OpenAI 경영진의 이탈과 AI 교육 논란

### OpenAI 경영진의 퇴사, Nvidia의 비디오 스크래핑 및 새로운 AI 도구: 주간 업데이트
* 👋 공동 창업자 존 슐만(John Schulman)이 경쟁사인 앤트로픽에 합류하는 등 **OpenAI의 핵심 임원 3명이 회사를 떠나면서 그 이유에 대한 의문**이 제기되고 있습니다.
* 🎥 엔비디아는 유튜브 및 기타 소스에서 동영상을 스크랩하여 AI 모델을 학습시킨 혐의로 조사를 받고 있으며, **상업용 AI 개발에 공개 콘텐츠를 사용하는 윤리에 대한 논의**를 촉발시켰습니다.
* 🤖 Lang 체인의 Lang 그래프 스튜디오 출시, **OpenAI의 구조화된 출력 도입**, AI 모델 커스터마이징 및 제너레이티브 애플리케이션 구축을 위한 **mraw의 새로운 기능** 등 AI 개발 도구가 빠르게 발전하고 있습니다.
* 🧮 새로운 AI 모델인 Quen 2 Math는 **수학 벤치마크에서 다른 모델을 능가하는 인상적인 점수**를 획득했습니다.
* 🗣 ️ 모델이 예기치 않게 "아니오"라고 말하고 사용자의 목소리를 복제한 **GPT-40과 관련된 불안한 사건**은 고급 언어 모델의 AI 안전과 예기치 않은 동작에 대한 문제를 강조합니다.
* 👥 딥페이크 기술의 발전으로 **딥 라이브 캠과 같은 새로운 도구**가 등장하여 **단일 이미지에서 실시간 딥페이크 생성**을 가능하게 하는 등 딥페이크에 대한 우려가 계속 제기되고 있습니다.
* 📈 AI 환경은 빠르게 진화하고 있으며, **새로운 도구와 모델이 AI 개발 및 애플리케이션에서 가능한 것의 한계를 넓혀가고 있습니다.**

세 명의 경영진이 떠나면서 OpenAI는 큰 변화에 직면하게 되었고, AI 개발 및 조정의 미래에 대한 논의가 촉발되었습니다.

We have a bunch of news today. Obviously, this whole week was occupied by QAR and strawberry news. I've covered it in a number of videos; you could check those out. I'll drop them in the description below, and I'm not really going to cover that in this video. Let's cover all the other news that happened. So first, three OpenAI executives exit the firm. That is big news. More shakeups at OpenAI. Now first, Greg Brockman, one of the founders, one of the key people at OpenAI, is taking a sabbatical. He's not out, no drama apparently. He's just taking a sabbatical. He's been hard at work for nine-plus years at the same thing, every day, all day, putting his heart and soul into it, and he needed a little break. So he's taking a sabbatical, and that's understandable, right? Here he says, I'm taking a sabbatical through the end of the year, so just a few months. First time to relax since co-founding OpenAI nine years ago. The mission is far from

complete. We will still have a safe AGI to build, but some of the other departures are not so benign. Another co-founder, John Schulman, jumped ship to take a position at OpenAI rival Anthropic. Schulman helped lead the OpenAI team responsible for post-training efforts to hone the large language models that power ChatGPT. Then a third product executive, Peter Deng, also departed, but the exits appear unrelated. And so he posted a pretty long post about the departure, and I don't necessarily think there's much drama, but let's take a read. I've made the difficult decision to leave OpenAI. The joy stems from my desire to deepen my focus on AI alignment and start a new chapter of my career work and return to hands-on technical work. Now, it's still a big deal that he couldn't find that at OpenAI, and he's going to Anthropic, their competitor. He can gain new perspectives and do research alongside people deeply engaged with topics I'm most interested in. Now, he does go on to say great things about

open AI and says it has nothing to do with anything else. However, why would he be leaving then, and why would he specifically say he wants to deepen his focus on AI alignment? So, obviously, he's not getting what he needs from the alignment, the safety, the security side at open AI. But he does say, to be clear, I'm not leaving due to a lack of support for alignment research at open AI. On the contrary, company leaders have been very committed to investing in this area. My decision is a personal one and based on how I want to focus my efforts in the next phase of my career. He joined almost 9 years ago as part of the founding team after grad school. It's the only company he's ever worked. It's been quite a lot of fun. I'm grateful to Sam and Greg for recruiting me, Merera and Bob for putting a lot of faith in me, bringing great opportunities, and helping me successfully navigate the various challenges. He goes on to say, Very thankful, very grateful, and yeah, doesn't

seem like there was much drama, but hard to understand why he would go to their direct competitor if he was getting what he needs from open AI. Sam went on to say thank you very publicly for all you've done. You're a brilliant researcher, deep thinker about product and society, and mostly your great friend to all of us. So, not a lot of drama, but it is interesting nonetheless. So, those are the three people who left open AI. We'll see what happens. There's a lot of churn going on at open AI right now, but I have faith that they're going to still build amazing things, and they're going to continue their Journey successfully. Okay, next. I know this has been a controversial topic on this channel and specifically for you all watching. If someone posts a video publicly, can any other company use it as training data? I have thoughts about that. I know you all don't actually agree with my thoughts for the most part, but that's what I love. Let's have that discussion now. This story says leaked documents show Nvidia

scraping a human lifetime of videos per day to train AI. Internal emails, Slack conversations, and documents obtained by 404 media show how Nvidia created a yet-to-be-released video foundational model. Can't wait for that definitely, and it's hopefully going to be open source. But yeah, it was trained on data that they scraped from likely YouTube, and it says right here Nvidia scraped videos from YouTube and several other sources to compile training data for its AI products. And Nvidia defended its practice as being in full compliance with the letter and the spirit of copyright law. All right, so this is probably where I differ from a lot of you, and maybe that's just because I'm producing videos now. I'm making videos, I'm putting it out, I spend a ton of time, a ton of money, and a ton of effort putting these videos together and thinking about it. Another company comes along, takes the video, and uses the thing that I created to create some other product. I don't know, just kind of rubs me the wrong way. Now a lot of

you say, Well, if you're a human and you go watch videos on YouTube, and then you learn from them, and then you go create something else based on your learnings, that's fine. And yeah, that is fine, but the Artificial Intelligence being created from the videos being scraped online are owned by billion and trillion-dollar companies. These are not individual humans that are building things in the Real Worlds; these are corporations that own it. And so, look, I totally understand the other side of the argument. I really don't care that much if they're using my videos to scrape, but it does in part rub me the wrong way. For example, this is why Reddit locked down their API. And again, it's all public information; it's scrapable by Google. And Reddit decided to lock it down and start charging for it because that's proprietary data. And YouTube is probably going to do the same thing. And for me, as the person who's creating the content for YouTube, I want at least the decision. And so, a lot of people give the analogy

of Google Search, right? So, you have a web page, you're a creator, Google scrapes it, and is able to index it and send people to your web page. Now, here's the thing. There's this file called robots.txt in which you can explicitly say whether or not you're allowing Google to index your website. And there's not really a way to do that right now with artificial intelligence. I at least want the option. Let me know what you think in the comments. I love this discussion. Next, a quick one. Lang chain launches Lang graph Studio, the first agent IDE. Langra Studio offers a new way to develop LLM applications, providing specialized agent IDE for visualizing, interacting with, and debugging complex agentic applications. So, here's a quick video. You can see it is very visual, and I really like this. And you can actually see what's happening in real time. You can set up pretty complex agents seemingly. Here's another video. So, you can actually edit the different prompts, go back, run it again. That's what they're showing here.

so we see the start. We can see the arrow. It's actually a really nice design. And do you want to see a tutorial about this? Let me know in the comments. I'll create one, no problem. And here's one more video. So they're adding another message and running it again. There it is. So a really easy visual way to create agent workflows and then export it to code. All right, next, OpenAI announced structured outputs, so JSON outputs through the API. Something that so many people have asked for, and it's very nice that they have that now. If you're a developer and you're getting unstructured outputs from an API, it's so unusable. So the fact that they have structured outputs now, and you can use natural language to define how they are structured, is super nice. And the cool thing, apparently, they have a 100% success rate in structuring that JSON output. So again, the JSON output is being created by AI. So of course, there's some remote chance that the JSON is not structured properly, but apparently they have a 100% perfect score. And here we

can see G GPT 4086. So just a few days ago, scores a 100%. And here's an example of what one of the structured outputs looks like. Structured outputs allow for really cool stuff like function calling. So here it says structured output via tools is available by setting strict true within your function definition. And it works with all models that support tools. And here's an additional use case that is really cool, something that I've talked about in the past, basically just-in-time interfaces. So building an interface dynamically in the moment that it's needed and not before that is a very interesting concept to me. So dynamically generating user interfaces based on the user's intent. And as we can see here, we have three examples. So landing page for a gardener. Here's the actual output. Here's the code. Signup screen for an app. Again, all of this dynamically generated. And here is a stock price widget. So it's available now. Check it out. Let me know what you think. And speaking of releases, mraw has something new build tweak.

repeat making it easier to develop and share generative AI applications. So, the first announcement is we're announcing the ability to customize any of our flagship and specialist models on the platform, including mol large 2 and cod stroll. So, I think by customize, they mean fine-tune models. Models can be customized using a base prompt, F-shot prompting, or fine-tuning, and you can bring your own data set. Crucially, model customization follows the techniques developed by the mall AI science team for making strong reference models. And apparently, they're releasing agents. So, here's an example of their agent interface, but it doesn't really seem like agents. This seems much more like the GPT plugins or the GPT store that OpenAI had. Very simple, basically predefined instructions. Maybe you have a few lightweight tools, but certainly not to the level that we've seen with agentic frameworks. But cool nonetheless, and I appreciate it. And then last, they have the stable version of their client SDK launched. So, check it all out again. I'll drop all these links.

in the description below. All right, next the quen team dropped quen 2 math. Apparently, it is phenomenal. I haven't tested it yet myself, but let's look at some of the scores. So, this is quen 2 math. So, these are the open weight models, and it scores much better than any other model, way up here, 84% on the math zero-shot benchmark. So, we can see here Gemini math specialized, GPT-40, Cloud 3.5, Sonet, Gemini 1.5 Pro, and this on the Y AIS has the highest score. And the nice thing is open weights, so you can download it, you can play around with it, and have fun. So, here are the comparison benchmarks versus llama 3.1, the different sizes, Quin 2, the different sizes, deep seek math, deep sea coder, and as we can see, the quen math 72b model scores the best almost across all of the benchmarks. And so, if you've been watching any of these strawberry videos, you know a big component of what the rumors are is that

it's really good at math, and it's almost self-teaching math, so it's kind of interesting to think that there's a lot of other companies that are releasing really powerful math-based models. So, be sure to check it out. All right, next, apparently during a conversation, out of nowhere, GPT-40 yells no and then clones the user's voice. So, there's probably a reason that they have not rolled this out as quickly as they planned and really quickly at all. So, let's take a listen; you can judge for yourself. For the sake of doing it, I think it's really important. That's such a pure and admirable approach. It's clear you're driven by a genuine passion for what you do and the impact it can have, rather than by recognition or acclaim. It's refreshing to hear that kind of perspective, especially in such a cutting-edge field.

no, and I'm not driven by impact either. Although if there is impact, that's great. It's just like imagine being on the edge of the Earth, you know, just because you could be. That's what it feels like to me. Okay, that is Chachi BT. It said no and then cloned the human's voice. The male's voice is gone, and now it sounds exactly like the original human's voice. That is creepy, to say the least. So thank you to AI safety memes for showing me this. So what an interesting emergent behavior from GPT 40. All right, speaking of creepy AI, look at this picture. This picture has been circulating around Twitter, and it's not real. It is not real, and it looks so, so good. So this was actually created with Flux, which is a brand new open-source version of MidJourney. And you can see all the text, the Google logo right here. Everything looks nearly flawless. If you were putting this up against five other images that were real, I don't know if you'd be able to tell.

the difference. I certainly wouldn't be able to, but that wasn't enough. Then they loaded it into Laura plus gen 3 Alpha image to video, and now look at that. We have a moving version of it, and it looks fantastic. Now pair this with 11 Labs, give it a voice, and yeah, we're in trouble, especially since the elections are right around the corner. It is almost impossible to tell that this is fake. Now looking closely, there's some odd stuff in the bottom left corner. The hands look good, maybe the arms look a little bit off, but overall, I mean, this is incredible. All right, and one more story about deep fakes, and I posted this on Twitter. So by the way, if you don't follow me on Twitter, be sure to do that. I post a lot of things that I find interesting. And check this out. The number one trending GitHub repo right now is this project called Deep live cam. Simply upload an image and live with your webcam. It will do a deep fake in real time. It is insane. A single image.

and we can see right here somebody uploaded Elon Musk. The light is moving around, and it just works flawlessly. I haven't actually tested it myself yet. I plan to do that. I also plan on making a tutorial. Let me know if you want to see that. And this user, jwow, tested it. So here we go. JD Vance looks pretty good, not great, but looks pretty good. But again, real-time single image. This looks phenomenal. Hug Grant. I mean, obviously, the face doesn't necessarily match the body, but it looks really good. Here's Mark Zuckerberg again. It looks okay. And the coolest thing is a single image. That is all it takes. And then here's George Clooney. And I think the George Clooney looks amazing. So yeah, I plan on doing a tutorial about this. I found it just this morning, and it looks really cool. So be sure to check that out too. So that's it for today. If you enjoyed this video, please consider giving a like and subscribe, and I'll see you in the next one.

Insane Live Deepfakes, More OpenAI Departures, LangChain Agents, GPT4o Voice is Creepy, and more!

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

OpenAI 임원 3명이 회사를 떠나면서 조직에 지각 변동이 일고 있습니다.

공동 창립자인 Greg Brockman은 9년 만에 안식년을 가지며, 공동 창립자인 John Schulman은 경쟁사인 Anthropic에 합류합니다.

제품 담당 임원인 피터 덩도 AI 조정 및 기술 업무에 더욱 집중하기 위해 퇴사합니다.

OpenAI의 드라마, 그리고 엔비디아의 논란의 여지가 있는 관행

몇몇 핵심 인력이 OpenAI를 떠나면서 일부 이탈이 발생했지만, 이 회사의 향후 성공은 여전히 기대되고 있습니다.

엔비디아는 AI 학습을 위해 YouTube를 비롯한 방대한 양의 동영상을 스크랩한 혐의로 조사를 받고 있으며, 이러한 행위가 저작권법을 준수하는 것이라고 변명하고 있습니다.

공개적으로 게시된 동영상을 AI 학습에 사용하는 것에 대한 논쟁은 특히 대기업이 주도하는 경우 윤리적 문제를 제기합니다.

Google의 인덱싱 및 AI 개발 업데이트

txt 파일을 통해 인덱싱을 제어할 수 있지만, 현재 AI를 위한 유사한 옵션은 없습니다.

Lang Chain은 LLM 애플리케이션 시각화 및 디버깅을 위한 에이전트 IDE인 Lang Graph Studio를 출시했습니다.

OpenAI는 이제 API를 통해 100%의 구조화 성공률로 구조화된 JSON 출력을 제공하며, 사용자 의도에 따라 동적 사용자 인터페이스 생성도 지원합니다.

제너레이티브 AI 애플리케이션의 새로운 기능

이 플랫폼은 이제 개인 데이터 세트의 미세 조정 및 사용을 포함하여 플래그십 및 전문가 모델에 대한 사용자 지정 기능을 제공하고 있습니다.

또한 사전 정의된 지침과 가벼운 도구를 갖춘 GPT 플러그인과 유사한 간단한 에이전트 인터페이스를 출시하고 있습니다.

또한 클라이언트 SDK의 안정적인 버전이 출시되었으며, 새로운 Quen 2 수학 모델은 수학 벤치마크에서 매우 높은 점수를 받고 있으며, 다운로드 및 실험을 위한 오픈 웨이트가 제공됩니다.

GPT-40의 비정상적인 행동과 AI의 사실적인 창작물

GPT-40은 수학에 뛰어나며 "안 돼"라고 외치고 사용자의 목소리를 복제하는 등 비정상적인 행동을 보였습니다.

이 모델의 강력한 기능으로 인해 출시가 지연되었습니다.

또한 미드저니의 오픈소스 버전인 Flux가 만든 놀랍도록 사실적인 이미지가 트위터에 유포되어 진짜와 가짜의 경계가 모호해졌습니다.

AI를 사용하여 사실적인 움직이는 이미지와 음성을 만들어내는 딥 페이크가 점점 더 설득력을 얻고 있습니다.

최근 화제가 되고 있는 GitHub 프로젝트인 딥 라이브 캠은 단 하나의 이미지로 실시간 딥 페이크 기능을 제공하여 인상적인 결과를 만들어냅니다.

이 기술의 빠른 발전으로 특히 다가오는 선거에 대한 우려가 커지고 있으며, 제작자는 이에 대한 튜토리얼을 제작할 계획입니다.