에이전트 AI의 부상: 미래 컴퓨팅 패러다임

### 에이전트의 미래: AI와 컴퓨팅의 혁신
* 🤖 엔비디아의 CEO인 젠슨은 **"수천, 수백만 명의 에이전트"**가 매일 엔비디아를 위해 일하게 될 미래를 상상하며, 이는 도구 산업에서 기술 산업으로 전환하는 계기가 될 것이라고 말합니다.
* 💼 그는 **에이전트가 도구를 활용할 뿐만 아니라 서로 협력**하여 복잡한 문제를 해결함으로써 에이전트 환경에 중요한 기회를 창출하게 될 것이라고 강조합니다.
* 📈 대규모 언어 모델과 에이전트 프레임워크의 발전으로 이러한 에이전트의 성능이 기하급수적으로 향상되어 **단기 기억과 장기 기억을 모두 보유**할 수 있게 됩니다.
* 🔍 젠슨은 데이터 라벨링에서 인간의 한계에서 벗어나 모델이 자율적으로 학습하고 진화할 수 있도록 하는 **비지도 학습의 혁신**에 주목합니다.
* 💻 이 대화는 무어의 법칙을 뛰어넘는 컴퓨팅 기술의 빠른 진화를 강조하며, **AI와 GPU가 병렬 컴퓨팅을 촉진**하여 개발을 크게 가속화하는 것을 강조합니다.
* 🛠 그는 마이크로소프트의 코파일럿과 같은 기존 AI 제품을 **실패한 클리피와 비교**하며 비판하는 한편, **누구나 AI 에이전트를 구축할 수 있는 보다 사용자 친화적인 솔루션**이 필요하다고 주장합니다.
* 📚 이 토론에서는 AI 에이전트가 처음부터 효율적으로 업무를 수행할 수 있도록 하기 위해 **사람의 훈련에 비유하여 효과적으로 온보딩하는 것의 중요성**도 다룹니다.
* 🌐 궁극적으로 젠슨은 컴퓨팅의 패러다임이 **사용자의 요구를 예측하는 동적으로 생성되는 소프트웨어**로 이동하고 있으며, 이는 우리가 기술과 상호작용하는 방식에 혁명적인 변화를 가져올 것이라는 기대감을 표명합니다.

젠슨은 강력한 AI 에이전트가 협업하고 혁신하며 고급 도구와 기술을 통해 산업을 변화시키는 미래를 예측합니다.

Jensen, the CEO of Nvidia, says the future is agentic. During an interview with Marc Benioff, the CEO of Salesforce, at the Dreamforce conference, he talks about how we are going to have thousands and even millions of agents working for us every day, all day. It is an agentic future, and we're going to break down the video right now. Jensen, come on out to have you. Our industry is, you know, call it about a trillion dollars, and it's an industry of tools, you know, computers, um, software, their industry of tools. For the very first time, this is going to be an industry of skills, and you capture that phrase, and you call it agents, but it's going to be for the very first time agents sitting on top of tools, agents using tools, and the opportunity for agents is gigantic. As we know, you know, we now know, we now know, I'm hoping so, by the way, so obviously I agree. If you've watched this channel at all, you know the phrase I always say, bullish on agents. I

really believe the best incarnation of AI in the future is going to be agents and agentic frameworks and Mark Benny off the reason why he's saying I really hope so is because of agent force which is again their new product that they're launching they really think this is going to be transformational to Salesforce which has been around forever at least forever in Internet terms and they're betting big on it and he actually gives a little dig towards Microsoft which we'll hear in a moment it sounds it sounds insane but here's the amazing thing we're going to have agents that obviously uh understand understand the subtleties of the things that we ask it to do but uh it can also uh use tools and it can reason and uh it can reason with each other and collaborate with each other and you know going to we're going to give we're going to give a problem to uh agent force and uh uh agents are going to go find other agents that that can help achieve this mission they'll work together assemble together

work together and solve this problem. And so, that is absolutely the agentic future that I envision. There are agents; agents can spawn other agents; they can collaborate with other agents. There's a set of tools, and I'm not talking about three, four, five tools—thousands, hundreds of thousands of tools—that these agents can pick and choose and also create tools themselves and actually test the tools, save the tools, use them later. These are all things that we're just starting to scratch the surface on right now. And as the core models get better, the large language models get better, as the agentic frameworks around those large language models get better, as the ability for those agents to build their own tools, to choose tools off the shelf, gets better, all of these agents become exponentially more powerful. And remember, the definition of an agent is a large language model with the ability to use tools, to collaborate with other agents, and that has memory—short-term and long-term memory. That's my favorite definition of an agent. All right, let's keep watching. I think the... the... the

Breakthrough, uh, for me was a moment when, um, we realized that unsupervised learning was going to be possible because the ability for humans to, uh, be able to label data at scale, uh, we would become the limit limiters of these digital AI to, uh, expand their capabilities. And when unsupervised learning came along, which allowed us to use language models to create language models, which codify human prior knowledge, using that to now learn multimodality, multimodal, uh, uh, data, uh, from that point forward, the scale was going to be exponential. All right, so he just said a lot in very few words, so let me break that down for you. First, he's talking about the fact that humans are really the limiting factor for artificial intelligence. The amount of data we create is massive, obviously, but the amount of data that is ingested by artificial intelligence is much more, and for the most part, all public data has been used already. It's all been used by these Frontier models and even open source and non-frontier models. All of

the open data is gone. It's been used, so now there's proprietary data, data behind pay walls or offals, stuff like Reddit and Twitter and Facebook, all of this data that these companies will use themselves to make their models distinct and better than their competition. So then he's talking about reinforcement learning without the need for humans because another limiting factor of humans is the ability to label data. If we have to label all the data that trains the AI, that is an extreme bandwidth limitation. And so we have projects like AlphaGo, where AlphaGo, the AI behind it, was not trained by humans. It was not trained with a bunch of Go games where humans labeled them good move, bad move. They simply spawned a bunch of different versions of the AI, played the game, see which one did best, and then iterated on that millions or billions of times. And then eventually, they surpassed the best human Go players in the entire world. And again, that's with basically no human input whatsoever. Then we have the notion of how are we

going to scale the data. So, I already talked about all of the public data being used. There's really two paths. One, we can either do a lot more with the data we have, and that's things like being able to scale test time compute, which means during inference time we do things like Chain of Thought. And now we have this entire new dimension of scalability on AI, which is super fascinating. So, previously we were just scaling the number of parameters, and that really scaled pretty well so far. But now we have two distinct dimensions that we can scale on: the number of parameters in kind of the training time compute, and then the post-training test time compute when it's actually running through Chain of Thought. And then the other part is synthetic data. So, what if we use AI to create really great data for other AI? And OpenAI has been rumored to be doing that through the QAR project, which eventually became 01. Maybe 01 is creating data for other models. All of this is fascinating, but the more we can

remove humans from being a limiting factor in AI. The more quickly AI is going to explode into the intelligence explosion. You know, for everybody here, this is an extraordinary time because in no time in history has computer technology not only moved faster than Moore's law. I mean, Moore's law, for example, over the course of a decade would be about 100x. We are probably advancing at somewhere near 100,000. That was our other neighbor in Hawaii, Gordon Moore. Oh, is that right? Great person, great leader, great person. And so we're at a stage now so quickly. Moore's law basically says that every 18 months, the number of transistors that can be fit onto a CPU essentially doubles. So your ability to compute essentially doubles every 18 months. But a while ago, we reached a physical limitation, and we really cannot shrink these transistors anymore, and they just don't work after that. And so we really did hit this

limitation, and then AI came around, and GPUs came around, and all of a sudden, rather than serialized compute, which is how CPUs work, we have parallel compute, and that's how GPUs work. It does a bunch of calculations simultaneously, and because of the way AI works, plus parallel compute, all of a sudden, Moore's law has actually been exceeded lately, and every 6 months, we are doubling the compute power of chips. Now, we're in an era now where we're moving way faster than Moore's law, and arguably, easily Moore's law squared. And the reason for that, of course, is at every single layer, computers went from CPUs to GPUs, from human-engineered software to machine learning software, and now this feedback loop that allows us to create new AIs, and these new AIs are helping us create new computer systems. So, he said something really interesting, but he said it so quickly, it almost went unnoticeable. He said that AI is creating more and more of our

software, and I've talked a lot about that on this channel. Again, removing the human limitation from the system that we're trying to scale up will allow us to scale up the system much more quickly. Just two years ago, humans were writing all software, so we were really limited in the amount of software we can write because it's a very manual and difficult thing for humans to do. But as large language models got better at writing code, they started writing more and more code, and then we have all of these projects like cursor and repet building infrastructure around large language models to make it even easier for AI to write code. And I believe in the short run and the mid-run, we are going to have an explosion in the number of people who write code because it won't take writing actual code. All you have to do is speak or type natural language, and the AI will write the code for you. Then, at a certain point, AI will just write code as needed, and then the end vision is really just model weights being shipped by the AI.

itself. So, rather than it writing code, it's just writing model weights, and at that point, we're probably not going to be able to read the code at all. I've been thinking about this idea that in the future, we probably won't be able to read the code that AI writes because the only reason code looks the way it does today is because humans are really bad at reading and writing code, so it has to look as much like natural language as possible. But maybe in the future, it doesn't have to, and probably in the future, it doesn't have to, especially if the compute platform of the future is simply model weights. Because we already cannot read model weights, we don't know what's in the black box of AI, at least not yet. Maybe we will in the future, but it's not looking likely. The challenges that we have in front of us are many and worthy to be tackled right away. Of course, some of the most inspiring advances has to do with fine-tuning supervised guard.

and Guard railing and and um uh and uh um all of the work that's surrounds safety um everything from using AI to curate the data to create a safe curriculum to teach the AI which then fine-tunes the AI on particular skills grounded on values then then all of this wonderful AI technology to guardrail it um and Guard railing and then reflection uh using using Chain of Thought to reflect on the quality of the answers that it's producing you know it's no longer AI this one shot producing information it's now reasoning about is this is the the the uh the answer I'm generating uh sufficiently safe and proper and is it the best possible answer I could be I could I am so excited about this particular concept that we're able to now scale Artificial Intelligence on two different dimensions and I've talked about it already in this video so just briefly I want to reemphasize how important it is that we now can scale compute and scale AI intelligence at the training time and the amount of data we feed that model.

and then also at test time, which is even more scalable in my mind, because the more time you give it, the more compute you give it, the more tokens it can churn through to generate the best possible output. And that's what he's talking about with reflection and Chain of Thought. That's exactly what OpenAI just dropped about a week ago. That is what it is doing and why it has really just jumped ahead of all the other models in terms of logic and reasoning. Could be providing its reasoning with itself, and it's reflecting on the answers before it answers. And so, I think the progress that we're making in safety is phenomenal. Isn't it so important that people can actually understand? This is what I call they need to get their hands in the soil. They need to understand the basic ideas to make this possible. And that's what I hope happens with the thousands of people here, that they can get in the code while they're here to really

understand. So, when they're going back to their offices, you know that exactly what you're saying, that we've demystified the AI, that you know they don't have to worry about how they're going to DIY their AI, that they can somehow put all the hard work that your team has done right into some very practical use because this leap of faith that you've had, this strategic motion of your company, has enabled so much of this. Now, how do we put it in the hands of everybody to just get to work and building an agent should not be some computer science fair project. It should be something that we can easily do because so when he's saying it's some computer science fair project, he is specifically referencing a lot of companies that have shipped products that are so difficult to really use in production environments that actually drive a lot of value. He specifically called out Microsoft with Copilot. He said it's basically their new version of Clippy, and Clippy was a complete failure of a product, and now he is equating

clippy to co-pilot, and that is a huge dick at Microsoft that I don't really agree with. This is much different, even though the first wave of co-pilot might not have been that valuable. I think a lot of what they're doing with integrating co-pilot into all of their office products and Outlook, and I think that's great. It's going to be a lot more like onboarding employees than writing software. All right, I just want to talk about that for a second. So, he talked about onboarding employees. Just last week, I had a conversation with Crew AI CEO Joe, and we talked about agents being onboarded as employees, and that is a critically important piece of implementing AI at your company, specifically. We drew an analogy to human onboarding. So, when you hire somebody, if you don't give them any context, any training, any onboarding materials, then they're going to be starting from scratch, and the onboarding period or the ramp-up period is going to be really lengthy. And the way to shorten that is

with training, with context, with onboarding materials, and it's the same thing with AI agents. If they simply start from nothing, then they have to learn everything as they go, and every prompt you give them has to be incredibly detailed. But when you work with a human, and the more you work with them, the more history you build with a human employee or colleague, the better shorthand you create with them. You start having to say less and describe less about exactly what you want, and you can kind of just directionally say what you want, and the human, your colleague, will understand. And that's going to be the same thing with AI. I think a big part of AI agents in the future is going to be defining how to onboard them quickly, giving them all of the memory necessary, giving them access to documentation, telling them exactly how you want them to respond and what tasks you want them to do, so that when they get started, they're ready to hit the ground running. The ambition of wanting to reinvent computing to

create what we now know is such a great endeavor. Um, so somewhere between the being inspired by all the stimulus, um, the incredible challenge ahead of us, you know, it's kept us propelled, um, inspired, and fired up for a long, long time. And now here we are. We now have the instruments, the tools, uh, this capability called artificial intelligence that goes to solve all of those other problems that we've, you know, been excited about ever since we were kids. And so that's pretty exciting times. So here we are at the cusp of a complete paradigm shift in computing, the base level computing, from computing pre-written software and hardcoded everything to just-in-time dynamically created predicted software, and maybe not even software, maybe just user interfaces, and just delivering what you need in the moment. And it's all predicted. It is a completely new way to think about computing, and it requires new chips, GPUs, and it requires new software model weights. And I'm so excited to be here and talk about it with you all. It is

such a cool time to be alive. I hope you enjoyed this video. Please consider giving a like and subscribe, and I'll see you in the next one.

NVIDIA CEO on Agents Being the Future of 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

엔비디아의 CEO인 젠슨 황은 수천 명의 에이전트가 툴을 사용하고 협업하는 미래를 상상합니다.

툴의 산업에서 기술의 산업으로 전환하는 과정에서 에이전트가 중요한 역할을 하는 혁신적인 변화가 일어나고 있습니다.

대규모 언어 모델과 도구를 구축하고 선택할 수 있는 기능을 기반으로 하는 에이전트 프레임워크는 이러한 에이전트를 기하급수적으로 더 강력하게 만들 것으로 예상됩니다.

비지도 학습이 AI에 미치는 영향

비지도 학습은 인간의 지식을 체계화하고 멀티모달 데이터로부터 학습하는 언어 모델을 만들 수 있게 해줌으로써 AI의 게임 체인저로 떠올랐습니다.

지금까지는 사람이 AI의 데이터 수집을 제한하는 요소였지만, 이제는 독점 데이터와 강화 학습이 모델 차별화와 개선을 추진하는 데 사용되고 있습니다.

AI 데이터의 확장은 추론 과정에서 컴퓨팅을 늘리고 AI가 생성한 합성 데이터를 다른 모델에 사용하는 방식으로 접근하고 있습니다.

AI는 6개월마다 칩 연산 능력이 두 배로 증가하는 무어의 법칙을 뛰어넘는 전례 없는 속도로 발전하고 있습니다.

CPU에서 GPU로의 전환과 머신러닝 소프트웨어의 부상은 피드백 루프를 만들어 AI가 더 많은 소프트웨어와 시스템을 생성할 수 있게 했습니다.

소프트웨어 개발에서 인간의 한계가 사라지면서 코드 작성자의 수가 폭발적으로 증가하여 결국에는 AI가 전체 코딩 프로세스를 대신하게 될 것입니다.

AI는 결국 코드 대신 모델 가중치를 작성하여 사람이 읽고 이해하기 어렵게 만들 수 있습니다.

AI 모델 가중치의 블랙박스 특성은 투명성과 이해도에 문제를 제기합니다.

AI 지능과 컴퓨팅 파워의 확장과 함께 리플렉션 및 사고의 사슬과 같은 기술이 AI 안전과 추론의 발전을 이끌고 있습니다.

연사는 AI를 이해하기 쉽게 설명하고 누구나 쉽게 사용할 수 있도록 하는 것이 중요하다고 강조합니다.

또한 Microsoft와 같은 기업이 프로덕션 환경에서 사용하기 어려운 제품을 출시하는 것을 비판하며, 자사의 코파일럿을 실패한 클리피와 동일시합니다.

연사는 AI 에이전트를 인간 직원과 유사하게 온보딩하여 원활한 전환과 효과적인 성과를 보장하기 위해 컨텍스트, 교육 및 문서를 제공하는 것이 중요하다고 강조합니다.

인공지능은 이제 오랜 과제를 해결할 수 있는 도구와 기능을 제공하며 컴퓨팅의 패러다임 전환을 이끌고 있습니다.

이 새로운 접근 방식은 적시에 동적으로 생성되고 예측되는 소프트웨어에 중점을 두며, 잠재적으로 사용자 인터페이스가 정보를 전달하는 방식에 변화를 가져올 수 있습니다.

이러한 변화에는 새로운 칩, GPU, 소프트웨어 모델 가중치가 필요하며, 이는 컴퓨팅 환경에 혁명적인 변화를 의미합니다.