제프리 힌튼: 노벨상 수상자이자 AI 선구자

### 제프리 힌튼의 노벨상 수상과 획기적인 AI 개발
* 🏆 인공지능의 대부로 알려진 제프리 힌튼은 **1986년 역전파 알고리즘에 관한 중요한 논문**을 공동 집필하는 등 이 분야에 획기적인 공헌을 한 공로로 **노벨 물리학상**을 수상했습니다.
* 🧠 그의 혁신에는 **볼츠만 머신, 시간 지연 신경망의 개발, 2012년 이미지 인식에 혁명을 일으킨 AlexNet의 창시** 등이 있습니다.
* 😟 이러한 찬사에도 불구하고 힌튼은 잘못된 정보와 군사적 활용 등의 위험성을 이유로 **2023년 5월 구글에서 사임**하면서 AI와 관련된 위험에 대해 우려를 표명했습니다.
* 🎥 다른 소식으로, 메타는 오픈 소스 접근성과 인상적인 시청각 동기화 기능을 약속하는 **300억 개의 파라미터를 가진 텍스트-비디오 모델인 MovieGen**을 소개했습니다.
* 🎨 또한 Halo AI는 **이미지-투-비디오 기능**을 출시하여 **픽사 영화에 필적하는 고품질 애니메이션 콘텐츠 제작**의 가능성을 보여주었습니다.
* 💻 엔비디아는 **비전 언어 작업에 탁월한 멀티모달 AI인 NVM 모델**을 공개했으며, 리퀴드 AI는 **더 적은 메모리 공간과 뛰어난 성능**을 자랑하는 리퀴드 기반 모델을 소개했습니다.
* 📱 Apple은 **10월 28일에 유용한 알림 요약과 함께 이메일 및 캘린더 관리의 개인 지원을 위한 기대되는 AI 통합 기능**이 갖춘 iOS 18.1을 출시할 예정입니다.
* 💰 마지막으로, OpenAI는 확장하는 OpenAI의 리소스 수요를 충족하기 위해 고군분투하는 Microsoft와의 긴장 속에서 **역사적인 66억 달러의 자금을 모금**했습니다.

인공지능의 대부인 제프리 힌튼은 인공지능의 잠재적 위험성에 대한 우려를 표명하면서 노벨 물리학상을 수상했습니다.

the first story today. I'm super excited to talk about Jeffrey Hinton, who just won the Nobel Prize in physics. If you're not aware of who Jeffrey Hinton is, he is widely regarded as the Godfather of artificial intelligence. Basically, if not for him, we wouldn't have AI as it is today. Let's just read a little bit about him. In 1986, he co-authored a highly influential paper on the backpropagation algorithm for training multi-layer neural networks. He co-invented Boltzmann machines, a type of stochastic recurrent neural network. He developed techniques like distributed representations, time-delay neural networks, and mixtures of experts. In 2012, recently, he and his students created AlexNet, a breakthrough in image recognition that significantly outperformed previous systems. In 2018, he won the Turing Award. He won the Honda Prize in 2019, the Royal Medal in 2022, and just now the Nobel Prize for physics in 2024. But Jeffrey Hinton is also quite nervous about AI. In May 2023,

Hinton publicly announced his resignation from Google, citing concerns about the risk of AI technology. He's become an outspoken critic of the potential dangers of AI, expressing worries about its ability to generate convincing misinformation and its potential military applications. And in response to the phone call about the Nobel Prize, he said, I'm in a cheap motel in California, which doesn't have good internet or phone connection. I was going to have an MRI scan today, but I'll have to cancel that. So, big congrats to Jeffrey Hinton, well-deserved absolute legend in the AI field. Next, Meta released an absolutely stunning text-to-movie product. I covered it in the live stream on Friday, but this is absolutely incredible, and it's not just text-to-video. Let's just read a little bit about it. First of all, MovieGen video, so that is text-to-video, is a 30 billion parameter Transformer model. So, a few things here. One, most likely it's going to be open source because it's coming from AI and Meta. And two, it's 30 billion

parameters, which means it could actually fit on consumer-grade hardware, which is incredible. Next, it also has a 13 billion parameter text-to-audio, which is really, really nice. Now, what's super interesting about it is you can load up a movie, and it will give you audio sound effects, music that line up perfectly well with the video. It's super impressive. Here's an example of a generated video with generated audio. Let's take a look. [Music] [Applause] Next, you can also do precise video editing. So, you could basically just describe what you want changed, and it will change it. And then, you could also simply take a picture of yourself and put it into a video or prompt it to create a video based on your likeness. And I can't imagine any way that that would be abused. So, here's an example of editing video with text. So, we have the original video here. Then, we simply say, Add blue pom-poms to the hands. We say, Turn it into a cactus desert. And we can also say, Turn it into a running

inflatable dinosaur. Here's another impressive one. Here's the original video. Dress them up as penguins from the Victorian era, add beach umbrellas, and then make it look like a drawing. All of these are incredibly impressive. I'll drop links to all of these in the description below. And then to produce those personalized videos, you take a picture of yourself and then prompt it, and it will insert you into the video. It is so, so cool. So here's another one. This never happened, simply a 2D image, and then the prompt, and there it is. There's the video. So really cool. It's not out quite yet. They're still doing some testing. Hopefully, this is not a SORA where we basically will never get it, but I hope we're going to get it soon. Speaking of text, the video, we have another one, Halu Halo AI. I don't know how to pronounce that, but Halo AI has now released their image-to-video feature, and you can use it right now, right away. Let's look just at a few.

examples so you can definitely see a few mistakes here and there, but overall the consistency, the physics, they look pretty darn good. Here's the prompt, a very, very long prompt in fact. Let's look at another example. So here we go. I mean, this honestly could be a Pixar movie easily, and this really shows what's going to be possible with movie creation, television show creation, all basically entertainment creation in the very near future. So check it out. Hello, haloo, haloo AI. I'll just drop it in the descript. 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, L4s, and more available right now. And you can get a fraction of a card or a fully dedicated bare metal system, 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/bman, and make sure to use code Burman300 at checkout to get that credit. Thanks again to Vulture for sponsoring this.

video. Next, we have two brand new incredible models. First, an open-source model by Nvidia. This is an open massive model multimodal that can rival GPT-4. The model is called NVM, and I actually haven't tested it yet. Do you want me to test it? Let me know. So, we introduced NVM 1.0, a family of frontier-class multimodal large language models that achieve state-of-the-art results on vision-language tasks. So, here is an example. So, first, a picture of Jensen. Who is this person in the image? Jensen Hang. Now, one interesting thing is a lot of models will just refuse to do identification tasks. The Pixol model did it really well, although Llama 3.2 refused to identify Bill Gates's picture. But obviously, NVM is totally willing to do it. So, here's one. Explain why the meme is funny. Here's the abstract. Here's the paper. And there. Here it is, the explanation. What is the difference between the left, middle, and right object in the image? Different types of chips. All the way to write code based on the

provided pseudo code written just in messy writing, and then we have actual code to implement it right there. The model demonstrates versatile capabilities in various multimodal tasks by jointly utilizing OCR, reasoning, localization, Common Sense, World Knowledge, and coding ability. This is an all-in-one model. I'm so excited to test it out. So, we get the code, we get the weights, we get a paper, we get the benchmarks. Nvidia absolutely crushed it with this release. Thank you to Nvidia, that is so awesome. I will be testing it soon. And for the next story, we have a new model from liquid AI. I've been hearing about liquid AI for a while, but only Rumblings about it. And apparently, it's from a team out of MIT, and people have been saying this is really the next Frontier of models. But let's take a look. So, with this little graphic, we can see here are all the models, and it comes in a family of different models: lfm 1B, 1 billion, 3 billion, and 40 billion mixture of experts, and all the models across the

board beat all the other models that they have on this Benchmark. Now they're calling these the liquid foundational models, a new generation of generative AI models built from first principles. It comes in 1 billion, 3 billion, and 40 billion, the 40 billion being a mixture of experts and maintaining a smaller memory footprint and more efficient inference. So, you can try it today on the liquid playground, Lambda, perplexity labs, and Cerebrus inference. And as we can see for the 1.3 billion parameter model, it outperforms Llama 3.2, but it does have only a 32k context window. Now moving up to the 3 billion parameter version, as we can also see, it outperforms essentially all the other models in its size class. Then at the 40 billion, let's compare it to Mixel quen 2, gem 2 y 1.5, Llama 3.1, and as we can see, the model performs extraordinarily well. But where do these models really shine, and that is in the

memory footprint. As we can see, the inference memory footprint is much lower for the LFMs compared to the other models on this chart. Now, they are using a different architecture than Transformers, and that's what makes them very different. I'm not going to dive too deep into all the details, but if you want me to do a deep dive into LFMs, let me know, and I'm happy to do that. Next, Flux 1.1 is now out. Now, if you don't remember, Flux is the completely open-source text-to-image generation model from a team that was previously at Stability AI. They created this incredible text-to-image generation model that is open source. It's used by Grok, and now they have version 1.1. Flux 1.1 provides six times faster generation than its predecessor, Flux 1.0, while also improving image quality, prompt adherence, and diversity. Here, we can see the ELO scores of different text-to-image models, and Flux 1.1 is way up here as the winner, and here's Mid Journey 6.1.

right there idog V2, idog V2 turbo. So yeah, it basically pretty widely beats the other models in ELO score as compared to cost. It also still does quite well, so definitely the highest ELO score and about right in the middle of the cost spectrum. And in terms of speed, it's really fast, and that's really where it shines. And so, it has not only the highest ELO score but one of the fastest inference times as well. So, check it out, let me know what you think. Next, Apple is reportedly releasing iOS 18.1, which is really the Apple intelligence launch later this month, specifically October 28th. I just got an update on the developer Beta release last night on my iOS device, but it doesn't include any of the features that we really have come to expect from Apple intelligence. You've seen commercials of incredible functionality where the actress from The Last of Us says, Who's this person? Apple intelligence looks through essentially all of her emails and calendars and gets that person's information pretty darn

impressive. I cannot wait to try this out. I hope it's going to be really great because this is the missing piece of AI in my life. I need AI that will actually be able to accomplish tasks based on things in my personal life: my calendar, my email, my drive, all of these things that I am right now willing to give AI access to. It just hasn't been able to do that quite yet. Now, that isn't to say that the current beta features aren't useful. Actually, one of the most useful ones is notification summaries. So, I get a ton of notifications, as everybody does, and rather than showing each one, it buckets them all together and then just gives you kind of a two or three-sentence summarization of what all of those notifications are. And I've actually found that super useful and use it all the time. I don't use any of the writing tools, and I do love the new Siri animation. I've also played around with some of Apple Photos features, such as all you do is describe an album, and it

will grab all of the photos and put together an album for you based on that essentially prompt. But what I really want is that deep integration into my emails, my calendar, etc. And hopefully, we get that at the end of this month. Next, I already covered this in depth on Friday, but I'm just going to talk about it briefly. OpenAI released Chat GPT Canvas, which is really a huge upgrade. Although it's still pretty basic in functionality, it essentially allows you to do coding and writing tasks in a separate window with all of this additional functionality. So, very similar to Claude's artifacts feature, but a little bit more simple. And it's a great step to giving people more ability to create and use large language models to code and write, and just giving the tools necessary for people to do things without having to just prompt everything over and over again. And last, the news is already a bit old. About a week ago, OpenAI has raised the biggest round of private funding in the history of business. They raised $6.6 billion of new

funding at a $157 billion post-money valuation to accelerate progress on our mission. Now, there's already been a bunch of drama around this rise. It was rumored that Sam Altman was going to get a massive payday from it, although he denied all of that. It's also been rumored that they told their investors not to invest in other AI companies like Anthropic, but we don't actually have firm confirmation of that either. Either way, they raised a ton of money, and they're going to need it, it seems. The relationship between OpenAI and Microsoft is getting ripe with friction. One, if OpenAI ever calls their product AGI, Microsoft no longer gets access to it. Microsoft already owns like 49% of the company. Plus, just a couple of months ago, in a public report, Microsoft declared OpenAI a competitor. So, there's just so much to think about in this awkward relationship between Microsoft and OpenAI. Plus, OpenAI is now moving to become a for-profit, and it's just a mess. But either way, they raised a ton of

money, and it's also been reported that Microsoft isn't even able to provide enough servers for OpenAI. OpenAI is just eating up all of their resources, and Microsoft cannot give it to them fast enough. So, I'm going to keep digging into this story, but for now, that's all we got. If you enjoyed this video, please consider giving a like and subscribing, and I'll see you in the next one.

AI News: Meta Movie Gen, Hinton Wins Nobel Prize, NVIDIA's New Open SOTA Model, Flux1.1 Pro

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

제프리 힌튼의 노벨상 수상과 메타의 텍스트-투-무비 제품

AI의 대부라 불리는 제프리 힌튼은 신경망과 이미지 인식 분야의 획기적인 연구로 2024년 노벨 물리학상을 수상했습니다.

하지만 힌튼은 이러한 성공에도 불구하고 AI의 잠재적 위험성에 대한 우려를 표명했고, 결국 구글에서 사임했습니다.

한편 메타는 300억 개의 매개변수 모델을 갖춘 놀라운 텍스트-투-무비 제품을 출시하여 정밀한 비디오 편집 및 오디오 동기화 등의 기능을 제공함으로써 잠재적으로 업계에 지각 변동을 일으킬 수 있습니다.

AI는 해변 우산을 쓴 빅토리아 시대의 펭귄과 같은 역동적인 장면에 사용자의 사진을 삽입하여 멋진 개인 맞춤형 동영상을 제작합니다.

Halo AI는 이미지-투-비디오 기능을 출시하여 인상적이고 일관된 결과를 제공하며 엔터테인먼트 제작에 혁명을 일으킬 수 있는 잠재력을 가지고 있습니다.

이 비디오의 스폰서인 Vulture는 업계 최고의 가격 대비 성능, 유연한 확장성, 첫 30일 동안의 $300 크레딧을 통해 GPU 워크로드를 제공합니다.

엔비디아의 NVM 모델과 리퀴드 AI의 새로운 오퍼링

엔비디아는 GPT-4에 필적하는 오픈 소스 멀티모달 대규모 언어 모델인 NVM 1.

0을 출시하여 비전 언어 작업 및 코딩에서 다양한 기능을 선보였습니다.

이 모델의 고유한 기능에는 식별 작업 수행, 밈 설명, 복잡한 의사 코드에서 코드 구현 제공 등이 포함됩니다.

MIT의 Liquid AI 팀은 10억, 30억, 400억 개의 파라미터를 가진 생성형 AI 모델 제품군을 도입하여 최첨단 결과를 달성하고 여러 전문가를 통해 효율적으로 추론할 수 있는 기능을 제공합니다.

LFM은 다른 모델에 비해 추론 메모리 풋프린트가 더 적으며, Transformers와는 다른 아키텍처를 사용합니다.

오픈 소스 텍스트-이미지 생성 모델인 Flux 1.

1은 이제 6배 빠른 생성 속도와 향상된 이미지 품질, 신속한 준수 및 다양성을 제공합니다.

1을 출시할 예정이지만, 개발자 베타 버전에는 예상되는 Apple 인텔리전스 기능이 포함되어 있지 않습니다.

개인 생활에서 AI 통합의 필요성

스피커는 캘린더, 이메일, 드라이브 관리와 같은 개인 생활과 관련된 작업을 AI가 수행하기를 열망합니다.

여러 알림을 통합하고 요약하는 알림 요약 기능이 유용하다고 생각합니다.

또한 새로운 Siri 애니메이션과 Apple 사진의 앨범 생성 기능도 만족스러워하지만, 조만간 이메일 및 캘린더와 더 긴밀하게 통합되기를 기대하고 있습니다.

OpenAI의 1,570억 달러 자금 조달과 마이크로소프트와의 마찰

OpenAI가 1,570억 달러의 자금을 조달하면서 샘 알트먼의 거액 연봉에 대한 소문과 투자자들과의 긴장이 고조되고 있습니다.

Microsoft는 OpenAI를 경쟁자로 규정하고 회사의 성장을 위한 충분한 리소스를 제공하기 위해 고군분투하는 등 OpenAI와 Microsoft의 관계는 긴장된 상태입니다.

영리 조직으로 전환하려는 OpenAI의 움직임은 Microsoft와의 파트너십을 더욱 복잡하게 만들었습니다.