AI 혁신: 최신 개발 및 미래 트렌드

### AI 개발에 대한 인사이트: 암호화된 트윗에서 새로운 모델까지
* 🔍 샘 알트먼의 최근 트윗은 **오리온이 겨울철 별자리로 유명한 만큼 올해 말 이전에 출시될 수 있음**을 암시하며 OpenAI의 오리온 모델이 곧 출시될 것임을 암시했습니다.
* 💻 Microsoft는 Co-Pilot 제품을 개선하여 사용자가 웹 검색 및 콘텐츠 제작에 AI를 활용하면서 문서에서 공동 작업할 수 있는 **Co-Pilot Pages**와 같은 기능을 도입하고 있습니다.
* 📊 Excel의 코파일럿은 **Python 코드를 생성**하고 자연어 프롬프트에 따라 **차트를 만들 수 있으며**, Microsoft Office 도구에 AI를 심층적으로 통합한 것을 보여줍니다.
* 📈 PowerPoint에서 Co-Pilot은 **사용자 프롬프트에 따라 프레젠테이션 데크를 자동으로 생성**하여 콘텐츠 제작 프로세스를 간소화할 수 있습니다.
* 📧 Outlook의 Co-Pilot은 **유용한 후속 조치 팁과 요약**을 제공하여 이메일 생산성을 향상시키며, **Co-Pilot Studio**를 사용하면 OpenAI의 제품과 유사한 사용자 지정 에이전트를 만들 수 있습니다.
* 🛡 메타의 AI는 **2007년부터 공개 Facebook 및 Instagram 데이터로 학습**한 것으로 알려지면서 사용자 개인정보 및 데이터 사용에 대한 우려가 제기되고 있습니다.
* 🎨 Adobe는 안전한 상업적 사용과 고품질 출력을 목표로 사용자가 프롬프트에 따라 동영상을 생성할 수 있는 **Firefly 텍스트-투-비디오 모델**을 발표했습니다.
* 🌍 F. Lee의 새로운 회사인 World Labs는 공간적 맥락에서 세상을 이해하고 상호 작용할 수 있는 모델을 만드는 것을 목표로 **3D 지각 능력을 갖춘 AI 개발**에 주력하고 있습니다.

샘 알트먼의 트윗은 연말에 출시될 것으로 예상되는 OpenAI의 오리온 모델을 암시합니다.

Our first story today is another cryptic tweet from Sam Altman. Now, last time he made a cryptic tweet, it was about two months out from the strawberry release. This video is brought to you by Vulture, the easiest way to power your generative AI startup with the latest Nvidia chips. Check out Vulture. I'll drop a link in the description below. This time he says, I love being home in the Midwest. The night sky is so beautiful. Excited for the winter constellations to rise soon. They are so great. So, from this, there's already so much information to decrypt. So, first, I love is his starting point. Now, if we look back to his previous strawberry tweet, I love all lowercase letters as well. So, following a very similar pattern as the last time that he gave hints for a new model or kind of a new tech coming from OpenAI. Now, I love being home in the Midwest. The night sky is so beautiful. Excited for the winter constellations. So, constellation is likely in reference to the new Orion.

model, which is rumored to be their next Frontier Model, and it says winter, so maybe we can expect it before the end of the year. And then he says they are so great, and somebody actually did the digging for us. Orion is indeed one of the most prominent and easily recognizable winter constellations. Orion, the model, it's definitely coming soon. Next, Microsoft is revamping their co-pilot product, and to be honest, it kind of needed a revamp. The most compelling feature, recall, was indefinitely delayed for a while, and then we just got information that might be released next month, and that was really the only feature there was. There was a dedicated co-pilot key, which you could push, and it would open up chat GPT natively in Windows, but it wasn't much more than you were getting with chat GPT through the web. So, I'm really glad to see Microsoft continuing to integrate AI into every aspect of Windows. Let's talk about all the features that are coming. So, this is co-pilot Pages. Let me show

you what it does. So, this feels very similar to perplexity. It's able to do web search. It's able to reference external web pages, so you know where they're coming from, and it gives you an answer. Plus, you can create a page from it. This isn't anything that we haven't seen before, but I think what's really unique is the interface from the AI creation. You can start collaborating with anybody else on your team, and you can continue using AI to iterate on this document. So, you could see somebody's typing right there, create a proposal outline as a table for a sales opportunity using the same structure. So, yeah, it's a really deeply integrated AI work product, and it's cool. I like it a lot. So, here is co-pilot in Excel now. So, you have a spreadsheet like normal, but now you have this co-pilot window on the right side. It writes python code for you. It automatically outputs it in the Excel document. Just really, really cool. So, you simply just like always type in natural language exactly what you want, and

Copilot will create it for you in Excel, and there it goes, lots of different charts. And again, you don't have to specify exactly what you want to see in the chart; you specify what you're trying to accomplish with the output. And here's the Python code that was written to create those charts. So again, super impressive, love to see it. Next, here's Copilot in PowerPoint. So probably going to be the same thing. You type in exactly what you want to see, and it will create the deck for you. Why commercial EV charging stations would be valuable for a retail company, including how it will help attract customers and increase foot traffic. It's doing all the research for you, puts together the deck, you can reconfigure it any way you want, you can add things. And look at that, just with a simple prompt and a click of a button, you get the entire deck created for you. And next, here's Copilot in Outlook. So it's giving you what seems to be little tips of how to follow up. It's giving you

a summary at the top again, not something super revolutionary, but cool nonetheless. But what I really want to show you is co-pilot Studio agent Builder. Let's see what the agents look like in Microsoft. So here's co-pilot Studio. This is a field service agent. This feels very, very much like custom gpts, so not full agents like we've built with other agentic Frameworks. This seems very much like a reskin of custom gpts, which is fine. I love custom gpts. So you can link to specific documents you need. You can give it web search ability here. You can actually let it see your documents internally, which is really nice. But yeah, it's basically custom gpts. And Satya Nadella commented on it, and usually I wouldn't just cover a single comment, but it's actually extremely telling and kind of a dig towards OpenAI. As AI becomes more capable and agentic models themselves become more of a commodity, and all value gets created by how you steer, ground, and fine-tune these models with your business data and workflow, and

how they compose with the UI layer of human to AI to human interaction. This is something I've talked about at length on this channel. Models are becoming a commodity. They're already a commodity, especially when Llama was released and was essentially able to match the performance of the top closed-source frontier models. And SAA, I've said it before, he's playing 4D chess. He is investing in every AI company. He is not putting all his chips into OpenAI's model. He has partnered with Meta on Llama. He's done so much. I'm just super impressed with how all-in he is on AI. So congrats to Microsoft on all their launches. Next, speaking of Meta and Llama, it turns out Mark Zuckerberg may not have been telling the truth about Llama not or never being trained on your Facebook data. The Verge published this article last week. Meta fed its AI on almost everything you've posted publicly since 2007. Unless you're in the EU, there's no ability to opt out of AI training settings that keep Facebook or

Instagram post public now. Maybe what Mark Zuckerberg really meant is he's not going to train it on private data, but if you've posted publicly, then they have trained on your data, which makes sense. Look, that is their major differentiation from OpenAI, from really any other AI company, is the fact that they have all of this unique high-quality data from their products. Meta has acknowledged that all text and photos that adult Facebook and Instagram users have publicly published since 2007 have been fed into its artificial intelligence models. Australia's ABC News reports that Meta's Global Privacy Director Melinda Clay initially rejected claims about user data from 2007 being leveraged for AI training during a local government inquiry about AI adoption before relenting after additional questions. The truth of the matter is that unless you have consciously set those posts to private since 2007, Meta has just decided that you will scrape all of the photos and all of the texts from every public post on Instagram or Facebook since 2007 unless there was a conscious decision to set them private.

that's the reality, isn't it? And she says, Correct. And it should have been clear that they were going to do this because in their privacy center, they essentially just say it: We use public posts and comments on Facebook and Instagram to train generative AI models for these features and for the open-source community. We don't use posts or comments with an audience other than public for these purposes. So, I definitely remember Mark Zuckerberg saying that they're not training on Facebook's data, but maybe he said private data and I just didn't notice. So, if you're on Facebook, your data is getting trained on if your content is public. Next, MRAW AI has released a new model, and in very MRAW form, all they did was drop a torrent link, and everybody had to figure out what it was. And it turns out it is a vision model. It's called Pixol 12B, and it is a vision model. You load it up with an image, and you can ask anything about that image. I've been meaning to play around with it; I haven't had a chance.

yet, but if you want me to do a test of it, let me know in the comments. Next, a new open-source project from a former OpenAI employee looks really cool. William says, I'm excited to announce the future of prompt engineering, E11, developed from ideas during my time at OpenAI. E11 is a light functional LM programming library, automatic versioning and tracing, rich local OSS visualization tools, multimodality native. E11 is built out of frustration for frameworks like Lang chain AI on three principles: prompts are programs, not strings; prompts are parameters of machine learning models; and every call to a language model is worth its weight in credits. Prompting should be readable and scientific. So here's an example of how it works. On the left side, what you're seeing is essentially the same thing with Lang chain and with E11. We have just a few lines of code. You specify the model, and then you define functions for the prompt that you want to use. So you define joke, you return tell me a joke about, and then you simply call the

method prompt engineering is an optimization process, and E11 automatically versions and serializes them, no custom IDE or editor required. Very cool, and they have an open-source UI that you can use, so prompt engineering goes from a dark art to a science with the right tools. E11 studio is a local open-source tool for prompt version control, monitoring, and visualization. So that's what we're seeing here, and version control for prompts is incredibly important when you're building robust and really production-ready systems using large language models, especially when you're using an agentic framework. This would plug in great into that, and it's multimodal natively, so you can see right here all the images are getting output. So if you want to check it out, I'll drop the GitHub link in the description below if it is open source, so enjoy. Next, Adobe is coming out with their own text-to-video model, and it is using the same brand name Firefly. Alexandre says, Welcome to the world Adobe Firefly video model announced today public beta later.

this year designed to be safe for commercial use, and that's really the distinguishing property of Adobe's AI Ventures, is the fact that they're only training on what they say is IP that they either own or licensed for great cinematic quality and fluid motion camera controls, and of course deep integration into our tools. So, here's an example of what it's going to look like. You can upload an image, you can have a prompt right here, and it generates it. So, here's a few examples, different prompts, and the videos look pretty good, not fantastic, but pretty darn good. And it feels very similar to the text-to-image product that they released earlier this year, except it's text-to-video. And here's one that I think is amazing. So, this one right here is real, so it's a little girl with a magnifying glass looking at a flower, and that's the reference video. Video, and then this down here is the generated clip. This is the clip based on what you may need to fill in your extra video that you may not.

have captured, so I think this is absolutely brilliant. It looks really good, and if you forget to capture a shot or you just need something extra as you're thinking about it when you're editing, now you can do it simply with a prompt. Next, the CEO of Clara is basically rewriting their entire tech stack themselves using AI, and I think this is extremely telling for the future of SaaS businesses. SaaS businesses have been overcharging for a lot of years. In a former life, I was a SaaS startup founder, so this is very familiar to me. Clara CEO said that the company is shutting down its software as a service provider Salesforce, and within a few weeks will shut down Workday, two of the biggest SaaS companies on the planet. There are large ongoing internal initiatives that are a combination of AI standardization and simplification. As an example, we just shut down Salesforce. Within a few weeks, we will shut down Workday. We are shutting down a lot of our SaaS providers as we are able to consolidate. Now I'm a little bit torn on

this on the one hand. I totally understand you have all of these new tools that allow you to create software so much more efficiently and at such a higher rate that maybe you don't actually need to pay for all these software as of services, which can be really expensive depending on how many employees you have. But on the other hand, engineers in general underestimate the amount of work it's going to take to build new systems, and I think this is going to be an example of that. At my previous company, the hard part wasn't actually the core software that we were building; it was things like integrations into other platforms, building those integrations, maintaining them, making sure that we had all the right ones. That takes a tremendous amount of work, and I don't know if AI is quite there yet, but it'll be interesting to see how AI affects the SAS industry. Next, Google Labs releases Notebook LM, which is basically a way to load up notes or documents, and then it converts it into a podcast for you, so you can actually

listen to a discussion based on whatever you've uploaded. Our new audio overview feature can turn documents, slides, charts, and more into engaging discussions with one click. So, here's an example which is based on a blog post: notebook LM goes Global with slide support and better ways to fact check. You ever get that feeling like you're just drowning in information, articles, PDFs, websites, all promising to unlock the secrets of the universe, or at least help you finally finish that research project you've been putting off? Exactly, it's like trying to drink from a fire hose, you know? It sounds really good, it's interesting, I'm trying to figure out what the actual use cases are. I can imagine loading up a research paper and listening to it, but I don't know, is it going to be as effective as just reading it? What do you think the use cases for this are? It's definitely cool tech, but I'm just not sure what the actual applications are. Next, of course, last week 01 was released by OpenAI. It made all the headlines, it made multiple

videos about it. Be sure to check those out and the Arc prize tested it. As a reminder, The Arc prize is a benchmark specific for AGI, and many of the other benchmarks in the world have really just been beaten by artificial intelligence at this point. But the arc prize is very unique in that if you're a human and you're looking at the test, it's very simple to figure out, but AI really struggles with it because Arc prizes test specifically for the ability to acquire new knowledge and then use that new knowledge in figuring out how to solve these puzzles. And 01 did pretty well, but it seems like just testing a model is not enough, and I'm going to talk about what that actually means. So here we can see an 8% score for Gemini 1.5, a 9% score for GPT 40, and here's a 133% score for 01 mini. And 01 preview and Sonic 3.5 are tied at 21%. Now, 21% doesn't sound all that great.

but just as a reference, 21% compared to GPT-40's 9% is a massive, massive improvement. Now, all these other companies have done much better. Mine's AI, for example, and they're using different techniques, not just the raw model, but they're using fine-tuning and they're using different agentic frameworks. And so now what I'm thinking is a company like Mine's AI is going to take 01, plug it into their existing framework for how they accomplish a 46%, and then get a much higher score. So again, it's not enough for just the raw model right now. Next, we have another glimpse into the future of video games. Tencent presents Game Gen O, open-world video game generation. Look at this video. So we have a paper and we have a GitHub page for it. So we introduced Game Gen O, the first diffusion transformer model tailored for the generation of open-world video games. High-quality open-domain generation by simulating a wide array of game engine features such as innovative character dynamics.

environments, complex actions, and diverse events. Now, just a couple of weeks ago, I covered Doom being created exclusively by a diffusion model, and now we have this. So, if you don't think video games are going to be vastly different in just a few years, you are going to be very surprised. Let's look at a few examples of the actual Generations themselves. Here are character generations: Geralt of Rivia. Here's one that looks very similar to Red Dead Redemption. Here's one that's a security guard, kind of looks like Snake from Metal Gear Solid. They also have environment generation. So, here's beautiful cherry blossoms. Here are some palm trees. Here's a pyramid. And it could also generate action scenes. So, here's one driving. Here's one flying, sailing. Here's a motorcycle. Looks really, really good. And here are characters walking through what looks to be any modern video game. So, here's a cyberpunk version and something that kind of looks like Destiny. So, another really cool project showing what the future video games are going to be like, and for the

last story of today. The Godmother of AI launches a new company. F. Lee has launched World Labs with $230 million of funding at a billion-dollar valuation. The company's focus will be on developing artificial intelligence with 3D perception. World Labs describes its primary product as large world models. In a blog post announcing the company's launch, it pointed out that current generative AI models can only interact with the world through text, audio, and video. This is something that Yan Laon has talked extensively about and why he doesn't think large language models are enough to actually model the world. But that is what they're hoping to do here: spatial artificial intelligence. To advance beyond the capabilities of today's models, we need spatially intelligent AI that can model the world and reason about objects, places, and interactions in 3D space and time. Now, the other company that is doing this, or at least has the data to go do this, is Tesla. They have a ton of real-world video data that can be used to train

World models, so I'm really curious what they're going to come out with soon. So that's it for today. If you enjoyed this video, please consider giving a like and subscribing, and I'll see you in the next one.

Sam Altman Teases Orion (GPT-5), NotebookLM, Pixtral, Meta Training on Facebook Data

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

샘 알트먼의 암호 트윗과 마이크로소프트의 코파일럿 개편

샘 알트먼의 최근 트윗은 연말 전에 출시될 가능성이 있는 OpenAI의 새로운 오리온 모델을 암시합니다.

Microsoft의 코파일럿 제품은 웹 검색, AI 문서 공동 작업, Excel의 Python 코드 생성과 같은 새로운 기능을 추가하는 등 대대적인 개편을 앞두고 있습니다.

또한 코파일럿은 PowerPoint 데크를 만들고 Outlook에서 후속 팁을 제공하여 Windows에 AI를 심층적으로 통합하는 모습을 보여줄 예정입니다.

마이크로소프트의 코파일럿 스튜디오와 메타의 AI 교육 논란

Microsoft의 코파일럿 스튜디오는 문서 연결 및 웹 검색 기능을 제공하는 맞춤형 GPT와 유사한 에이전트 빌더를 출시합니다.

사티아 나델라의 발언은 AI 모델이 비즈니스 데이터 및 워크플로우와의 통합을 통해 가치를 창출하는 상품이 되고 있음을 시사합니다.

한편 메타의 AI는 2007년부터 공개적으로 게시된 사용자 데이터로 학습되어 개인정보 보호 및 동의에 대한 우려가 제기되고 있습니다.

Facebook의 공공 데이터 활용과 새로운 AI 모델

페이스북과 인스타그램은 페이스북의 데이터를 사용하지 않는다는 기존 주장과는 달리, 공개 게시물과 댓글을 사용하여 생성형 AI 모델을 학습시키고 있습니다.

MRAW AI는 사용자가 업로드된 이미지에 대해 질문할 수 있는 비전 모델인 Pixol 12B를 출시했습니다.

OpenAI의 전 직원이 버전 관리 및 멀티모달 기능을 갖춘 언어 모델 사용을 최적화하기 위해 오픈소스 프롬프트 엔지니어링 라이브러리인 E11을 출시했습니다.

Adobe의 AI 벤처는 안전한 상업적 사용에 중점을 두고 있으며, 영화 같은 품질과 유연한 움직임을 위해 소유 또는 라이선스된 IP를 학습합니다.

이 회사의 텍스트-투-비디오 제품은 이미지 업로드와 프롬프트 기반 비디오 생성 기능을 제공하여 추가 영상을 캡처할 수 있는 솔루션을 제공합니다.

한편, 클라라의 CEO는 소프트웨어 제작의 통합과 효율성을 목표로 AI를 사용하여 기술 스택을 다시 작성하고 Salesforce 및 Workday와 같은 주요 SaaS 제공업체를 폐쇄하고 있습니다.

새로운 오디오 개요 기능은 문서를 매력적인 토론으로 전환할 수 있어 독서에 비해 그 효과에 대한 의문이 제기되고 있습니다.

OpenAI의 01 버전은 AGI 부문 Arc 상 벤치마크에서 우수한 성적을 거두며 화제를 모았지만, 원시 모델 테스트 이상의 필요성을 강조했습니다.

Tencent의 Game Gen O는 오픈월드 비디오 게임 생성을 위한 확산 트랜스포머 모델을 도입하여 고품질의 혁신적인 캐릭터 역학을 위한 게임 엔진 기능을 시뮬레이션합니다.

제너레이티브 모델을 사용하여 비디오 게임의 사실적인 캐릭터와 환경 디자인을 만드는 새로운 프로젝트가 게임 개발의 중요한 변화를 보여주고 있습니다.

Lee는 3D 인식 및 공간 세계 모델링을 갖춘 인공 지능 개발을 목표로 2억 3천만 달러의 자금을 지원받아 월드 랩을 출범시켰습니다.

테슬라의 실제 세계 비디오 데이터는 공간 지능 AI를 개발하려는 회사의 목표를 발전시키는 데 중요한 역할을 할 수 있습니다.