주간 AI 업데이트: 로봇, 영향력 있는 운영 및 동반자 문제

### AI의 주요 발전: 휴머노이드 로봇, 인플루언스 운영 및 정서적 의존성
* 🤖 유니티 로보틱스는 인상적인 민첩성과 균형감을 자랑하는 **새로운 휴머노이드 로봇**을 공개하며 휴머노이드 로봇이 가정에서 보편화될 수 있는 미래를 예고했습니다.
* 🐕 또한 오프로드 타이어를 장착해 다양한 지형에서 기동성을 높인 **강아지 모양의 로봇 '언리 고토 W'**도 소개해 많은 잠재 사용자들의 관심을 모았습니다.
* 🌐 OpenAI는 미국 정치에 대한 오해의 소지가 있는 콘텐츠를 생성하기 위해 ChatGPT를 사용한 이란의 영향력 행사를 방해하여 **잘못된 정보에 대응하기 위한 AI 도구의 시급한 필요성**을 강조했습니다.
* 💔 AI 동반자 관계가 성장함에 따라 개인, 특히 청소년들이 **AI 목소리에 정서적으로 의존**하게 되면서 실제 인간관계에 미치는 영향에 대한 우려가 커지고 있다는 보고가 있습니다.
* 🛠 OpenAI는 개발자가 **100만 토큰당 25달러로 설정된 훈련 비용**으로 응답을 맞춤화할 수 있는 GPT-4에 대한 미세 조정을 시작하여 다양한 애플리케이션에서 사용성을 향상시켰습니다.
* 🚗 웨이모는 **주간 자율주행 운행 횟수 10만 건**이라는 이정표를 달성하여 자율주행 기술에 대한 강력한 수요와 운영상의 성공을 보여주었으며, 잠재적으로 **월 1,000만 달러의 수익**을 창출할 수 있습니다.
* 📈 캐릭터 AI는 **a16z의 상위 생성 AI 앱 목록에서 2위**를 차지했는데, 이는 점점 높아지는 인기와 사용자들의 AI 중독 트렌드를 반영한 결과입니다.
* 🔍 메타는 향후 모델 학습을 위해 공개 데이터를 수집하는 새로운 웹 크롤러를 출시하여 **콘텐츠에 대한 액세스를 허용하거나 거부할 때 퍼블리셔의 선택이 중요**하다는 점을 강조하고 있습니다.

휴머노이드 로봇은 점점 더 민첩해지고 저렴해지면서 가정 내 통합을 향한 변화를 예고하고 있습니다.

We had another big week in AI, and I'm going to go over all of the news. Let's just get right into it. We have a brand new video from Unitary Robotics showing off their incredible humanoid robot, and it looks so agile that it almost seems unreal. But let's take a look. So here it is. It looks to be kind of a new design, and I mean its jumping ability, its moving ability, its balance is unreal. We are almost certainly in the year of humanoid robots, and not only that, this robot is not that expensive either. So it seems like we're closer than we actually thought to reaching humanoid robots in every household. And right before this, it said it's using deep reinforcement learning to basically figure out how to navigate its terrain. Here's the other robot dog from Unitary Robotics, and I don't know why they always have to push the robot. It just seems mean. I will never do that because one day when AI becomes sentient, I don't want it to have any record of that, and that's not the

only robot announcement from Unry. They've also announced Unry Goto W, which is essentially that same dog-like form factor, except it has off-road tires. So now it could pretty much go anywhere. So here's the original one, here's the one with tires. Obviously, it's also going to be able to move a lot faster, so it can walk while it's going. It could easily crawl over large ledges. Now it's at a skate park. I mean, this just looks so, so cool. And yeah, I definitely want one in my house. Next, it seems OpenAI has disrupted a covert Iranian influence operation. And this isn't the first, and it's definitely not going to be the last. AI has the potential to be used for incredible harm, but we're hoping OpenAI has the tools to block, filter, find, and disrupt all of these types of operations. OpenAI said we banned accounts linked to an Iranian influence operation using ChatGPT to generate content focused on multiple topics, including the US presidential campaign. We have seen no indication that this content reached a

meaningful audience. This is the same thing that Facebook had to deal with with bots in the 2016 election, the 2020 election, and now it's just going to ramp up from here because all of a sudden we can create enormous amounts of synthetic content. Previously, the main bandwidth was all of this content took a lot of time to manually create. They had entire buildings full of people just churning out fake content, but now they don't really have to. They can generate text, images, even video synthetically and spread it, and there's really no bandwidth limitation whatsoever. And so this is actually something that Eric Schmidt talked about in the video I made a few days ago, where as a country we just have to get better at critical thinking, and other countries and the world in general needs to get better at looking at things, questioning what we're seeing online, and hopefully we can teach in school how to detect fake images, fake text, fake video. And an even better solution will be to use AI to detect AI, and hopefully that's what we're going to

get this week. We identified and took down a cluster of chat GPT accounts that were generating content for a covert Iranian influence operation identified as storm 2035. And there's actually a research report by Microsoft detailing everything that they found, and it's quite interesting. So here's past Iranian election influence efforts: 2020 America, 2021 Israel, and so on. So you can find all of this information in this PDF. I'll drop a link to it in the description below. So it does say that they created commentary on candidates on both sides in the US presidential election, which is interesting. So they don't have favorites; they're just trying to create chaos. Our investigation revealed that this operation used chat GPT for two purposes: generating long-form articles and shorter social media comments. The first workstream produced articles on US politics and global events published on five websites that posed as both progressive and conservative news outlets. The second workstream created short comments in English and Spanish, which were posted on social media, a dozen accounts on X and one on Instagram.

involved in this operation, and again, the accounts on X POS as both progressives and others as conservatives. The operation generated content about several topics, mainly the conflict in Gaza, Israel's Presence at the Olympic Games, and the US presidential election, and to a lesser extent, Venezuela. And here's a couple of examples. Here's one fake article, here's another, another, here's some comments, and even these images, which it says they ran through their Dolly classifier. They found to be fake. So yeah, look at what you're reading online with a very critical eye. Next, in an article on vox.com, people are falling in love with and getting addicted to AI voices. And I have been talking about this for a while. Character AI is incredibly popular, and there are numerous Reddit posts and posts elsewhere that talk about how addicted, especially teenagers, are getting to the conversations they're having with Character AI. And here, even Open AI warns that chatting with an AI voice can breed emotional reliance. So look, if someone's lonely and they start having a conversation with an artificial

intelligence that can easily pass the Turing test. Of course, they're going to become emotionally bonded with that AI, and I don't think this is going to be a good thing in the long run, especially because we have declining birth rates throughout the world. We have an epidemic of loneliness, both of which I've talked about at length in other videos. And when you add AI to the mix, AI companionship, that is, it becomes obvious that people are going to choose that instead of the harder path, which is to go out and meet real people. And that's only going to exacerbate the declining birth rate issue. And especially when we start getting voice mode from ChatGPT, it only increases the likelihood that people will become reliant on that emotional interaction. OpenAI says that the AI's ability to have a naturalistic conversation with the user may heighten the risk of anthropomorphization, attributing human-like traits to a non-human, which could lead people to form a social relationship with the AI. So basically, her and right here, it goes on to say

open AI isn't the only one creating sophisticated AI companions. There's character AI, which young people report becoming so addicted to that they can't do their schoolwork. And Google just introduced Gemini live, which charmed Wall Street Journal columnist Joanna Stern so much that she wrote, I'm not saying I prefer talking to Google's Gemini live over a real human, but I'm not not saying that either. And what I like about this article is they actually talk about the reasons why this is concerning beyond what I just mentioned. So first, chatbots make it seem like they understand us, but they don't, and that's important. As of now, they don't understand us. Their validation, their emotional support, their love, it's all fake, just zeros and ones arranged via statistical rules. Second, there's legitimate concern about entrusting the most vulnerable aspects of ourselves to addictive products that are ultimately controlled by for-profit companies from an industry that has proven itself very good at creating addictive products. They could easily

have just been talking about social media, and it would have made just as much sense here. Finally, there's the worry that people will get addicted to their AI companions at the expense of getting out there and building relationships with real humans. That's probably the one I'm most concerned with, but actually sharing your most intimate secrets with an AI that is going to record it and give it to a for-profit company, that's very concerning as well. So look, I'm not judging at all, but if you're going to use AI as a companion, just know it's not a substitute for a real human. It doesn't really understand you, and anything you say to it probably is being recorded. Next, from OpenAI, fine-tuning is now available for GPT-4. So their most cutting-edge frontier model now allows for fine-tuning, which is fantastic. And I get this question all the time: when would I fine-tune, and when would I use RAG? And nine times out of ten, your use case likely requires RAG, not fine-tuning, if you're trying to

influence the tone of the response, the types of responses that you're going to get, put guard rails on those responses, then use fine-tuning. But for anything else, and especially for giving a large language model additional knowledge, fine-tuning is not the right choice. It is all about RAG, retrieval-augmented generation, and I've made multiple videos about this topic, so check them out on my channel. So GPT-4.0 fine-tuning is available today for all developers on paid tiers. To get started, visit the fine-tuning dashboard. GPT-4.0 fine-tuning training costs $25 per million tokens. So I don't actually know how that compares to other fine-tuning. That seems pretty expensive. Inference is $3.75 per million input tokens and $15 per million output tokens. And as we've talked about before, GPT-4.0 mini fine-tuning is also already available. And interestingly enough, I talked about Cosign Genie, a new model that was announced a week ago. I didn't actually realize it was a fine-tune of GPT-4.0. But listen to this with a

fine-tuned GPT-4 model, Genie, achieved a state-of-the-art score of 43.8 on the new SBench verified benchmark. So, right here, obviously, you can get a lot of great results just by fine-tuning GPT-4. All right, next, Grok 2 large is now available. So, as we could see here, we already had Grok 2 mini in beta. Now we have Grok 2 large in beta, and I already recorded a video testing it out. I'm excited to share it with you. Give me another day or so, and I will publish it. So, all you need is a paid X account, and you get access to Grok 2. And, of course, with Grok 2, you also get all of the amazing images that have been created since its release. Next, LM Studio. One of the easiest ways to run inference locally on any open-source model has a new version coming, and it is a complete redesign. So, LM Studio 0.3.0 is a total UI refresh. I've

actually been testing it out for the past couple of weeks, and the UI refresh is really nice. It's very beautiful, and it makes everything just much easier to use. It also has built-in chat with documents, AKA rag, open AI compatible structured output API, which they just announced. Serve on network, which makes llms available to other devices on the network. So definitely check out LM Studio next, and something that I've been talking about for a long time in a leaked recording. Amazon Cloud Chief tells employees that most developers could stop coding as soon as AI takes over. This is something I truly believe. We're going to have an absolute explosion of developers, billions of developers, and then we're really not going to have any. We're going to have this really cool interim period where anybody's going to be able to code using natural language with the assistance of artificial intelligence. And then once it becomes good enough, there's really not going to be a reason to create applications anymore. You're just going to tell the AI what you want, and it's

going to deliver it to you. There is no need for an application layer at all. This is my own prediction. I could be wrong, but I do strongly believe it. What he said is, If you go forward 24 months from now or some amount of time, I can't exactly predict where it is. It's possible that most developers are not coding, said Garmin, who became AWS's CEO in June. Coding is just kind of like the language that we talk to computers. It's not necessarily the skill in and of itself. The skill in and of itself is like, 'How do I innovate? How do I go build something that's interesting for my end users to use?' So, I actually have a more extreme thought on this. There is no need to build something for your end user. The end user is simply going to talk to the AI directly and get what they need. Let me know what you think about this. I know a lot of you disagree. I love having this conversation. I admittedly could be wrong about it, but I really think this is going to be the future. All right, next.

California, the state in which I live, has been discussing an AI bill that was pretty bad, and apparently, OpenAI thinks so as well. Now, obviously, if OpenAI is supporting some kind of legislation, some kind of bill, I always look at it with a bit of cynicism, but in this case, I kind of agree. And already, I should mention the AI bill has changed. OpenAI is opposing a bill in California that would place new safety requirements on artificial intelligence companies. OpenAI said the bill would hurt innovation and the AI industry and argued that regulation on this issue should come from the federal government instead of the states, which I kind of agree with. The bill specifically is SB 1047, introduced by Scott Wiener, aims to enact what his office has called common-sense safety standards for companies that make large AI models above a specific size and cost threshold. Now, that's already the issue in my mind. If it's above a certain cost and a certain size threshold, both of those things, both of those measures are coming down.

tremendously. So let's think about cost. Although right now, cost seems to be exploding to train a new large model, there could be a new technology that comes out that makes it so it doesn't cost as much, or a new chip innovation. Think of it this way: Sam Altman talked about larger and larger models likely not being the future of AI. We should be able to do more with the existing data that we have, and so having kind of a fixed number, a fixed size, that above it you have to register, or you have some kind of guardrails, doesn't really make sense. And cost is the same; cost could go way up, it could go way down, and it doesn't really signal whether you have something dangerous or not. So the bill would require AI companies to take steps to prevent their models from causing critical harm, such as enabling the development of bioweapons that can cause mass human casualties, or by contributing to more than 500 million in financial damage. It needs to take reasonable care that artificial intelligence models don't cause catastrophe and disclose a

statement of compliance to California's attorney general, and it may face civil penalties if it doesn't. So, it always makes me a little bit nervous when somebody who is not in the AI industry whatsoever tries to regulate AI. It just doesn't make sense to me. And Scott Weiner obviously got advice from different people. I don't know who he got advice from, but in this instance, the way that they proposed the bill just doesn't make sense to me. Next, finally, Mid Journey has a website. I can't believe it took them this long to make it. Going through Discord to create your Mid Journey images is the reason I stopped using Mid Journey about a year ago. I just can't stand that. And now they actually have a really nice-looking website. So, maybe it's time to dust off my Mid Journey account. But I am pretty happy with Dolly and now Flux, but it's a beautiful website nonetheless, and I definitely want to try it out. Have you used it already? Let me know what you think in the comments. Next, Audiogram 2.0.

launches are the most advanced text image model now available to all users for free. So here's what's new. So you can see a bunch of image examples, and let's see what else is available now. Choose from five distinct styles: General, Realistic, Design, 3D, and Anime. And it lets you create images that convincingly resemble photographs with dramatically improved textures. Let's look at a few examples. So this is graffiti with a realistic-looking low rider in the front, and the text looks fantastic. Here's another one of a wizard. Look at those hands. Look at the detail in this bubble. It's so cool. The design style significantly improves text rendering and Audiogram 2.0. It enables you to create premium graphic designs, including greeting cards, t-shirt designs, posters, illustrations with longer and more accurate texts. So this is great if you are trying to create a logo, and you could also specify different palettes to use. Look at that really cool interior design. And you can use ideogram search to explore over a billion public ideogram images for

inspiration it offers industry-leading image generation capabilities. Human evaluations consistently rate ideogram as significantly better than Flux Pro and Dolly 3. It also has an API and an iOS app, so definitely check out ideogram. I plan to as well. Next, there is just so much news this week. Waymo achieves 100,000 weekly autonomous trips. If you're not familiar with Waymo, it is Google's autonomous driving company. They have been all over Los Angeles, all over San Francisco for a while now. I think they're in a couple of other cities as well, but they're running this as a full-scale operation. This is a real business that is only going to grow, and they are hitting 100,000 autonomous rides per week. So, if we do some back-of-the-envelope math, 100,000 autonomous trips per week is 400,000 trips per month, multiplied by let's say an average of $25 per trip, and that's $10 million a month in revenue. So, that's pretty darn

good for just a few cities, so congratulations to wh. I actually haven't taken a whmo ride myself. I don't really live in the more crowded areas of Los Angeles, which is where they are, so I haven't been able to use it. And I haven't been to San Francisco in a little while, but I really want to. It seems really cool. Apparently, they're super safe, but there was that problem about a week ago where late at night in the middle of a parking lot, they all started beeping at each other and waking up everybody who was around them. So, that's pretty funny, and obviously, they're going to have to fix stuff like that. But overall, I think it's really cool. Next, remember we were just talking about AI addiction? Well, character AI ranks number two in a16 Z's gen AI apps. If you're not familiar with a16z, it's Andreessen Horowitz, one of the most renowned venture capitalists in Silicon Valley, and they ranked the top 50 gen AI

consumer web products. Obviously, number one is ChatGPT, but number two is Character AI, and number three is Perplexity, which I just got into. I'm embarrassed to say I know I'm late. I made a video comparing Search GPT to Perplexity to Google Search. Perplexity won, and I just started paying for the pro version, so I'm really excited to use it. But yeah, look at that Character AI, number two. So no wonder they are just absolutely blowing up. They have a ton of usage, and congrats to them. Next, Neuralink's second patient, Alex, successfully played the first-person shooter game Counter-Strike 2 on his laptop using the link. So imagine this, look at this. This is being played through his thoughts, nothing else, no mouse, no keyboard, completely through thoughts. And he'd probably beat me, but I think this is super impressive. All of the movements look realistic. They're not jittery or jumping all over the place, so really, really cool. And you know what? I think this is amazing. Everybody should be able to play Counter-Strike. All right.

just a couple more stories. Jamba 1.5 just dropped. If you're not familiar with Jamba, it is a model built on the Mamba architecture that is different from the Transformers architecture. It's one of the only really capable AIs based on something else than Transformers. So, they have Jambo 1.5 mini, which is 12b and 52b, and Jambo 1.5 large, which is 94b and 398b. If you want to see me test this, you know what I'm going to say, drop a comment so I know you're interested in seeing a test of these models. And here are its scores. So, here's Jambo 1.5 large Arena Hearts 65.4 MMLU Chain of Thought 81.2, and so on. So, you can see all the scores here on the hugging face page. And since it's on hugging face, it is open source, so definitely check it out. And our last story, Meta has quietly launched a new web crawler to go get new training.

data for future models. The crawler named Meta external agent was launched last month. According to three firms that track web scrapers and Bots across the web, the automated bot essentially copies or scrapes all the data that is publicly displayed on websites. Meta says, Like other companies, we train our generative AI models on content that is publicly available online. We recently updated our guidance regarding the best way for Publishers to exclude their domains from being crawled by meta AI-related crawlers. And I like that. As long as they give the option for Publishers and content creators to decide whether they want to be crawled or not, that's all I would ask for. I know a lot of you disagree with my copyright stance, but as long as we have the choice, that's all I care about. And so, if you're a publisher, if you're a content creator, all you need to do is update robots.txt, which is the same way you prevent Google from scraping your website or allow it. And you can allow or deny the open AI crawler.

You can allow or deny meta crawler. So, that's it for the AI news. If you enjoyed this video, please consider giving a like and subscribe. And I'll see you in the next one.

Humanoid Robots, Grok 2 Large, Midjourney Web, Ideogram 2.0, Autonomous Taxis, AI Used For Harm

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 뉴스 하이라이트: 휴머노이드 로봇과 비밀 작전

유니티 로보틱스는 내비게이션에 심층 강화 학습을 사용하는 민첩한 휴머노이드 로봇과 새로운 오프로드 로봇 개를 공개했습니다.

OpenAI는 ChatGPT를 사용해 정치 콘텐츠를 생성하는 이란의 은밀한 영향력 행사를 방해하고 관련 계정을 금지했습니다.

이 작전은 진보와 보수적 관점의 기사와 소셜 미디어 댓글을 게시하여 혼란을 야기하는 것을 목표로 했습니다.

특히 청소년의 경우 AI 목소리에 중독되어 정서적 의존과 잠재적인 사회적 고립을 초래할 수 있습니다.

OpenAI는 AI와의 자연스러운 대화가 의인화와 사회적 관계 형성으로 이어질 수 있다고 경고합니다.

인간의 취약한 측면을 영리 기업에 맡겨 외로움과 출산율 저하를 악화시킬 수 있다는 우려도 있습니다.

이제 유료 티어의 개발자는 100만 토큰당 25달러의 교육 비용으로 GPT-4.

0 미세 조정을 이용할 수 있습니다.

Grok 2 대형은 베타 버전으로, 유료 X 계정으로 해당 기능을 이용할 수 있습니다.

LM Studio는 UI를 완전히 새롭게 디자인하고 오픈 소스 모델에서 로컬로 추론을 실행할 수 있는 기능을 추가한 새 버전을 출시했습니다.

캘리포니아의 AI 법안에 대한 OpenAI의 반대 입장

OpenAI는 AI 기업에 새로운 안전 요건을 부과하는 것을 목표로 하는 캘리포니아의 SB 1047에 반대하며, 혁신을 저해할 수 있으므로 연방 차원에서 규제해야 한다고 주장하고 있습니다.

이 법안은 대규모 AI 모델을 보유한 기업을 대상으로 하며, 심각한 피해를 예방하기 위한 조치를 취하고 법무장관에게 규정 준수 여부를 공개하도록 요구하고 있습니다.

기술과 비용의 급격한 변화 가능성을 고려할 때 규제를 위한 고정된 규모 및 비용 임계값을 도입하는 것은 문제가 있는 것으로 간주됩니다.

아이디어그램은 이미지 생성 분야에서 경쟁사들을 제치고 선두를 달리고 있으며, API와 iOS 앱을 제공하고 있습니다.

구글의 자율 주행 회사인 웨이모는 현재 주간 10만 건의 운행 횟수를 달성하여 월 1,000만 달러의 수익을 창출하고 있습니다.

캐릭터 AI와 퍼플렉시티는 a16z의 차세대 AI 앱 순위에서 2위를 차지했으며, Neuralink의 환자는 생각만으로 카운터스트라이크 2를 성공적으로 플레이하는 데 성공하는 등 주목을 받고 있습니다.

12b에서 398b에 이르는 미니 및 대형 모델을 갖춘 독특한 아키텍처를 갖춘 Jamba 1.

이 모델의 점수는 오픈 소스인 Hugging Face에서 확인할 수 있습니다.

메타는 향후 모델을 위한 학습 데이터를 수집하기 위해 웹 크롤러를 출시하여 퍼블리셔가 자신의 도메인을 크롤링 대상에서 제외할 수 있는 옵션을 제공합니다.