### 주간 AI 업데이트: 음성 기능, 새로운 모델 및 혁신

### AI 주간 업데이트: 플랫폼 전반의 새로운 기능 및 개선 사항
* 🗣 OpenAI는 모바일 및 데스크톱 앱에 이어 웹의 유료 사용자에게도 **고급 음성 모드**를 출시했으며, 곧 **무료 사용자에게도 확대**할 계획입니다.
* 🆕 GPT-4의 업데이트는 **창의적인 글쓰기 기능을 향상**시켜 더욱 자연스럽고 매력적인 글쓰기가 가능해졌으며, **업로드된 파일에 대한 성능도 개선**되어 더 나은 인사이트를 얻을 수 있게 되었습니다.
* 📸 흥미롭게도 최신 베타 버전에서 **라이브 카메라 기능과 시각 인식 코드**가 발견되어 **실시간 처리와 음성 모드를 통합할 수 있는 향후 기능**을 암시하고 있습니다.
* 📄 Anthropic은 사용자가 **Google 드라이브에서 직접 문서를 추가**할 수 있는 기능을 Claude에 도입하여 사용자 경험과 문서 관리를 개선했습니다.
* 💾 이제 구글의 제미니에는 메모리 기능이 포함되어 **사용자의 선호도와 맥락을 기억**하여 저장된 정보를 기반으로 맞춤형 응답을 제공함으로써 상호작용을 개선할 수 있습니다.
* 🌍 YouTube는 동영상을 여러 언어로 자동 더빙하는 기능을 출시하여 **크리에이터가 추가 노력 없이도 콘텐츠 도달 범위를 크게 넓힐 수 있도록 지원**합니다.
* 📊 DeepSeek R1 Light라는 새로운 모델은 **로직 기반 작업에서 유망한 결과**를 보여주며 **OpenAI의 GPT-1과 경쟁**하고 있으며, **Mistral의 Le Chat은 비슷한 기능을 무료**로 제공하여 예산에 민감한 사용자들에게 어필하고 있습니다.
* 📅 Microsoft는 **하퍼콜린스와 책에 대한 AI 교육을 위한 계약, 컴퓨터 활동을 추적하는 리콜 기능, Teams 모임을 위한 새로운 음성 복제 기능** 등 사용자 상호 작용과 생산성을 더욱 향상시키는 여러 가지 발표를 했습니다.

웹 사용자에게 OpenAI의 고급 음성 모드가 출시되어 ChatGPT의 기능이 향상되었습니다.

As with most weeks lately, it has been a really busy week in the world of AI, and I don't want to waste your time, so let's go ahead and break it all down right now. [Music] Starting with the fact that OpenAI is finally bringing their advanced voice

mode to the web. Now we've already had advanced voice mode available on the mobile app for a little bit now. They recently rolled it out on the desktop app. Well, now it is going to be available inside of the browser app. This is based on an announcement from Kevin Weil here

on X. He says, "Rolling out to ChatGPT paid users this week: advanced voice mode on the web." Down here at the bottom of his post, he says, "Give it a try at chatgpt.com if you're a Plus, Enterprise Teams, or Edge user subscriber. We'll look to

roll it out to free users in the coming weeks. Let me know what you think." That seems to imply that it's already in there, but when I head over to my own ChatGPT account, I'm not seeing any options here to use the new advanced

voice mode yet. So it seems to be a roll out. Maybe some people have it already and it is still rolling out, but I don't have it in my browser account yet to show off. But something that did definitely roll out this week was a new

update to GPT-4. According to OpenAI, the model's creative writing ability has leveled up. It's more natural, engaging, and it's tailored writing to improve relevance and readability. It also apparently works better with uploaded files, providing deeper insights and more

thorough responses again. If you head over to ChatGPT, if you're using the GPT-4 model, you'll be using the latest version of the 4 model. I've played around with it a little bit. I find the updates to be very subtle, but I also use

Claude a little bit more than I use ChatGPT, so I'm not sure I would totally know the difference at the moment. Since we're talking about ChatGPT, remember back when they showed off the advanced voice feature in that demo video? Well, they

showed it off with the ability to actually get live video and have ChatGPT actually see what you're looking at to be able to inform its responses. Well, that feature never really rolled out with the advanced voice mode like we thought it would. Well, it looks like it's going

to be coming very soon. I found this out from Rowan Chong here, who spotted it from Android Authority, but it looks like the code for those features was actually discovered in the latest beta version of GPT. They discovered in the code live camera functionality, real-time

processing, voice mode integration, and visual recognition capabilities. I'll make sure this tweet thread is linked up in the description below, but you can actually see screenshots of the code here that make it look like it's going to be coming pretty soon. Moving on to Anthropic news, they just rolled out a

new feature. If you use Claude, you can now directly add documents from Google Drive. If I head on over to my Claude account here, you can see there's a new little Google Drive button here. If I click this button here, it shows me my recent files.

inside of Google Drive and I can also do a search for specific files in Google Drive and whatever file you pull in will work the same way as if you were to just drag and drop a file into Google Cloud. Gemini got an update this week, it can

now actually remember things, so ChatGPT has had a memory feature for a little while now. Now Google's Gemini has it as well. When you're logged into Gemini here, you can actually come down to your little settings cogwheel and you'll see a button that says saved info. If you

click in there, you can actually see all of the info that's saved. You can add info or you can ask Gemini to remember something while you're actually chatting with it. At the moment it says this feature is only available in English, but here are some examples that Google gave

of the types of things you might want to save. Use simple language and avoid jargon. I'm a vegetarian, so don't suggest recipes with meat. After responding, include a Spanish translation when trip planning. Include the cost per day. I can only write code in JavaScript. I prefer

short, concise responses. You can also give it info about yourself like your name and your career and where you live and things like that so it can be better informed when it's actually giving you its responses. And since we're talking about Google, I want to point out

something really cool that YouTube is doing. I've mentioned in the past but it looks like it's actually finally starting to roll out. They're now going to start doing automatic dubbing on videos into multiple languages, and the YouTube creators don't have to do anything extra. So for someone like me,

all I need to do is upload my video to YouTube like normal and it will automatically create a translated version of it in Spanish, Portuguese, German, French, Italian, Hindi, Indonesian, and Japanese. That's super super exciting to me and it should be for any content creator that's publishing on YouTube.

because it means like every single piece of content you make is going to be available to an audience that's like 10 times bigger than what you're already getting. That's going to be huge. There's a new model out from a company called DeepSeek called DeepSeek R1 Light.

preview, which is a model that's going head-to-head with OpenAI's GPT-1. Now if you remember the OpenAI GPT-1 model is the model that sort of thinks things through before it gives you a response and is much better at sort of logic and

questions that require some detail thought behind them. And up until now, the OpenAI GPT-1 model was really the only model out there that had that sort of focus. Well now there's a model out of this DeepSeek Company from China that can seemingly act in the same way, and we

can see on these benchmarks here that in things like math and code, it actually outperforms the GPT-1 preview model. Now keep in mind this GPT-1 preview model isn't the full GPT-1 model. It's like a checkpoint of the model. We actually

haven't seen the full OpenAI GPT-1 model. In fact, rumor has it it's going to be coming out within the next few weeks, but again that's just a rumor. But this DeepSeek model that is beating GPT-1 preview in a few of these benchmarks here, it's not

really a comparison head-to-head with the full 01 model that we're going to be seeing soon now the French AI company Mistral also has some updates to their chatbot called Le Chat It now has search capabilities, vision capabilities, ideation and coding, and you can use it

all for free so if I come up here to Le Chat we'll go ahead and sign in and very similarly to ChatGPT we've got the option to use Canvas, we've got the option to search the web and we've got the option to generate images.  Theat now

has some new features: web search with citations, Canvas for ideation and inline editing, state-of-the-art document and image understanding, image generation powered by BlackForce Labs Flux Pro. And we got a little breakdown here comparing ChatGPT, Perplexity, Claude and showing that most of the features that are paid

in these other platforms are actually free here in Le Chat. Give me at least 10 YouTube channels to follow to stay up to date with AI. We've got The AI Coach, Two Minute Papers, Some Bum Conrad Cording, Deep Learning AI, Robert Miles AI, Explain AI, Yannic Kilcher, Vsauce, Valov, and

AI in Games. Some of these are actually channels I hadn't heard of yet so more for me to check out. If you're on a budget and you don't want to spend any money on a chatbot like ChatGPT or Claude but you want a lot of the same features

give Mistral's Le Chat version a try cuz it does a lot of the same stuff that the paid versions of those do, but for free. And just for fun let's see how it does with image generation. Generate an image of a cat eating a taco. And that's what

We get it took about 18 seconds to generate, and again, it's using Flux One Pro, which if you've used the AI generator over in Growth, it's the same one that that one uses, but again, it's free to use tools like ChatGPT and Claude have

been absolutely amazing for businesses for brainstorming ideas for marketing and branding and business strategy, but it's not always super obvious what prompts are going to give you the absolute best results for the goals that you have in your business, and that's why for this video I partnered with HubSpot

to share their AI prompt library. Now, this prompt library was custom curated by the CMO and the SVP of HubSpot, Kip and Kieran, also the host of Marketing Against the Grain. Now, as somebody who has a podcast on the HubSpot Network, I

know these guys personally, and they know their stuff. They know how to craft amazing prompts around things like marketing and sales strategy, branding and advertising, business strategy and analysis, pitching and investing, website strategy, and so much more. Want to write headlines like David Ogilvy or sell your

products using a proven storytelling framework? All of these prompts are available to you completely for free, and they're designed so you don't need to be a prompt engineer to get amazing results from these AI tools. I've gone ahead and put the link to this 100% free resource

down in the description below. You're going to want to check this out, and thank you so much to HubSpot for sponsoring this video. There was a lot of news out of Microsoft this week because Microsoft held their annual Ignite event, an event where they typically make a ton

of announcements, I actually don't know if this announcement came out of the Ignite event, but Microsoft signed a deal with HarperCollins, which is a book publishing company, to essentially get permission to train AI models on the books from HarperCollins. Now HarperCollins is going to get

permission from each individual author, as it sounds like. But with all the lawsuits that have been happening over news sources and authors and people like that actually suing these AI companies, it's starting to sound like the big companies are starting to ask for permission a little bit more frequently.

Now, if you remember back at Microsoft Build earlier this year, they announced a feature called Recall, which was like your web history but for your entire computer. So you can go back and remember what you were doing in certain apps throughout the day or you know, several

days ago. Maybe you were doing something in Photoshop, but you wanted to go back to that exact moment. You can actually go back in history on your computer. Well, there was a lot of concerns when they announced that. They went back to the drawing board and really, really dialed

in like the security and just overall functionality of it to make people a little more comfortable with it. Well, as of today, November 22nd, co-pilot PCs with the Snapdragon chip in them are now going to have the ability to access this Recall feature, finally. Now, in order to

access this, you need to have the Windows 11 Insider Preview build and of course, you need to be using one of the newer co-pilot PCs. Here's what the new feature looks like. You have this little scroll window across the top of your Recall page here and you can scroll back to any

moment you can see you've got today, October 21st, October 20th, October 19th so you can go back to any day in time and remember what you were doing back on that day and actually get screenshots of that moment that you're

looking for. You can also type to search your history as well. There's also a little icon that will pop up down in your system tray here that will allow you to pause the recording. So if there's websites you're looking at that for whatever reason you don't want Recall to

remember that you're looking at that website, you can always pause it and make sure it's not remembering that specific moment that you're looking. And if you do happen to record moments on your computer that you wish weren't recorded, maybe you were doing some gift shopping

or something like that, you can go back through and actually delete those moments from your Recall history so that they're completely removed. There's also security measures in place so to even open the Recall window and see what's in there, it's going to double check and

make sure it's you through whatever security measures you have set up on your PC. And if you really, really don't want to use it, you can turn it off completely and not use Recall at all on your computer. So if you're concerned about buying a co-pilot plus PC and

you're considering not getting one cuz you don't like the Recall feature, it is possible to disable it completely. There's also a new feature called Click To-Do where you can click on an image and it'll give you options like visualize, search with Bing, blur the background of

The photo erase objects in the photo, remove the background with paint and things like that, so when you click on an image, you have all sorts of new AI related options. And again, these features roll out today. This video is going live November 22nd, so if you already

have a co-pilot PC with one of the Snapdragon 8 or better chips in it, you should be able to upgrade to the new Insider preview build of Windows and start using these features. Another thing Microsoft announced this week is that it's going to let you clone your voice

on Teams meetings. This is really cool for if you're speaking to somebody in a different language. You'll be able to speak in your native language and it will actually translate it and localize it to whoever you're talking to, and it'll still sound like your

voice. Review the progress of the project. Can you please tell us how the tasks we should take are going? So I'm super excited to see that come out inside of Microsoft Teams. I would love to see that feature roll out into things like Zoom and Google Meet as well soon. 11 Labs

rolled out a new feature this week to build conversational AI agents. So if I log into my 11 Labs account, you can, there's a new button over here called conversational AI. Let's go ahead and try a call and see what it sounds like. I'll

use the video game character here. Let's just see how this goes. Greetings traveler, how can I assist you on your quest? What is your name? Greetings traveler, I am Callum, a wizard here to help you on your quest to collect the 11 gemstones scattered

across Labs onia I can help you find them but I can't do it all for you what can I do to help you today what are these gems for ah the 11 gemstones they are said to hold immense power capable of granting wishes and even altering the very fabric

of reality Legend has it that they were scattered okay that's fascinating let's go ahead and try this myself now so I actually just told it to create that character for me and you can see it dumped it into this AI agents list over here and I can see what it looks like to

actually build an agent now so the agent language we want English we can have it set on the very first message when somebody starts talking to it there's a system prompt with the details wizard name Etc which large language model do you want to use and it's got quite a bit

of options Gemini 1.5 flash is the default cuz apparently it's the fastest model in here but you've also got GPT 4 clae 3.5 Sonet and even a custom llm which looks like you can set up through

using an API key but I'm not going to go there for this one let's just leave it on the default you can set the temperature and then it's got a base so we can add data to this knowledge base so if I wanted to make a AI newsbot I

could take the transcriptions from all of my old videos throw them in this knowledge base and people can ask questions about all of my old videos you can add tools which apparently uses web hooks here you can add Secrets okay so this is probably for adding you know API

keys and things like that I'm guessing you can select the voice it used Callum but I have my own Matt Wolfe voice that I trained in 11 Labs so I can make it sound like me if I want We've got the normal settings you get from 11 Labs all

sorts of really cool features to go ahead and build out a bot that has the information and the voice of whatever you want it to be So that's that's kind of cool and fun I can see the main use case probably being like customer support bots on websites and things like

that but this is something I definitely want to do a deeper dive in and try to build some characters with later So maybe that'll be a video we circle around to in the future 11 Labs also rolled out the ability to auto-regenerate so if you're using their long

form text editor to generate audio it looks like if something doesn't quite sound right you can highlight that selection and tell it to regenerate it and it will do it again and ideally it gets the sort of tone and inflections right the second time around Soundraw got a

huge upgrade this week as well with their new V4 and I actually got early access to this and was able to play with it a little bit before they released it and honestly it is pretty dang impressive It actually generates music that I want to listen to and a few of

these songs I've actually gone back and listened to multiple times One of the things I was having fun doing and I hope I don't get in trouble for this but I was taking lyrics from existing songs that I like and changing the style like "Damn It" from Blink 182 This one I told to

make it an electronic poppy dance version and it reminded me of Owl City, so that's why I named it that. But listen to this. It's all right to tell me what you think about me. I won't try to argue or hold it against you. I know that you're

leaving. You must have your reasons. The season is calling. Your pictures are falling [Music] down. I actually think it's good. I actually like listening to that version of this song, and then I did it like a poppy dub style, and then I did it like a

country version. Like it's all right, tell me what you think about me. Here's why I subscribed to Matt Wolfe's song. That's a punk [Music] version. Tech whispers through the digital trees. AI news brings you to your knees. Finally, this was actually a song

that I did with the version 3. But if you do a song with Sona version 3, and you want to sort of bring that song up to date with the new version 4, you can go to your old songs, click on these little three dots here, click on "Create Over"

here, and there's actually a remaster, and it'll convert the song into a V4 song and actually make it even more impressive sounding. So Sona V4 has been really, really fun to play with. We got a pretty big update in the world of AI art

as well. Black Forest Labs released Flux One tools, which actually gives Flux One all sorts of new abilities. We're getting closer and closer to being able to do with Flux One what we have always been able to do with Stable Diffusion AI art things like inpainting and outpainting.

we can see this picture here where he's starting with this gray suit they mask out the suit and then he turns it into like a jean jacket here or this sign that says burgers and beers mask out the burgers change it to coffee and beers and you can see it maintains that look

that style and then here's an example of out painting where they gave it a picture of just an eyeball and then it out painted the rest of the face they also have what they're calling structural conditioning with flux with canny and depth models now if you're

familiar with control Nets with stable diffusion this is essentially the same thing as the control Nets you've got a depth map so you can give it an image and it tries to estimate the depth in that image like you see here in this gray image and then you give it a prompt

it will create new images trying to model this same sort of depth here and then this is the canny model here where you upload an image it sort of makes like a traced outline of the image and then your new images follow that same sort of traced outline but in a

completely new style they also added image variation and restyling with flux so here's some examples here where you combine an input image with a prompt like they uploaded this duck here and then said fine glass sculpture on a shelf and it made this image of the same

duck but as a glass sculpture or a simple 3D model or actual duck on a desk or epic anime art and all of it modeled the look of this duck but in the new style that was prompted here's a guy taking a selfie like out on a hike or

something flat Vector simple material art large hand puppet with brown fur in the forest pen and ink sketch and detailed anime artwork all from this original image and then here's this one with bread and looks like milk and cheese and they were able to make like

different variations in angles of the same image and all of this is available inside of the Black Forest Labs API so tools that leverage Black Forest Labs will probably start rolling out a lot of these features soon there's a handful of websites that you can use some of

this there's file.ai replicate together AI free pick and CR AI if I jump over to file.ai the very top says introducing flux tools so if I click try it now it looks like all of these tools are available here as well let's see if I

click on flux one fill I'll upload this image of my head shot here and then it asks for a mask URL let's go ahead and create one and now let's just go ahead and you know mask off my face see what it does then use mask and you can see

here's the mask that I just did and let's run it and see what happens I accidentally left the prompt up here a knight and shining armor holding a great shield with fall on it so it turned my face into a shield with fall on it let's

try the cany model real quick once again I'll upload my head shot here I'll leave the prompt as a pink owl and let's just run it it should basically make a pink owl that looks like the shape of my head and I guess that's its interpretation of

a pink owl, but really what it did was it found the outline in the original image and then it used that for this. So if it doesn't see an outline of an owl in my original image, it's going to struggle. So some fun stuff that you can now do with

Flux, giving it a lot more controllability, the same kind of controllability we have with something like Stable Diffusion. Coca-Cola is in the news related to AI this week, and that they created an AI-generated ad that resembles their old Christmas ads, but just generated with AI.

This time around, here's a peek at what that AI-generated ad looks like. Even the people in the ad were AI-generated. This whole thing, fully AI-generated. Uh, I don't know the copyright of the music, so I'm getting rid of the audio here, but

they put out that ad, and well, a lot of the AI people thought it was really cool that Coca-Cola is putting out ads with AI, and a lot of anti-AI people were very pissed off that they used AI instead of hiring people to make an ad

the normal way. They basically said that they wanted to create an ad that found the sort of nostalgia of the old Coca-Cola Christmas ads, but they wanted to bring it into the modern era, and they thought using AI was a really good way to do that. Since we're on the topic of

AI video, The Tool Haen now has an iOS app. Haen is a tool where you can sort of train an AI avatar of yourself, and it can speak like you and look like you and lip-sync like you and translate and still sound like you, and now a lot of

This stuff you can do with an iOS version of it now. Every week before I make these videos, I go on X and I go on Threads and lately I've been going on Blue Sky and LinkedIn as well. And I've been sharing most of the news that I'm

going to share on these videos and I'm asking people for help, like what news am I missing that's not on this list so that I can make sure I get full coverage of what people are talking about in the world of AI this week. Sponsorship Guru

pointed out something that I hadn't seen yet. He said, "Pickle", the AI tool that allows you to join Zoom teams, meet with your avatar and speak from elsewhere. I actually hadn't seen this one yet so I had to look up like what is this Pickle

thing he's talking about and I found their Twitter account and it's actually pretty interesting. So check out their demo video here. As you can see, I'm currently participating in a Zoom call, but what if I tell you that I'm actually not in front of the camera right now and

this is an AI generated live screen that in sync with my voice, Pickle generates real-time video of talking you based on your audio inputs which can be streamed directly to Zoom. So you can have an AI avatar looking like they're on a Zoom meeting the whole time and you could be

in a different room talking into a microphone and in real time it syncs up your voice to the avatar that's actually on the Zoom call. So with something like this, you don't have to worry about, you know, brushing your hair or putting on clothes before getting on a Zoom call.

Perplexity has a new feature because Perplexity just keeps on shipping new features. This time around, you can actually purchase products straight from the Perplexity app. So you can use Perplexity to do research before buying a product, which is something that I

actually use Perplexity for quite a bit. And when you find the perfect product that you want to purchase, you don't even need to go to the e-commerce store to buy it anymore. You could just press a button and purchase it straight through Perplexity. So you can see, if I go to

Perplexity here and say "compare the available versions of the Meta Quest", it gives me a response comparing these various versions, the three, the two, and the Pro. But then, if I scroll down a little bit, look, it's got all of them available for purchase right here

in the same Perplexity app without me needing to leave. And I could simply click "buy with Pro". I would then add my shipping address, my billing address, and then check out. Once you've added these details once, you don't need to keep adding them in the future. It becomes

AI News: New FREE AI Challenges ChatGPT!

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

Matt Wolfe

OpenAI's advanced voice mode is coming to the web for paid users.

It's already available on mobile and desktop apps.

Rollout to free users is planned for the coming weeks.

GPT-4 received an update improving creative writing, relevance, and readability.

The update also enhances handling of uploaded files.

Live video integration with ChatGPT's advanced voice mode is expected soon.

Other AI Updates: Anthropic, Google, and DeepSeek

Anthropic's Claude now integrates with Google Drive.

Google's Gemini gains a memory feature, similar to ChatGPT.

DeepSeek R1 Light, a new model, competes with OpenAI's GPT-1, showing superior performance in some benchmarks.

Mistral's Le Chat chatbot adds search, vision, and coding capabilities, all for free.

HubSpot's AI Prompt Library and Microsoft News

HubSpot offers a free AI prompt library curated for marketing and business use.

Microsoft partnered with HarperCollins to use books for AI model training, seeking author permissions.

Microsoft's Recall feature (web history for computers) is now available on select PCs.

New Click To-Do feature in Windows offers AI-powered image manipulation options.

Microsoft Teams will allow voice cloning for cross-language communication.

11 Labs introduces a tool to build conversational AI agents with various customization options.

Soundraw V4 significantly improves music generation quality.

Black Forest Labs' Flux One tools enhance AI art capabilities with inpainting, outpainting, and structural conditioning.

Miscellaneous AI News and Concluding Remarks

Coca-Cola released an AI-generated Christmas ad, sparking debate.

Tool Haen's AI avatar tool gets an iOS app.

Pickle allows AI avatars to participate in Zoom meetings.

Perplexity adds product purchasing directly within the app.

Vercel's VZ tool gets updates for AI coding.

Rabbit R1 gains a teach mode for computer automation.

Startup H raises $220 million for agentic AI Runner H.

Brave search adds recommended follow-up questions.

Elon Musk predicts AGI arrival within the next year or two.

Google DeepMind's Alpha Cubit uses AI to improve quantum computing reliability.

Leonardo AI announces a Black Friday sale.