주요 AI 발표: Microsoft의 에이전트 및 새 모델

### 주요 AI 개발 사례 Microsoft, IBM, Anthropic 등
* 🚀 어제는 인공지능 뉴스에서 중요한 날로, 다음 달 공개 프리뷰에 들어갈 예정인 **Windows 환경에서 자율 에이전트를 만들고 실행하는 코파일럿 스튜디오**를 Microsoft가 공개한 것을 시작으로 수많은 발표가 있었습니다.
* 💼 Microsoft는 영업, 서비스, 재무 및 공급망 팀을 대상으로 **10개의 새로운 자율 에이전트**를 Dynamics 365에 도입하고 있으며, 곧 **수백만 명의 에이전트가 기업 인력에 침투**할 것으로 예상하고 있습니다.
* 🔥 이와는 대조적으로 Salesforce의 CEO 마크 베니오프는 Microsoft의 노력을 비판하며 **Co-pilot을 "실패작"**으로 규정하고 효과적인 기업 인텔리전스에 필요한 데이터와 보안이 부족하다고 주장하며 **"클리피 2.0"이라는 유머러스한 별칭**을 붙였습니다.
* 📊 IBM은 또한 기존의 미세 조정 없이 핵심 모델을 개선하는 **새로운 기술을 갖춘 Granite 3.0**을 비롯한 여러 오픈 소스 모델을 출시하여 화제를 모았습니다.
* 🤖 앤트로픽은 몇 가지 실험적인 단점이 있긴 하지만, AI 모델이 컴퓨터를 제어할 수 있는 **새로운 컴퓨터 사용 도구인 Claude 3.5**를 선보였습니다.
* 🌍 메타는 새로운 텍스트 음성 변환 언어 모델과 함께 **이미지를 쉽게 분할할 수 있는 Segment Anything 2.1**과 같은 기능을 출시하며 오픈 소스 프로젝트에 크게 기여했습니다.
* 💰 전 OpenAI CTO인 미라 무라티가 독점 AI 모델에 초점을 맞춘 새로운 벤처를 위해 **1억 달러 이상의 자금**을 조달할 것이라는 소문이 돌면서 업계에 경쟁이 더욱 치열해질 것으로 보입니다.
* 🎨 마지막으로, **Stability AI, Runway, 11 Labs와 같은 기업의 혁신**으로 크리에이티브 도구가 대중화되면서 사용자들이 고품질 미디어 콘텐츠를 쉽게 제작할 수 있게 되어 크리에이터들에게 흥미로운 시기가 되고 있습니다.

Microsoft는 Co-pilot Studio를 통해 새로운 자율 에이전트로 기업 환경을 혁신하고 있습니다.

Yesterday may have been one of the biggest days in artificial intelligence news since the beginning. There were numerous announcements, and we're going to cover everything from yesterday and the past few days. There's so much stuff; let's get right into it. First, Microsoft released agents. They've now released previews of their agent studio called Co-pilot Studio, where you can create agents, run agents, and basically put agents into every aspect of your Windows environment. And they are betting big on it. And there's a few interesting bits from the announcement. Let me read a few parts. The ability to create autonomous agents with Co-pilot Studio will be in public preview next month, so not quite released yet, but very soon. Then they're introducing 10 new autonomous agents in Dynamics 365 to build capacity for every sales, service, finance, and supply chain team. So they are building agents for the enterprise environment. And based on this blog post, based on a keynote Satya Nadella, the CEO of Microsoft, just gave, based on another post that Microsoft just released, they are betting that

agents are going to infiltrate every aspect of the corporate environment. They believe that millions of agents are going to enter the workforce in the next few years, and I kind of agree with them. It's probably going to be pretty simplistic at first, a lot of basic automation, but as the models get better, as the agentic frameworks get better around the raw intelligence, we're just going to see these agents become more and more proactive rather than just being reactive and doing the task that is assigned to them in that moment. But not everybody thinks the Microsoft announcements are awesome. In fact, Mark Benioff has been flaming Microsoft about all of their launches. And if you're not familiar, Mark Benioff is the CEO of Salesforce. He definitely has a very strong personality. And if you remember way, way, way back at the beginning of Salesforce, it started with a very famous motto: No software. Previously, all software was delivered on-prem and you would install it and it would become stagnant. And he was really one of the original people to bring software to the

cloud SAS, and he had this famous pin where it was exactly like this, but rather than saying co-pilot, it said software. But now he's using it again to say co-pilot. Look at how aggressively he is going after Microsoft. Microsoft rebranding co-pilot as agents, that's panic mode. Let's be real, co-pilot's a flop because Microsoft lacks the data, metadata, and enterprise security models to create real corporate intelligence. Now, that's not really true. Microsoft has all of the data. They are in essentially 98% of all enterprise companies right now. Every business uses Microsoft in some way, so they have the data. And to say that they don't have the security, maybe, but really, they have the talent to build the security if they don't already have it, which is up for debate. And he's calling it clippy 2.0, and I think that's hilarious. I really actually love this corporate jabbing, even though it's all puffery, and Mark Benioff might not actually believe the things he's saying.

and of course he's coming up with competing products with agent force. So, you know, as soon as co-pilot Studio comes out, I'm going to play around with it. I'll let you know how it is. Next, IBM has released multiple open-source open-weight models to the world. I was just in New York last week visiting IBM. They showed me a bunch of previews of these announcements, and now they're here. So, we have Granite 3.0 8B and 2B. We also have a mixture of experts version of the Granite model. And again, completely open source, completely open weight, Apache 2.0 license. They also have a new technique to add knowledge to these core models that isn't really Rag and isn't really fine-tuning. It's somewhere in between, and I have an entire video coming out going over the granite model, going over all of these techniques coming soon. Next, of course, the biggest news from yesterday, Anthropic released Claw 3.5 Sonet, the new version. They also

announced Claude 3.5 Hau, which is their newest small model but better than clad 3.0 Opus, which is kind of nuts to think about. Opus was their biggest model, and 3.5 Sonic new is performing extraordinarily well. If you haven't already seen my video about it, check it out. I'll drop a link in the description below. I also did a full test of cloud 3.5 Sonic new. Alex is putting the final touches on the edits, and we should release that in the next day or so. So, if you're not already subscribed, make sure you subscribe so you can see that video and see how this new awesome model does. And the most compelling announcement from yesterday from anthropic was the computer use tool, which is absolutely mind-blowing. Basically, it's anthropics' attempt to give these AI models the ability to control your computer. We've already seen open-source projects attempt to do it, like open interpreter and a few others, but it hasn't always worked well. And now anthropics giving it

a try. They have a caveat with it: it is experimental, and it makes mistakes. And they've actually shown examples of weird behavior, like when the model all of a sudden stopped doing what it was tasked to do and started researching Yellowstone out of nowhere. Kind of cool, kind of weird. So, it's an API. You can try it out already. I haven't created a video about it yet, but if you want to see a tutorial, or if you just want to see me play around with it, let me know in the comments. Thanks to the sponsor of this video, Vulture. Reap the benefits of the world's largest independent cloud provider when you bring your GPU workloads over to Vulture. They have the latest Nvidia GPUs spanning 32 locations across six continents. Vulture delivers industry-leading price to performance and serious accessibility and reliability. Vulture's global fully composable cloud infrastructure moves your applications closer to your users and frees you from vendor lock-in, allowing you to bring your own networking and database solutions. And if you need to scale beyond just a single

cluster vulture's Kubernetes engine allows you to take full control over your deployment, offering up a 100% free control plane. So, if you're tired of waiting for GPUs from other providers, make sure to use Vulture. You can deploy at any scale immediately, and they have H100s, L40, and more available right now. You can get a fraction of a card or fully dedicated bare metal systems, which gives you full control over your hardware and your throughput. They also have one-click installation of all the applications you might need for advanced machine learning workloads, allowing you to get up and running in minutes, not hours. So, experience the Vulture difference. Don't get bogged down by severe wait times or limited locations. Try it free today with a $300 credit for your first 30 days when you visit GetVulture.com/Burman and make sure to use code Burman300 at checkout to get that credit. Thanks again to Vulture for sponsoring this video. Next, last week, Meta released a ton of open-source projects. I continue to be

impressed with Meta's open-source contributions. So, here are all of the features segment anything 2.1. If you haven't already seen that it was out, you can basically take any image or video, click on it, and it automatically will segment anything in the image. So, if you want to take a person out of the image, if you want to take a balloon or a building or anything at all, it's so simple to do and highly accurate. Then, they also released Spirit LM, which is an open-source language model for text-to-speech. And then, they also released a bunch of highly technical projects aimed at making training easier, making inference faster, and so on. So, definitely check it out. I'll drop the link to this blog post in the description below. Next, former CTO of OpenAI, Mira Murati, fresh off just leaving OpenAI, is already rumored to be raising a bunch of money and starting a new company and hiring people. The talks are in the early stages, and Mira Murati's new venture could raise over 100 million given her

reputation and the capital needed to train proprietary models. And so, of course, she's building proprietary models. She might build products on top of those models. But look, if anyone has the knowledge to do it, it's definitely the former CTO of OpenAI. And she has the reputation to hire the best people, raise a ton of money. So, you know what? I'm all for it. The more competition, the better. Next, a Chinese copycat company of the Optimus robot has released a new video and a new humanoid robot that seems to be pretty good. It looks a little bit awkward at times, and it's definitely not as smooth as the Optimus robot, but cool nonetheless. And again, the more competition, the better. Now, a lot of people are saying this is almost an identical copy of the Optimus robot. It definitely looks pretty similar, but I don't know if it's a copy necessarily. Plus, the Optimus robot hasn't been around that long. This company hasn't been around that long. So, the idea that they maybe stole secrets

and then reproduced it in such a short period of time seems unlikely to me. But you know what? Again, I love humanoid robots. I love all the progress, and good for this company NeXT. Stability AI released stable diffusion 3.5, the new version of their text-to-image generation model. They have stable diffusion 3.5 large and stable diffusion 3.5 large Turbo, and they also said they're going to release a medium-sized version on October 29th. Now, the great thing about stable diffusion is it's open source. You can download it, you can run it on your computer, and anytime a company contributes to the open-source community, I am very thankful to them. So here are some examples. These people look incredible. They look real. I wouldn't be able to tell the difference, and the light is great, the hair is fantastic, everything looks right. Now, obviously, there are no hands in these photos, so that's really the seemingly hardest thing for AI to get right. Here are some other examples. This is obviously a piece of art. This looks like a pretty

realistic mug. Although the handle of the mug looks a little awkwardly positioned. And then another piece of art right there. And you can access these models through the stability AI API, through replicate, through comfy UI, locally on your computer, and from Deep infra. And if that's not enough, AOG released a brand new feature called AOG canvas. It's basically like having an infinite canvas board. You start with an image, and you can keep adding to it or adding around it. And yeah, it just seems really cool. Bring your face or brand visuals to ideogram canvas and use industry-leading magic fill and extend to blend them with creative AI-generated content. So, it has a brand new interface. Here's what it looks like. You can generate images, select the images, expand on them, add to them. It's really cool. Here's what magic fill looks like. So, you can take an image, you can select part of it, and basically change it to be anything that you want. Here is extend, which is an outpainting tool. So again, you take the

image, and then you can add anything you want around it. Andre Karpathy says, Love it, eager to try, and I am as well. I can't wait to try this next LM Studio released another awesome update, 0.3.5, and now it has headless mode. So, very similar to how AMA Works, where there's basically no interface, you get the interface with LM Studio, but then you can essentially make it disappear and just use it in your taskbar, for example. And it's just an API at that point. So, if you're a developer, that is perfect for you. You know I love LM Studio, I'm an investor in LM Studio, and I really do use it all the time. Here, you can minimize to the system tray, so you don't have to keep the UI open if you don't want to, and again, that is perfect for development efforts. Next, one thing the CEO of Perplexity, Rvin, said that they are adding in the ability for their models to think, and most likely they're using the 01 model model now, and so if

you're doing deep research or anything requiring that level of intelligence, which you know how I feel about that. You can obviously use that feature now, the thinking feature. So, it is a reasoning-powered search agent; that is what they're calling it. But not only that, now that Clad 3.5 son is out, they have it available. If you're a pro user in perplexity, you can use that new model. And of course, I turned it on. I really don't see much of a difference between the models in perplexity. Perplexity usually is just really good, and I don't care about which model it's using, but it's cool to be able to try it out nonetheless. And one thing that's super fascinating to me, and I don't really understand how they do it, is how do they just add these models with no issues? The day that the model's announced, they have it available in perplexity. How do they not have any issues? How do they not have to redo the prompting or the caching or

anything else. It just seems like they're ready to go, and whatever they've built, it's a very robust system. Next, we have a brand new open-source text-to-video model. It is called Moi 1 and is by Genmo. I think it's Genmo, Genmo AI, Genmo, whatever. And so, it looks really good, and everything stays consistent, and the physics looks good. And I haven't played around with it more than just looking at the demo videos, but if you have, let me know what you think in the comments below. It's open-source, so you can technically download it and get it running on your own computer, which you know I love. And if you want to try it, you can try it right now for free. And of course, here's the model available on GitHub if you want to download it and play around with it. And another one, Runway, has come out with a brand new feature. Now, Runway is not open-source, but it's still cool nonetheless. They just introduced Act One, which is a new

way to generate expressive character performance inside Gen3 Alpha using a single driving video and character image. Basically, you take your face or a video, anything about yourself, and then you can use text-to-video to overlay on top of it, and you're basically animating the character very similar to how Gollum was portrayed by a real human. But they needed a ton of technology with all of this face tracking and all of this really expensive technology. But now, you just need a video of yourself, and that's it. So, really, the cost to create incredible videos is coming down drastically. So, very cool. Congrats to Runway on this launch. Next, a quick story. XAI finally has an API available, and it's pretty competitive pricing. If you like the Grok models, definitely try it out. We know we're getting another Grok model very soon, and it has a single model, the Grok Beta, priced at $5 per million input tokens and $15 per million output tokens. So, if you're a developer and you need another model in your

quiver of models for your AI application. Definitely check this out next, and I love all this generative AI coming out right now. 11 Labs just released the ability to describe a voice, and then you can use the voice with a script. So, rather than having to clone an existing audio file or voice, you basically just describe what you want, and then you can use it. So, here's an example: a large yeti with a deep, rumbling voice. Let's listen. I have guarded these sacred peaks since before your ancestors first dreamed of climbing them. Here's a scary witch with a hushed voice. The moon whispers secrets tonight. The stars have told me of your fate. So, again, really awesome. The amount of creative tools that people are getting today is going to allow anybody to create incredible movies, TV shows, podcasts. It's just such an exciting time to be a creative person. So, that's it for today. Those are all the stories. It's only Wednesday as I'm recording this, so who knows what's going to happen the rest of the week. Maybe I'll create

another news video this week because this week has been absolutely bonkers. So, if you enjoyed this video, please consider giving a like and subscribe, and I'll see you in the next one.

AI News: Copilot Agent Builder, IBM Granite 3.0, NEW Claude Sonnet, Open-Source Text To Video!

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

AI 에이전트를 위한 마이크로소프트의 대대적인 추진

Microsoft는 Windows 환경의 모든 측면을 위한 자율 에이전트를 만들고 실행하는 것을 목표로 하는 코파일럿 스튜디오의 미리 보기를 출시했습니다.

이 회사는 영업, 서비스, 재무 및 공급망 팀을 지원하기 위해 Dynamics 365에 10개의 새로운 에이전트를 도입하고 있으며, 다음 달에 에이전트 생성의 공개 프리뷰가 예정되어 있습니다.

그러나 Salesforce의 CEO인 마크 베니오프는 코파일럿을 에이전트로 리브랜딩하는 것은 공황 상태의 신호이며 진정한 기업 인텔리전스를 제공할 수 있는 능력에 의문을 제기하며 Microsoft의 노력을 비판했습니다.

최근 AI 모델 릴리스 및 업데이트

Co-pilot Studio가 곧 출시될 예정이며, IBM은 Granite 3.

0과 전문가용 버전을 포함한 여러 오픈 소스 모델을 출시했습니다.

5 Sonet과 Hau 모델이 발표되었으며, 특히 Sonet의 성능이 매우 뛰어납니다.

Anthropic의 가장 주목할 만한 업데이트는 실험적인 컴퓨터 사용 도구로, AI 모델이 사용자의 컴퓨터를 제어할 수 있는 기능을 제공합니다.

Vulture의 Kubernetes 엔진과 Meta의 오픈 소스 기여도

Vulture의 Kubernetes 엔진은 100% 무료 컨트롤 플레인과 모든 규모의 즉각적인 배포를 제공하며, H100 및 L40을 사용할 수 있습니다.

고급 머신 러닝 애플리케이션을 위한 원클릭 설치와 함께 하드웨어 및 처리량에 대한 완전한 제어 기능을 제공합니다.

Meta는 최근 이미지 및 동영상 분할을 위한 Segment Anything 2.

1, 텍스트 음성 변환을 위한 Spirit LM 등 여러 오픈 소스 프로젝트를 출시했습니다.

Stability AI는 다양한 모델 크기와 오픈 소스 접근 방식을 갖춘 Stability Diffusion 3.

AOG는 이미지 생성 및 확장을 위한 무한 캔버스 보드인 AOG 캔버스를 도입했습니다.

5 업데이트에는 개발자를 위한 헤드리스 모드가 포함되어 인터페이스를 시스템 트레이로 최소화할 수 있습니다.

5 출시와 함께 추론 기반 검색 에이전트를 도입하여 프로 사용자들이 새로운 모델을 사용해 볼 수 있도록 했습니다.

Genmo는 일관된 물리를 제공하는 오픈 소스 텍스트-투-비디오 모델인 Moi 1을 출시했으며, 무료로 다운로드할 수 있습니다.

런웨이의 새로운 기능인 1막을 사용하면 하나의 주행 영상과 캐릭터 이미지로 표현력 있는 캐릭터 연기를 생성할 수 있어 인상적인 영상을 제작하는 데 드는 비용을 크게 절감할 수 있습니다.

제너레이티브 AI의 흥미로운 발전

11 랩스에서는 기존 오디오를 복제할 필요 없이 음성을 묘사하고 스크립트와 함께 사용할 수 있는 새로운 기능을 출시했습니다.

이 도구는 깊은 목소리의 설인이나 조용한 마녀와 같은 창의적인 음성 옵션을 제공하여 사용자가 독특한 캐릭터를 만들 수 있도록 지원합니다.

제너레이티브 AI 도구의 급증으로 크리에이터들은 놀라운 영화, TV 프로그램, 팟캐스트를 제작할 수 있게 되어 예술적 표현을 위한 흥미로운 시대가 열리고 있습니다.