will be a lot of heterogeneity in how the business models work.
Business model은 매우 heterogeneous할 것입니다.
I think the thing that people sometimes miss is that frontier labs are not the only ones who will make money from AI.
사람들이 놓치는 점은 frontier lab만 AI로 돈 버는 게 아니라는 것입니다.
There will be a whole ecosystem around them.
주변에 전체 ecosystem이 있을 것입니다.
There will be many applications built on top of the models.
모델 위에 많은 application이 만들어질 것입니다.
There will be many infrastructure companies.
많은 infrastructure 회사도 있을 것입니다.
So the pie is going to be very large and there will be many winners.
파이가 매우 크고 많은 winner가 나올 것입니다.
I think frontier labs will capture a reasonable fraction of that pie, but not all of it.
Frontier lab들은 그 파이의 상당 부분을 차지하겠지만 전부는 아닙니다.
If AGI is imminent, why not buy more compute?
AGI가 임박했다면 왜 더 많은 compute를 사지 않나요?
I think we are buying a lot of compute.
많이 사고 있습니다.
We’re one of the largest buyers of compute in the world.
세계에서 가장 큰 compute buyer 중 하나입니다.
We’re investing billions of dollars in compute.
수십억 달러를 compute에 투자합니다.
But there are constraints on how quickly you can actually deploy that compute.
하지만 실제로 deploy할 수 있는 속도에는 제약이 있습니다.
There are supply chain constraints, there are power constraints, there are data center build-out constraints.
Supply chain, power, data center build-out 제약이 있습니다.
So we’re moving as fast as we can, but it’s not instantaneous.
최대한 빠르게 움직이지만 즉각적이지는 않습니다.
We’re also not the only lab doing this.
우리만 하는 것도 아닙니다.
All the frontier labs are scaling aggressively.
모든 frontier lab이 aggressively scaling 중입니다.
I think the amount of compute that’s being brought online is growing very rapidly.
온라인으로 들어오는 compute 양이 매우 빠르게 증가하고 있습니다.
But there are real physical constraints in the world.
현실적인 물리적 제약이 있습니다.
It’s not like we can just snap our fingers and have 10x more compute tomorrow.
내일 당장 10배 compute를 snap fingers로 만들 수는 없습니다.
We’re pushing as hard as we can on all of those fronts.
모든 front에서 최대한 push하고 있습니다.
Will regulations destroy the boons of AGI?
Regulation이 AGI의 boons를 파괴할까요?
I think regulation is a real risk.
Regulation은 진짜 위험입니다.
There are ways that regulation could go wrong and slow things down or make things worse.
Regulation이 잘못되어 속도를 늦추거나 악화시킬 수 있습니다.
But I also think that there are smart ways to regulate that could mitigate risks without destroying the benefits.
하지만 risk를 완화하면서 benefit을 파괴하지 않는 smart regulation도 있습니다.
I think the key is to have regulation that is focused on the highest risk applications, like biological weapons or cyber weapons, while allowing the beneficial uses to proceed.
가장 높은 risk application (biological weapons, cyber weapons)에 집중하면서 beneficial use는 진행하게 하는 regulation이 핵심입니다.
I think we need to have a balanced approach.
Balanced approach가 필요합니다.
Why can’t China and America both have a country of geniuses in a datacenter?
중국과 미국이 둘 다 data center에 country of geniuses를 가질 수 없는 이유는?
I think they can both have very powerful AI systems.
둘 다 매우 강력한 AI system을 가질 수 있습니다.
But there are differences in the ecosystems.
하지만 ecosystem에 차이가 있습니다.
The US has advantages in terms of talent, in terms of certain kinds of innovation, in terms of alliances with other countries.
미국은 talent, innovation, 동맹국 측면에서 장점이 있습니다.
China has advantages in terms of scale, in terms of manufacturing, in terms of government coordination.
중국은 scale, manufacturing, government coordination에서 장점이 있습니다.
So it’s not going to be one winner takes all.
One winner takes all은 아닙니다.
But I do think the US has a significant lead right now and we should work to maintain it.
하지만 미국이 현재 상당한 lead를 가지고 있고 유지해야 한다고 생각합니다.
I think the competition is good in some ways because it pushes everyone forward.
Competition은 모두를 앞으로 밀어붙이는 면에서 좋습니다.
But we also need to be careful about the risks of proliferation and arms races.
Proliferation와 arms race risk에는 주의해야 합니다.
There’s a lot of discussion about whether we should have export controls on AI technology.
AI 기술 export control에 대한 논의가 많습니다.
I think some level of control is necessary to prevent the most dangerous capabilities from spreading too quickly.
가장 dangerous capability가 너무 빨리 퍼지지 않게 어느 정도 control은 필요하다고 생각합니다.
But we have to be careful not to stifle innovation or international collaboration on beneficial uses.
Beneficial use에서의 innovation이나 국제 협력을 stifle하지 않도록 조심해야 합니다.
Overall, I remain optimistic that we can navigate these challenges.
전반적으로 이러한 challenge를 잘 헤쳐나갈 수 있을 것이라고 낙관합니다.
The benefits of this technology are so enormous that it’s worth the effort to get it right.
이 기술의 benefit이 워낙 거대해서 제대로 하는 노력이 가치 있습니다.
That’s why I’m sending this message of urgency.
그래서 urgency 메시지를 보내는 것입니다.
We are near the end of the exponential, and the decisions we make in the next few years will shape the future for a very long time.
우리는 exponential의 끝에 가까워졌고, 앞으로 몇 년 동안의 결정이 오랫동안 미래를 결정할 것입니다.
I want to make sure that we’re all paying attention and acting with the seriousness that this moment deserves.
모두가 주의를 기울이고 이 순간의 seriousness에 맞게 행동하기를 바랍니다.