주요 AI 개발 현황 OpenAI, Google, Microsoft 등

### 주요 AI 개발 현황 OpenAI, Google, Microsoft 등
* 🚀 이번 주는 AI 분야에서 기념비적인 한 주였습니다. **OpenAI, Google, Microsoft, 그리고 다양한 AI 아트 및 동영상 제작 업체에서 주요 발표**가 있었습니다.
* 🗣 OpenAI의 개발자 데이에서 샘 알트먼은 **독립적으로 작업을 수행할 수 있는 AI 에이전트가 내년에 출시될 예정**이라고 밝히며, **2025년까지 대대적인 출시**를 암시했습니다.
* 💻 캔버스라는 새로운 ChatGPT 기능이 도입되어 사용자가 편집을 제안하고, 콘텐츠 길이를 조정하고, 문법을 확인할 수 있는 등 UI가 개선되었으며, **이제 모든 ChatGPT 플러스 가입자가 사용할 수 있습니다.**
* 🖼 Microsoft는 사용자 활동을 추적하는 리콜 기능과 AI 기능으로 이미지 상호 작용을 향상시키는 **클릭 투 할 기능** 등 Co-Pilot PC의 새로운 기능을 발표했습니다.
* 📱 Google 렌즈는 사용자가 콘텐츠에 대해 질문할 수 있도록 **동영상 이해 기능**을 확장하고 이미지에서 제품을 식별할 수 있는 **쇼핑 기능**을 도입했습니다.
* 💡 메타는 사용자가 인터넷에 접속하지 않고도 기기에서 AI 모델을 실행할 수 있는 **Llama 3.2**를 출시하여 개인 정보 보호를 강화하고 AI 기능을 크게 향상시켰습니다.
* 📈 특히 OpenAI는 **66억 달러의 자금을 조달**하여 비영리 단체에서 영리 단체로 전환할 가능성을 시사했으며, Microsoft는 **검색 기능에 제너레이티브 AI를 계속 통합**하고 있습니다.
* 🤖 AI 기술이 발전함에 따라 새로운 규정과 윤리적 고려 사항이 점점 더 중요해지고 있으며, 최근 캘리포니아에서는 **AI 딥페이크와 기업의 책임에 관한 법률**이 제정되는 등 그 중요성이 더욱 커지고 있습니다.

내년에 OpenAI에서 출시가 확정된 AI 에이전트는 독립적인 작업 수행에 있어 상당한 발전을 약속합니다.

so it's been an absolutely insane week in the world of AI. We've got announcements out of OpenAI, out of Google, out of Microsoft, out of Meta, out of all of the AI art generators, and all the video generators. Everybody decided to drop new announcements and roll out new features this week, so there is a ton in this video. I'm going to try my best to Rapid Fire it and not ramble too much and share all the crazy AI news that happened with you this week. So let's just get right into it. Let's start right here with OpenAI. This week was AI Dev day for the most part. Most of the announcements were more geared towards developers and not really end users of ChatGPT, but there was still some interesting stuff that came out of Dev day. Sam Altman did a little fireside chat at the end of Dev day and allowed the audience to ask questions, and a lot of the questions were about when are we going to get AI agents. And according to this article over on Tom's Guide.com, it

says open AI confirms AI agents are coming next year. Open AI is on target to launch agents next year. These are independent artificial intelligence models capable of performing a range of tasks without human input and could be available in chat GPT soon. Now, as far as I know, Open AI hasn't actually released the recordings from DevDay yet. However, I did find this YouTube channel called Kyle Kabas. I'm sorry if I mispronounced your name, where he actually filmed the entire sort of fireside chat with Sam Altman. And the closest thing I can find to them saying that agents will be here by 2025 was this little clip here. Maybe talk to us a bit more about how you see agents fitting into OPI's long-term plans. Are you a huge part of the... I mean, I think the exciting thing is this set of models, 01 in particular, and all of its successors are going to be what makes this possible because you finally have the ability to reason, to take hard problems, break them into simpler problems, and act on them. I... I mean, I

think 2025 is going to be the year that this really goes big. Yeah, so they did mention from stage that they think 2025 is going to be the year, and Sam Altman kind of sort of confirmed it. That was the only clip that I can find that seemed to kind of confirm that they think agents by 2025. Again, I don't think the full video is online, but if you do want to watch this entire fireside chat, they did check out Kyle's channel here. I'll make sure it's linked up in the description as well. I'm going to talk more about the dev day announcements in a second, but before I do, let's talk about canvas, which is actually a chat GPT feature that they just rolled out this week as well. And canvas is kind of a complete overhaul of the UI inside of chat GPT. We can see that chat GPT will start suggesting edits, it'll adjust the length to be shorter or longer, change the reading level, add final polish, and check for grammar and clarity, and you

can even ask it to add emojis. They've also added some coding features with this new canvas, like review code, add logs, add comments, fix bugs, or port to a different language. So, you can go from JavaScript to Python or whatever. And according to Sam Altman here over on X, he said that the new canvas feature is now live for 100% of ChatGPT Plus subscribers. So, if you are a paid subscriber to ChatGPT, you should have this feature now. Now, when I pop into my ChatGPT account, it defaults still to ChatGPT 4.0. But if I click this drop-down, you can see we now have an option to switch to GPT 4.0 with canvas. We don't get canvas with the new 0.1 preview mode that thinks through things, but we do have it with the GPT 4.0 model. So, if we click into here, we now have the ability to call on the canvas. If I just give it a prompt like write a short story about wolves learning to use a computer, and

hit enter. You can see that it just completely changed the whole interface, and it put my chat over here on the left sidebar, and it put the story that it just wrote over here in the right window. Now, if you use Claude and you've seen Claude's artifacts, this feels very similar. The biggest difference is in artifacts. You can actually have it generate code, and then you can actually preview what the code does without actually going to another screen. This will just output the code, but you still have to copy and paste the code over somewhere else to actually execute the code. Now, what this did when it created this new right side of Chat GPT is it made it so I can sort of select text here, and when I select text, I can ask Chat GPT about just that paragraph. I can tell it to rewrite just this paragraph, and if I submit it, you'll see that it will actually automatically fix just that paragraph and leave the rest of the document alone. In the past, you would have basically had to tell it to

rewrite the whole thing but just fix the first paragraph and it would have completely done the whole thing over again. If I select this text, I can also format it, you know, make it bold, metallic, change it to headings, things like that. And then over on the right sidebar, you can see this little pencil icon down here. If I hover over it, it opens up a new menu here with suggest edits, adjust the length, reading level, add final polish, and add emojis. So if I click suggest edits, it actually reads through the story that it wrote and then suggests its own edits to its own story. And if you like the edits that it's suggesting, just like in a Google doc or something like that, you click apply, and it actually tweaks it and edits it with the new version that it suggested to itself. I can click on adjust length, and we've got a little slider here to make it super long or super short. So if I go to shortest and just let go, you'll see it'll just rewrite the whole thing, but

in a much shorter version. If I click adjust the length, bring it all the way up to longest, it reworks the entire story, this time making it quite a bit longer. We can change the reading level with a slider as well, all the way down to a kindergarten level, up to a graduate school level. So, if I bring it down to a kindergarten level here, you can actually see that the writing style is more like a kid-friendly writing style. Deep in the forest under the tall trees, the Wolfpack found something strange, etc., etc. If I change the reading level all the way up to a graduate school level, you can see it's much more descriptive. Deep within the dense forest ecosystem beneath the expansive canopy of towering conifers, a Wolfpack encountered an anomalous object. So, it completely changes the reading level. So, if you're trying to get it to explain some sort of complex concept to you, and you're not quite understanding it, you could essentially tell it to dumb it down for you and keep on dumbing it down until you finally get it. You can

add some final polish. You can see it goes through and formats it. We got like a headline and some subheadlines, and it just sort of broke it up and cleaned it up, so it's a little bit easier to read. And then if we click add emojis, it will do exactly what you expect it to do and just fling some emojis in there as well. And uh, yeah, make it look like that. I'm going to go ahead and clear this and go to a new chat window because there's something else that they've recently added that's pretty cool. If I come down to the chat window here, they actually have like quick shortcuts now. If I type slash, you can see I've got reason, search, and picture. So if I select picture, anything I put after this, it will call on Dolly 3 to generate. If I do search, it will make sure that it searched the web before it finishes The Prompt. And if I click on reason, it's going to make sure it uses the new 01 model that really thinks things through. So some pretty

handy new updates to chat GPT this week. All right, jumping back to some of the other stuff they talked about at DevDay. I'm going to kind of go quickly through it because this is more designed for the developers that are using the API, but they did introduce Vision to fine-tuning the API. So developers can now fine-tune GPT-4 with images and text to improve the vision capabilities. So if you're using the OpenAI API and you want to build a tool that uses their Vision models, you can actually upload some of your own images and give it some additional context, and it will actually get better on the specific type of images that you train into it. Again, something more focused towards developers. Another thing focused towards developers is that they introduced the real-time API. They recently rolled out the advanced voice mode where you can have a much more conversational chat with chat GPT. Well, they rolled out the ability to use those conversational Bots inside the API. So now other apps can use that same technology within their app.

and if you want, you can even test this yourself outside of the chat GPT app. You can actually go to the OpenAI Playground here: platform.open.com/playground. And over on the left, they added a real-time box here, and I can actually start a session and have a conversation with the GPT-4 model here using the sort of new advanced voice features, but outside of chat GPT. Hey, hey, how are you doing today? I'm fantastic, thanks for asking. How about you? I'm doing great, just recording a video breaking down all the news for the week in the AI world. That sounds exciting. There's always so much happening in the AI world. Any particularly big stories you're covering this week? Oh, wouldn't you like to know? Fair enough. I'll just have to wait for the video then. They also rolled out model distillation in the API. This lets developers easily use the outputs of frontier models like 01 preview and GPT-4 to fine-tune and improve the performance of more cost-efficient models like GPT-4 mini. They also added

prompt caching into the API. This is something that significantly reduces the cost of using the API. If you're a developer, this is something that Cloud's had for a little bit now, but we're finally getting it in the OpenAI API. And those were pretty much the big announcements that came out of OpenAI's Dev Day. But there's other OpenAI news this week in the fact that they also got new funding to scale OpenAI. Now it's looking more and more likely that OpenAI is going to convert from a nonprofit to a for-profit company. And this past week they managed to raise $6.6 billion in funding at $157 billion post-money valuation. This makes them the third largest startup on the planet. I believe last week we got Meta Connect, and they showed off some of the new Meta Ray-Ban sunglasses. During Meta Connect, they announced a new feature that these glasses were going to be getting memory. So you can look at something and say like, Remember where I parked, or Hey, remind me in 10 minutes to call my mom.

or whatever. Well, those features are rolling out right now into the sunglasses and glasses that you have. If you've got a pair, the new update will also allow them to recognize QR codes and open them on your phone, and make phone calls based on a phone number that is seen in front of the camera. I actually tested these features; they work great. You look at a QR code, and just by telling your glasses to scan this QR code, you can pull out your phone, and it just opens the app on your phone real quick. Pretty handy. The memory feature is really cool too. I actually haven't tried it with an image yet where I take a picture of like a parking spot number, but I have tried it where I told it to remind me to do something in 5 minutes, and then that reminder came through. Pretty handy feature. And while we're on the topic of meta for today's video, I partnered with meta because they just released llama 3.2, and it's a significant leap forward in AI technology. Whether you're a developer or

you're just simply interested in AI. This update is worth your attention. One question I get asked constantly is how can I use an AI model without actually sending my data to like the big corporations. Well, llama 3.2 has a pretty solid answer for you. You can run it directly on your device. In fact, you can use it without even being connected to the internet if you want. So, what's new in llama 3.2? Well, first, the larger models, 11b and 90b, they both now have Vision capabilities. This means that not only can they understand text, but they can understand images now as well. You can ask the AI about a chart in your report or to describe a photo, and it would understand the visual context. But meta actually put out some lighter weight models as well with 1B and 3B text-only models. These are perfect for on-device AI applications and even on mobile phones. Imagine a personal assistant that can summarize your messages or manage your schedule, all while keeping your data on your device.

one of the really cool features about these models is that it supports a 128,000 token context window. That's like being able to put a whole book's worth of information inside of a single conversation. This llama 3.2 is actually optimized for Qualcomm and MediaTek hardware right out of the gate. Now, this is super important for anybody that wants to actually develop AI-powered applications for mobile phones. And one of the most important aspects of these llama models is that they're open source, so you can download these models from llama.com or from Hugging Face and start building immediately. They're compatible with platforms like AWS and Google Cloud and Microsoft Azure and a ton of others. Now, personally, I believe open sourcing these models is really, really important. It encourages innovation and it allows for more diverse applications. This year alone, llama has seen 10x the growth and has become one of the preferred large language models for AI development. And if you're a developer, this means that you have access to the most cutting-edge models that you can modify and adapt to

your specific needs, and well, if you're not a developer, it means that the apps and services that you're going to use could be significantly more intelligent and more helpful. I've also to mention that meta really prioritized safety with this new release. They've introduced new safeguards, including Lam Guard 3, which is designed to ensure that these powerful models are actually used responsibly. So, whether you're aiming to create innovative AI applications or you're simply excited about the future of technology, llama 3.2 is actually really worth exploring. It's powerful, it's really flexible, and it's available for everybody to use and improve upon and iterate on. So, if you're ready to get started, you can go to llama.com or you can go to hugging face and download the models to begin your journey with llama 3.2. The future of AI is open source, and it's here now. Thank you so much to meta for sponsoring this video. All right, moving on to Microsoft now. There was a ton of announcements to

come out of the world of Microsoft, especially if you have one of the new co-pilot plus PCs with the NPU (Neural Processing Unit) built into them. These are pretty much all of the new laptops and computers that are coming out from Microsoft these days, or this new version of the co-pilot PCs. Some of the new features you're going to get are the recall feature. This was a feature that was supposed to roll out when the new co-pilot PCs rolled out, but a lot of security and privacy concerns popped up, and they sort of put it on the back burner for a little bit to fix it and improve some things. And now they're finally rolling out this recall feature, which is essentially like your internet browsing history, but for everything you do on your computer. It remembers you editing videos or writing documents in Word or browsing through your photos, anything that you did throughout the day. It sort of saves it as a history so you can go back to that moment and remember what you were doing on your computer, and

yes, you have the option to turn it on and off, and no, they don't actually send the information that they're collecting back to Microsoft. This is all just on-device. They're also adding this click-to-do feature where if you have an image open on your computer, you can click on it, and you can see it gives the option to visual search with Bing, blur background with photos, erase objects with photos, remove background with paint. So, just by clicking on the image, you get a whole bunch of new sort of AI-related options. It says here it also assists with text-related actions such as rewrite, summarize, or explain text in line, opening in a text editor, sending an email, web searches, and opening websites. Click-to-do is context-aware and accessible from any co-pilot plus PC screen. They're also improving the Windows search with some AI. You can see they did a search up in the top here where they searched Barbecue Party. Notice that all of these images are just titled like image 1123, image 1111. Windows figured out what the context of

these times were and pulled up all the images related to a barbecue party that were on this computer, and it says here it works even when you're not connected to the internet. So, this isn't like an online feature; this is just going to use your laptop's NPU that's built into it. I'm not sure if this works with videos or if it's really only for pictures. I really want this for video; that would help so much with organizing B-roll, but getting it with pictures, I imagine it's only a matter of time before we're getting it with videos as well. They're adding a feature called super resolution inside of photos, so you can open an image inside of photos on Windows and actually upscale the image. They're adding generative fill and erase inside of Microsoft Paint, so you could erase things in the background and generative fill inside the image, just like you can in Adobe Photoshop. But now you can do it in Microsoft Paint as well, so a lot of cool new features that are going to be available on these co-pilot plus PCs.

Microsoft also introduced co-pilot labs and co-pilot Vision. The first feature available in co-pilot Labs is think deeper, which gives co-pilot the ability to reason through more complex problems. It sounds to me like this think deeper is essentially going to use the new OpenAI model that uses that Chain of Thought prompting where it really thinks things through, but it looks like we're going to get that inside of co-pilot Labs. There's also co-pilot Vision. It says if you want it to, it can understand the page you're viewing and answer questions about its content. It can suggest next steps, answer questions, help navigate, whatever it is you want to do, and assist with tasks, all the while you simply speak to it in natural language. And they say it's an entirely opt-in feature, so it only works if you turn it on. But here's a little demo that they put out of what that looks like. They say, Hey co-pilot, I'm looking for a place to stay. They're on

this website stay.com co-pilot starts making recommendations. What do you think of this LOF house? Now in the video, they're speaking back and forth, but there's music on the video. I don't know the copyright status of that music, so I'm not playing the video for that reason. But this is an audio conversation that's happening. The user says, hm, it's a bit pricey. The AI calls that person bougie, and they say, I'm not, I'm just looking for something nice. A little color on the walls, you know. And then the AI says, This one definitely has some color. The user says, Wow, it's giving me a headache. Ahha, we don't want that. Wait, this one looks perfect. Minimal, modern, you. The user says, You're right, I love it. We're booking it. So that's their little demo of what this Microsoft co-pilot Vision looks like. Microsoft also updated the Bing generative search feature. They say, Today we're rolling out an expand of generative search to cover informational queries such as how to effectively run a

one-on-one, and how can I remove background noise from my podcast recordings? Whether you're looking for a detailed explanation, solving a complex problem, or doing deep research, generative AI helps deliver a more profound level of answers that goes beyond surface-level results. To use it, you simply type Bing generative search into the search bar, and you're met with some queries that you can use. There's also a deep search button on the results page, and they do say it might be a bit slow right now. Let's just try Bing generative search, and sure enough, when we test it out here, we get a whole bunch of different potential prompts that we can use. I click on reduce podcasting noise, and you can see we get an AI-generated response here along with a table of contents. Another thing Microsoft is doing is they're starting to pay publishers if their content is surfaced in some of these generative search results. Right now, it looks like it's just big companies like Reuters, Axel Springer magazine, USA Today, and the financial

times I'm not clear if this is going to be something they roll out for smaller content creators 'cause that'd be kind of cool if you write blog posts or make YouTube videos or something like that, and it responds with information that it pulled from smaller creators. It'd be cool for them to get compensated as well. I don't know if that's on the road map or not, though. Right now, it looks like all of the big news media outlets, though, they're trying to work with them so that they can show results from their websites and also pay them when they do. And in the last bit of Microsoft news for the week, the head of Microsoft AI, Mustafa Soloman, here wrote a letter sharing his thoughts on where he thinks all of this is going and what he describes is essentially co-pilot turning into more and more of an agent for you. He says co-pilot will be there for you in your corner, by your side, and always strongly aligned with your interests. It understands the context of

your life while safeguarding your privacy, data, and security. Remembering the details that are most helpful in any situation. It gives you access to a universe of knowledge, simplifying and decluttering the daily barrage of information, and offering support and encouragement when you want it. Over time, it will adapt to your mannerisms and develop capabilities built around your preferences and needs. We are not creating a static tool so much as establishing a dynamic, emerging, and evolving interaction. It will provide you with unwavering support to help you show up the way you really want in your everyday life. A new means of facilitating human connections and accomplishments alike, co-pilot will ultimately be able to act on your behalf, smoothing life's complexities and giving you more time to focus on what matters to you. So what he's describing essentially sounds like an AI agent that's trained on you and what you want to use it for most, which I think is something that most people can probably get behind as long as it's done in a

safe and ethical way without impeding too much on your privacy or sharing too much personal data with the big companies. All right, moving on to Google, 'cause Google had a handful of announcements this week as well. They've been making some updates to the Google Lens tool, a tool where you can upload images and have it sort of search out those images of where they are in the web and give extra information around those images, things like that. Well, now it can actually understand videos as well. We can see in this demo here some fish schooling, and they talk to it, and they say, Why are they swimming together? And it looks at the video, understands what's in the video, and then actually gives an AI response based on what it saw in the video. They're also adding that voice questions feature where you can talk to Google Lens. So they take a picture of the sky here, and then they say, What kind of clouds are these? And then it gives them an AI response, but they asked that question vocally. It wasn't them typing the question. They're

also adding this feature to shop what you see. So, you see a backpack, you take a picture of it, and then it finds where you can actually purchase that backpack online. They're adding the ability to identify songs in their Circle to search, kind of like the Shazam app, where you just hold it open, listen to a song, and then it tells you what the song is. It sounds like that exact feature is going to be rolled out into Android devices. They're also going to organize your search results using AI. And if you're still using Google to do your searches, you're probably going to start to see some of these changes take effect pretty dang soon. But Google makes most of its money through advertising, and so in this new world of AI, they have to figure out how to make money off of the AI responses as well. So, we're going to actually start seeing ads inside of the AI overviews. We can see right here that it showed some sponsored messages. So, somebody searches, How do I get a grass stain out of jeans? and then it gives an

AI response of how to do it, and then when they scroll down a little bit, you can see right below the response, there's some sponsored results for things like Tide pen and Oxy Clean related to their search. And in large language model news this week, we also got a new version of Gemini, Gemini 1.5 Flash 8B. This is a new small large language model that's 50% cheaper, two times higher rate limits, and lower latency on small prompts. This is really for developers that use the API here. And on benchmark tests, it looks like it performs pretty well compared to other models in a similar size. Since we're talking about large language models, Nvidia announced a new large language model this week called NV LMD 72B. And this is an open-source large language model that is also capable of vision tasks. And according to this article, it rivals the leading proprietary models like GPT-40. And if we look at the benchmarks here, we can see that this NVMD 1.72B is actually pretty on par with

GPT 4 Vision model, and in one benchmark, even outperforms GPT 40 and CLA Sonic here, which is pretty impressive given the fact that this is an open-source model and not a closed model like Anthropic or OpenAI or Gemini. Pinterest is rolling out generative AI tools for product imagery to advertisers, pretty much the same thing we've seen in tools like Shopify and Amazon. You can upload an image of your product, and it can remove the background or put it in a different scene, things like that. We've seen this roll out in all sorts of e-commerce platforms at this point. Well, now you're going to get it directly inside of Pinterest. There was some huge news in the world of AI imagery this week. Black Forest Labs released a new model called Flux 1.1 Pro, and they also made their API available. So Flux 1.1 Pro, you can use it right now over on Together AI, Replicate, File.A, and FreePick, and it's quite a bit improved if you've seen

things on like Twitter or X where people are referring to Blue blueberry was sort of the code name for flux 1.1. Here's a little comparison that my buddy angry penguin put together. You should definitely be following him over on Twitter if you're not already. He shares all sorts of cool AI announcements, but we can see the difference here of flux Pro, the old model, find me the stars align, the new model, find me where the stars align. Seems to be much better with text. You can actually see his prompt here if you want to duplicate that. Here's another one, sky the limit, Sky uh schint. Here's another one, sky the limit. So it's much more understanding of what you're looking for, at least from the text side of things. Here's another example. Feel free to pause on any of these if you want to look at them more closely and grab the prompt, but we can see that the text and even the image is quite a bit better. In this image, we can see the barbell is kind of going into

the cat's head, or maybe behind it. I can't really tell this one. You got the whole thing on the screen, and the text is exactly what was asked for. Here's some more examples. Here's another example of a Ghibli-style old Japanese city, blue sky, sunny background, Japanese temple, Japanese traditional. Here's the original one. Here's the new one. Looks quite a bit better. A lot better color palette, in my opinion. Obviously, what's aesthetically pleasing is very subjective, but I find this to be a little bit more aesthetically pleasing. Here's another example. We can just really see how prompted here it is, 'cause look at how big this prompt is. A vector illustration, a group of adorable smiling ghosts wearing different color witch hats. We're getting that each ghost has a unique expression, cheerful pumpkins with carved faces, background should be dark purple. It pretty much nailed every single element from this long prompt here. Here's another example. In the first one, it didn't even get the eye lift to eat. In the second one, it kind of nailed.

It may be one of these straws is meant to be the eye. I don't know a handwritten letter written in Old English and signed flux pro at the bottom. Look at that. Here's another example and another example and one final example. So thanks again to angry penguin for sharing all of those with me and giving me permission to share them in this video. Feel free to go back and pause on any of those if you want to see the specific prompt that I didn't cover. Angry penguin also gave me a quick tip. He said that you can actually use flux Pro 1.1 right now for free if you use it over on the glyph doapp website. Some of the other sites that are allowing you to use it for free are only allowing you to use it for free for like a day. This one it's free for now and from what I understand it'll be free for a few weeks but I don't know how long it's going to be free for but at the time this video is going live you can actually play

around with flux Pro 1.1 for free on glyph. So let's go ahead and sign in real quick. Let's just try a monkey holding a sign that says subscribe to Matt Wolf. It got the two Matt Wolf part but kind of missed the Subscribe. Let's run it one more time. This time it got it right with no issues. I actually love making images over here on glyph because they show up in this feed down here. And now anybody that goes and looks at this feed is going to have a monkey telling them to subscribe to Matt Wolf. But that's flux Pro 1.1, and I will link up where you can use that below. You can use Angry Penguins integration of it, or you can go build your own glyph with it in the workflow. We also got some updates out of Leonardo AI this week. Now this is a company that I am an adviser for, so just keep that in mind when I talk about them. But when they do really good stuff, I talk about the good stuff when they do.

stuff that I don't really like. I point out the stuff that I don't really like, so I try to remain fairly unbiased. But this is just the news. This week, they rolled out a new style reference feature. You can upload up to four reference images to direct the aesthetics of your image output. You can also adjust the strength of the reference image. They also rolled out a new image-to-image feature using the Phoenix preset. They've had image-to-image in Leonardo for a while, but the feature wasn't available to use with the Phoenix model, which is the model that's probably the best inside of Leonardo. But now, image-to-image is available using the Phoenix model. So, if I jump in here, I go to image creation up here in my prompt box. You can see a new little image icon. If I click on this, it gives me the option for style reference or image-to-image or a content reference, which is coming soon. But even if you don't have a style reference already, what's really cool is so I can click on style reference here.

go to the Community Feed, and if there's an aesthetic of an image that I really like, I can pull in that same aesthetic for the images that I'm about to generate. So, let's say I really like this painterly look here. Let's go ahead and pull that in, confirm it, and it's going to use that as a style reference. And I'll just put a simple prompt: a robot looking into the camera. And then hopefully, it will do it in a similar style. And there we go, it gave me four generations. This one's probably the best looking, but you can see it looks like a painted image that models the style that we have up here, but with a robot looking into the camera. Also, a sort of newer feature they rolled out is under the generation mode. They have this Ultra mode, and what that's actually doing is it's upscaling all of these images right as they generate. So, if I actually look at this image at full size, you can see it's a fairly large image. It's actually been upscaled right within the pipeline.

of generating the image. That's pretty cool as well. There's also lots of new features rolling out for Leonardo soon, but as they roll out, I'll show them off. I'm not quite allowed to talk about them yet, but there is some exciting stuff. I'll share that as it rolls out. Adobe rolled out some new AI features inside of their Photoshop Elements and Premier Elements products. These are sort of stripped-down versions of Photoshop and Premiere that don't have all of the features, but they're for more like casual users. You've got like object removal, new AI color correction features, depth of field simulation, and a handful of other smaller AI-related features that were in the bigger platforms, but now the sort of more casual elements version of these platforms are getting these AI features as well. Lum's Dream Machine, which is one of the more popular AI video generation models, got an upgrade this week. They now have hyper-fast video generation, their 10x faster inference, so you can now generate a full-quality Dream Machine clip in under 20

seconds Pika made a bunch of waves this week with their new Pika 1.5 model. But most of what we've seen from this 1.5 model have been more of these types of videos where there's an object and you can see the object getting squished or here's some of my Generations where it shows me sitting there and I get blown up and then float away like a balloon or getting crushed by a hydraulic press or getting exploded. We've been seeing a lot of these types of videos, but it also seems like it should be able to do text to video because all of these are like text to video Generations or possibly image to video Generations. But for me, for whatever reason, I have not gotten text to video to work. I actually starting to generate like a monkey on roller skates and a wolf howling at the moon, and these have been trying to generate for about 36 hours now at this point. I'm not confident they're ever going to actually generate, but these

like meme-type videos where you can squish yourself or cut something open like a piece of cake. Those all work perfectly. I just got to figure out how to get the text image to actually work 'cause that's not working for me anymore for some reason. But the videos they did show off, they look pretty dang impressive, possibly cherry-picked. I mean, most of the time when you're going to see stuff on social media, it's going to be cherry-picked stuff. It looks really, really cool, so I'm excited to actually be able to generate with text-to-video. Just still haven't really quite gotten it to work yet. B dance, the company behind TikTok, also revealed a new AI video generator this week that is said to rival Sora. That seems to be the benchmark that everybody compares the video generator models against. It's a model that none of us have actually gotten our hands on yet, but here's some examples of what it can do. Here's a woman taking off her sunglasses, standing up, and then walking away from the camera.

It looks pretty good. I mean, you can tell it's AI-generated, but it looks pretty good, and it's generating at 10 seconds as well. Here's another one of a man bowing down to a woman here and then looking back up at her, and then she's crying. That one looks pretty good too, but it sort of feels like it's in slow motion still. Here's another example of a black-and-white video zooming in on a woman's face who's wearing sunglasses. I mean, they're looking pretty good. My buddy Tim over here at Theoretically Media actually did a breakdown video all about this new model that's about 9 minutes. I'll link that up below if you want to sort of take a deeper dive into this model. Here's something that's pretty cool that's coming to Steam. So, if you're a gamer and you have the Steam game engine on your computer, there's this new thing called Dream World coming out where you can create any 3D asset and just drop it.

into the world that you're playing in. Here's the demo they put out around that. They type in giant King Kong, and then a giant 3D King Kong is now just like in that world. Black and gold Anubis statue. I think I mispronounced that, but you can see it put that big statue there. Anything that they can imagine, they can just drop into the world. Now, this looks like it's a sort of bigger game where like open world and challenges and things that you actually do. But then, like, one of the sort of cool interesting things that make this game novel is that you can just think of anything and then drop it into your world. Whether or not you can actually use those things, I'm not sure. Like, if you make a boat on the ocean, can I then jump into that boat and sail across the ocean? If I, you know, generate a car in this world, could I jump in the car and drive it around? I don't know. It seems like you just kind of can create the stuff, and it's just there.

and added to your world, and you got like all the crafting and open-world and sort of elements you get out of like a Valheim or a Minecraft or something like that, but with also the ability to just sort of drop 3D objects anywhere in the world. I don't know. I'll probably grab it and play around with it once it's available, though. We also got the news this week that the governor of California, Gavin Newsom, vetoed SB 1047. I've sort of talked about that one enough, but it was the bill that would hold responsible the AI companies that made the model if somebody else took that model and did something that caused catastrophic harm. So if somebody took the Llama model, tweaked with it, and then figured out how to make a chemical weapon that had catastrophic impact, Llama would be held responsible, as well as the person who made the actual chemical weapons, and all of the AI companies were fighting against this bill because they were basically saying we just want to make better and better.

models. We don't know what people are going to use these models for in the future, and Gavin Nome vetoed the bill. Most likely, some regulation is going to come around this stuff, just not that bill specifically. It's only a matter of time before another one gets drafted up and goes through Congress, and hopefully, it's something that more people can agree upon, I guess. But while we're speaking of AI legislation, a judge actually blocked another AI bill that was related to deep fakes, the AB 2839 bill, which was signed by Governor Nome. It was slapped down by the courts. AB 2839 targets the distributors of AI deep fakes on social media, specifically if their post resembles a political candidate and the poster knows it's a fake that may confuse voters. The law is unique because it does not go after the platforms on which AI deep fakes appear, but rather those who spread them. The judge basically said that this goes against the freedom of speech and that the only thing that's going to stick from this bill is that if you are going

to spread a deep fake message with political figures, you have to say that it was generated with AI. You could still spread them and share them; you just have to say that they were made with AI without trying to fake people. That's the only part of the bill that's stuck. Everything else from the bill, they basically said no, that's against freedom of speech. Amazon's rolling out some new fire tablets that are going to have AI tools built into them. Things like writing assistance, getting webpage summaries, and creating wallpapers from a prompt. I mean, I'm pretty sure at this point, like every tablet that comes out, no matter who makes it, is going to start rolling out with AI features. It's kind of become like a necessity or like an expectation of these devices these days. But we're getting it inside the fire tablets. And then finally, I want to end with this one because I thought this was sort of one of the cooler things I saw this week, which is a robust ladder climbing quadrupedal robot. We can see in

this video. They actually created one of these four-legged robots and designed it so that it can now climb ladders. Right now, that it can only climb up ladders. It can't climb down ladders. But if we look at the robot itself, what really makes it unique is sort of the claw-like hand that it's got so that it can grip over ladders and climb them. And the idea being we can send robots into very high dangerous places up ladders that we would normally send humans and for safety reasons. Wherever it seems smarter to send a robot, they would do that over putting a human's life at stake. So I think that's pretty cool. We can see here the sort of digital twin world where they're all being trained on these ladders. And I just love robots. Robots are really, really fun. Whenever I come across new robots doing novel things that I haven't ever seen robots do, I'm probably going to talk about it 'cause yeah, robots are just cool. Hopefully, they don't rise up and destroy us all once.

They get smarter and smarter, AI, but let's not think about that right now. Oh, and one more thing before I wrap up here. I am going to be helping to judge an AI hackathon up in Santa Monica on October 12th and 13th. It should be pretty cool because you don't actually have to be a developer yourself to participate in this hackathon. You can actually use AI to help you code, or you can be someone that actually knows code, and we'll just see who comes up with the coolest product at the end of it. It should be pretty fun. It's happening again October 12th and 13th. You can go to hack.cerebralbeach.com to learn more. Again, I'll be one of the judges. I'll be there, so it'll be fun to meet some people in person there. And that's what I got for you today. If you haven't already, check out Futur Tools. This is where I share all of the coolest AI tools I come across. I share all of the AI news that I come

across, and I have a free newsletter where I'll share just the coolest tools and coolest news that I think you need to know about directly to your inbox. It's all free. You can find it at future tools. Thank you so much for tuning in. If you want to stay on The Cutting Edge and stay looped in with AI and the latest AI tutorials and how they're doing the stuff and the latest AI news and all that kind of stuff, like this video and subscribe to this Channel, and I will make sure more of that stuff keeps showing up in your YouTube feed. I'll try to help you stay on The Cutting Edge of everything that's happening in this world. Thank you once again for tuning in and hanging out with me. I know this video is a little bit long; there was a lot that happened this week, but I appreciate you sticking with it and hanging out with me. Thank you once again to Meta for sponsoring this video. I'll see you guys in the next one. Bye-bye.

AI News: A TON Happened This Week - Here's What You Missed

AI, No-Code, Tech, Futurism - I'm a tech nerd and talk about tech nerd stuff

Matt Wolfe

이번 주에 발표되는 흥미로운 AI 소식

OpenAI의 개발자 데이에서는 내년에 출시될 AI 에이전트에 대한 발표가 있었고, 샘 알트먼은 2025년을 그 영향력의 중추적인 해로 제시했습니다.

ChatGPT의 새로운 캔버스 기능은 UI를 개편하여 편집 제안, 코딩 기능 및 특정 단락을 다시 작성할 수 있는 기능을 제공합니다.

캔버스 기능은 이제 모든 ChatGPT 플러스 가입자에게 제공되며, 텍스트 선택 및 편집 옵션으로 사용자 경험을 향상시킵니다.

OpenAI는 텍스트 길이 및 읽기 수준 조절, 이미지 생성 및 웹 검색을 위한 빠른 단축키, 대화형 봇을 위한 실시간 API 등 새로운 기능을 ChatGPT에 도입했습니다.

또한 66억 달러의 자금을 유치하며 영리 모델로 나아가고 있습니다.

한편, Meta의 새로운 Ray-Ban 선글라스에는 메모리 및 QR코드 인식 기능이 탑재되어 편리한 알림과 앱 열기 기능을 제공합니다.

2와 마이크로소프트의 새로운 코파일럿 PC

2는 대형 모델에는 비전 기능을, 모바일 애플리케이션에는 더 가벼운 텍스트 전용 모델과 함께 온디바이스 AI 모델 사용을 제공합니다.

이 업데이트에는 128,000개의 토큰 컨텍스트 창이 포함되어 있으며, 오픈 소스 액세스와 혁신을 촉진하는 Qualcomm 및 MediaTek 하드웨어에 최적화되어 있습니다.

Microsoft의 새로운 코파일럿 PC에는 내장된 신경 처리 장치로 구동되는 리콜 기록, AI 관련 작업을 위한 클릭 투 할 옵션, 향상된 Windows 검색과 같은 기능이 도입되었습니다.

Microsoft의 새로운 코파일럿 기능 및 Google 렌즈 업데이트

Microsoft는 복잡한 문제 해결을 위한 '더 깊이 생각하기' 및 웹 페이지 탐색을 위한 자연어 지원과 같은 기능을 갖춘 Co-Pilot Labs 및 Co-Pilot Vision을 도입했습니다.

또한 더 깊고 유익한 답변을 제공하기 위해 Bing 생성 검색 기능이 업데이트되었으며, Microsoft는 검색 결과에 사용되는 콘텐츠에 대해 게시자에게 비용을 지불하고 있습니다.

한편, Google 렌즈는 동영상 이해 기능을 확장하고 더 많은 대화형 사용자 참여를 위해 음성 질문 기능을 추가하고 있습니다.

AI 및 온라인 쇼핑의 새로운 기능 및 업데이트

새로운 기능을 통해 사용자는 백팩과 같은 상품을 사진으로 찍어 온라인에서 구매할 수 있게 됩니다.

Google은 AI 기반 검색 결과 구성을 출시하고 AI 응답 내에 광고를 표시하기 시작할 예정입니다.

또한, 오픈 소스 비전 작업 기능을 제공하는 Gemini 1.

5 Flash 8B와 NV LMD 72B를 포함한 새로운 대형 언어 모델을 출시할 예정입니다.

또한, Pinterest는 제품 이미지를 위한 제너레이티브 AI 도구를 도입하고 있으며, Black Forest Labs는 개선된 AI 이미지 모델인 Flux 1.

1은 현재 글리프 도앱 웹사이트에서 무료로 제공되며, 몇 주 동안 사용해 볼 수 있습니다.

레오나르도 AI는 스타일 참조 이미지 업로드와 피닉스 모델을 사용한 이미지 간 생성 등 새로운 기능을 출시했습니다.

Adobe는 포토샵과 프리미어 엘리먼트 제품에 AI 기능을 도입했으며, Lum의 드림 머신은 초고속 동영상 생성 기능으로 업그레이드했습니다.

Pika의 새로운 모델과 AI 동영상 생성기

5 모델은 인상적인 오브젝트 스퀴징과 밈 형식의 동영상으로 주목을 받았지만, 텍스트에서 동영상으로 생성하는 데는 문제가 있었습니다.

틱톡의 개발사인 B Dance는 소라에 대항하는 새로운 AI 동영상 생성기를 공개하며 인상적인 AI 생성 동영상을 선보였습니다.

또한 게이머가 3D 에셋을 만들어 게임 세계에 배치할 수 있는 Dream World가 Steam에 출시되고, 캘리포니아 주지사 개빈 뉴섬은 AI 기업의 유해한 모델 사용에 대한 책임을 묻는 법안에 거부권을 행사했습니다.

개빈 놈은 인공지능 모델과 관련된 법안에 거부권을 행사했고, 인공지능 딥페이크 법안은 언론의 자유를 침해한다는 이유로 판사에 의해 저지당했습니다.

아마존은 AI 도구가 탑재된 파이어 태블릿을 출시하고, 위험한 작업을 수행하기 위해 사다리를 오를 수 있도록 설계된 새로운 4족 보행 로봇을 출시했습니다.

10월 12일과 13일 산타모니카에서 열리는 AI 해커톤의 심사위원을 맡고 있으며, AI 도구 및 뉴스를 위한 무료 뉴스레터를 홍보하고 있습니다.