Claude 3.5 하이쿠 출시: 가격 논란과 혁신

Claude 3.5 하이쿠는 이전 모델보다 인상적인 성능을 보였지만 예상치 못한 가격 인상에 대한 반발에 직면했습니다.

### AI의 최근 발전: Claude 3.5 하이쿠, OpenAI의 컴퓨팅 과제, 로보틱스의 혁신
* 📰 클로드 3.5 하이쿠는 **신형 클로드 3.5 소네트를 제외한 이전 모델보다 성능이 뛰어나면서도 훨씬 더 저렴**하다는 주장이 제기되며 논란 속에 출시되었습니다.
* 💰 그러나 클로드 3.5 하이쿠의 가격은 **입력 토큰 100만 개당 1달러, 출력 토큰은 5달러**로 책정되어 많은 이들에게 충격을 주었으며, 이는 이전 가격보다 상당히 인상된 가격입니다.
* 📈 앤트로픽은 **테스트 기간 동안 이전 플래그십 모델을 능가하는 향상된 인텔리전스**를 강조하며 가격 인상을 옹호하고 있지만, 커뮤니티에서는 이러한 움직임에 대해 호의적이지 않은 반응을 보이고 있습니다.
* 🤖 OpenAI는 컴퓨팅 한계로 인해 제품 출시가 지연되고 있으며, CEO인 샘 알트먼은 **더 많은 리소스가 필요함**을 시사하고 **자체 데이터 센터를 구축할 계획**을 암시했습니다.
* 🧺 Physical Intelligence는 인공 물리 지능 개발을 목표로 하는 **Pi Pi0 모델**을 기반으로 세탁과 설거지 같은 가사 작업을 수행할 수 있는 새로운 로봇을 공개했습니다.
* 🎮 게임 업계에서 획기적인 행보로, **에픽과 데카르트 AI가 협업하여 디퓨전 모델로만 생성된 플레이 가능한 마인크래프트 게임**을 개발하여 콘텐츠 제작의 미래를 선보였습니다.
* 💡 최근 Reddit AMA에서 OpenAI의 경영진은 **API 비용 절감 계획**과 **모델에 멀티모달 기능**을 도입할 예정 등 다양한 주제에 대해 논의했습니다.
* 🌐 한편 메타는 **터치 감지 기능**을 통합하여 로봇이 주변 환경과 더욱 자연스럽게 상호작용할 수 있도록 함으로써 로봇 공학을 발전시키고 있으며, 이는 **섬세한 조작이 필요한 작업**에 필수적입니다.

We have a ton of news from the last few days. Let's go over it all. First, Claude 3.5 Haiku is now here and not without some drama. Now, if you remember, there was an announcement a few weeks ago about Claude 3.5 Sonet new, and as part of that announcement, we got computer control, and we had this announcement that Claude 3.5 Haiku was coming, and now it's here. So, according to Anthropic, Claude 3.5 Haiku surpasses all previous Claude models except the new 3.5 Sonet on coding and a genetic task while being significantly more affordable, a fraction of the cost. Here are the S-bench verified scores, and I want to share something with you. I'm actually going to be interviewing the team behind S-bench, so stay tuned for that. But looking back here, here we go: Cloud 3.5 Sonic new, Cloud 3.5 Haiku beating even Cloud 3.5 old. You get a 200k context window and a 50% discount.

with the Batch's API, and here's where the drama lies: the price is actually a lot more expensive than people were expecting. I got Chat GPT's search feature to tell me exactly what was going on with the pricing. The model is priced at a dollar per million input tokens, $5 per million output tokens. A significant increase from the previous rates of 25 cents and $1.25, respectively. Seven times more expensive than GPT-4 mini and much more costly than Gemini 1.5 Flash. Anthropic justifies the hike by citing the model's enhanced intelligence and performance, stating that during final testing, it surpassed even its previous flagship model. Now, a lot of people called them out for this, and then not only that, on Halloween, they put out a blog post, The Case for Targeted Regulation. I'm not going to go too deep into that; if you want me to do a deep dive on this, let me know in the comments. But for now, just know it was not well received. So, between the model's price and then what a lot of

people are calling their aim for regulatory capture. This entire announcement was just not well received. 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 solution ions. 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 fully dedicated.

bare metal systems, which give 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, we finally know why OpenAI has been delaying their newest products, and the reason? 'Cause they don't have enough compute. Maybe Microsoft is not giving them enough. Maybe they shouldn't depend entirely on Microsoft. We're not really sure exactly what's going on, but Sam Altman made it clear compute is the bottleneck. And now it makes sense why he wanted to spend $7 trillion trying to build out his own data centers. Why it has been rumored that OpenAI is going

to create their own chips, a lot of different things need to happen to open up the compute necessary to actually get these next-generation models in our hands. Next, we have a new robot. Physical Intelligence has released a new robot, and check this robot out. It is made to do your laundry and your dishes and everything in between. So, we've seen demos like this in the past, but this just looks incredibly impressive. Check out this robot folding laundry really well, very smooth motions, everything looks really good. And what it actually is, is the model itself, the model that is being used to power the robot. They're not actually selling the robot, as far as I can tell. In their blog post, they say, Over the past eight months, we've developed a general-purpose robot foundation model that we call Pi Pi0. We believe this is the first step toward a long-term goal of developing artificial physical intelligence so that users can simply ask robots to perform any task they want, just like they can ask large language models and chatbot assistants. So, the website has a bunch of examples.

videos where you can check out how capable this robot and really the underlying model is. And next, in a huge step towards the future of video games, another huge step, we have Oasis, the company Etched, partnered with Deckart AI to create a Minecraft game entirely using diffusion models. So essentially, there's no underlying code, no underlying logic, no game engine, nothing, and it is fully playable right now. Here's an example of what it looks like. Look how cool this is. Obviously, the resolution is very small, but still, it just shows you another glimpse into the future of what content and video games are going to look like. Next, Sam Altman and the OpenAI team, the leadership, did an AMA on Reddit, and Chubby put together a bunch of the highlights, so let's take a look at some of them. So first, are you planning to reduce the API cost of advanced voice? We've been reducing the cost of our APIs for 2 years now, and so it seems like they're going to continue to do that. Then somebody asked about inference cost specifically in regards to 01 and Chain.

of Thought reasoning and so how fast does open AI see inference cost reducing. We expect inference costs to keep going down. If you see the trend over the last year, it's come down like 10x. And my favorite, what do I always say, bullish on agents. What's the next breakthrough in GPT line of product, and what's the expected timeline. Sam Alman, we will have better and better models, but I think the thing that will feel like the next giant breakthrough will be agents. Somebody asked why 01 doesn't support image input, and here we go. The chief product officer says we focus on getting it out to the world first, first waiting to make it fully featured. Image input is coming in 01. That's great news. And in general, the O series of models will be getting things like multimodality, tool use, etc., in the coming months. They are going to be so powerful once they get all of the features that we've become used to with the other Frontier models. Somebody asked what's one thing you wish Chachi.

BT can do but can't yet. I'd love for it to understand my personal information better and take actions on my behalf. If you've watched this Channel at all, that is my dream. I want AI agents to be able to actually accomplish things on my behalf based on my personal information. That is why I also say Apple and Google are incredibly well positioned to still win the AI race, because they own the hardware in our pockets and they also really own all of our private data between Gmail, Google Docs, iCloud, everything. So definitely check out the full AMA on Reddit. I'll drop the link in the description below. Next, chat GPT search came out. I already made a full video of it, but it seems really cool and it is a competitor to perplexity. You search, it can scrape the internet, and it displays the results based on real-time information. It really is a huge step forward for chat GPT. Now I use perplexity all the time. I'm also going to be using chat GPT all the time. I'll let you know which one I like best next.

Meta is releasing touch robots that can actually sense touch and feel. It's pretty amazing. And so, Meta is giving the ability to robots to actually feel, and this is important. In fact, having all five senses as embodied AI, a robot, is really important for understanding and being able to navigate the world around you. It allows robots to be able to grab things more accurately, to apply the right amount of pressure based on whatever they're touching. For example, if a robot is petting a dog, you don't want it to use its full force. So, how would it actually know? It could obviously predict what the right force is, but if it actually had the feedback of touch, it would be able to do so so much more accurately. So, really cool. Good job, Meta. Next, there was a model that was topping the text-to-image leaderboards, which was called Red Panda. And now we know what it was. It is a company called Recraft, and they released a new text-to-image model. Here are some of these examples, and they really look absolutely

incredible. Here's a breaking bag cake, an alien taking a selfie. Here's another one. As I always say, more competition, better. Let's get more text-to-image models, text-to-video models, large language models. I can't get enough of them. Congrats to recraft. All right, next. Here's a story that blew me away and I think went under the radar. It turns out 25% of the code being written by Google Engineers now is written by AI. That is a huge increase in productivity of the engineering organization that is Google. Sundar Pichai, the CEO of Google, said it in an earnings call. More than 25% of all new code at Google is now generated by AI. Now, a Google engineer actually clarified what that means. Now, yes, AI is writing a lot of code, but it's not writing end-to-end code yet. What he said it's actually doing is more like tab completion. So, you start and you can architect a piece of code and it will help you complete that code. Now, that is still amazing and increases productivity.

greatly, but it's not quite end-to-end code generation by AI. Again, I still see the future of code being completely AI-generated, and what this just means is that a single engineer will just be that much more efficient and that much more productive. Next, in a minor update, Chat GPT now lets you search through your chat history. So, if you have thousands of different chats with Chat GPT, now it's easy to find the one you're looking for. Now, what I think might actually be cool is being able to chat with your chat history, but maybe that's getting too meta. So, a nice little quality-of-life update there. 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.

AI News: Claude 3.5 Haiku, AI Minecraft, OpenAI AMA, Google Reveals AI Writing Tons of Code

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

5 소네트를 제외한 모든 이전 모델을 능가하면서도 가격은 훨씬 더 저렴하다는 주장으로 클로드 3.

이 모델의 가격이 급격히 인상되자 반발이 일어났고, Anthropic은 이 모델의 향상된 인텔리전스와 성능을 강조하며 가격 인상을 정당화했습니다.

이 발표는 표적 규제에 대한 논란의 여지가 있는 블로그 게시물과 함께 대중의 부정적인 피드백을 받았습니다.

Vulture의 베어메탈 시스템과 OpenAI의 컴퓨팅 병목 현상

Vulture는 하드웨어에 대한 완전한 제어와 빠른 애플리케이션 설치를 위해 베어메탈 시스템을 제공하며, 첫 30일 동안 300달러의 크레딧을 제공합니다.

OpenAI의 제품 지연은 컴퓨팅 부족으로 인한 것으로, 자체 데이터 센터와 칩을 구축할 계획으로 이어지고 있습니다.

물리 지능은 작업 수행을 위한 인공 물리 지능 개발을 목표로 범용 기반 모델을 갖춘 로봇을 출시했습니다.

기본 코드나 게임 엔진 없이 디퓨전 모델만으로 Minecraft 게임을 만들었으며, 현재 완전히 플레이할 수 있습니다.

샘 알트먼을 비롯한 OpenAI의 경영진은 Reddit AMA를 개최하여 API 비용 절감 계획과 01 모델의 이미지 입력에 대한 예상 일정에 대해 논의했습니다.

O 시리즈 모델은 앞으로 몇 달 안에 멀티모달리티 및 도구 사용 기능을 추가하여 더욱 강력해질 예정입니다.

AI 에이전트는 개인 정보를 더 잘 이해하고 사용자를 대신해 조치를 취할 것으로 예상되며, 하드웨어와 개인 데이터에 대한 통제권을 가진 Apple과 Google이 AI 경쟁에서 유리한 위치를 선점할 것으로 보입니다.

Chat GPT의 새로운 검색 기능은 실시간 정보를 제공하며 퍼플렉시티와 경쟁하고, 메타의 터치 로봇은 더 나은 탐색과 작업 수행을 위해 오감을 재현하는 것을 목표로 합니다.

Recraft의 새로운 텍스트-이미지 변환 모델인 Red Panda는 인상적인 결과로 순위표에서 1위를 차지하고 있습니다.

AI가 Google 엔지니어링 및 채팅 GPT 업데이트에 미치는 영향

이제 AI는 Google 엔지니어가 작성하는 코드의 25% 이상을 담당하며 생산성을 크게 향상시키고 있습니다.

AI의 역할은 코드를 처음부터 끝까지 생성하는 것이 아니라 탭을 완성하는 것과 비슷하며 코드를 설계하고 완성하는 데 도움을 줍니다.

Chat GPT는 채팅 기록을 검색하는 기능을 도입하여 사용자가 이전 대화를 더 쉽게 찾을 수 있도록 했습니다.