AI News You Missed this Week! Suno V4, Auto Agents, & More!

Hello everybody. Welcome back to the Matvid Pro AI YouTube channel. So, as I explained in my last video, it was election week and it really seemed like the world of AI has taken a break for it. But actually that's not the case. A lot of stuff just got swept under the rug and there's some really cool things to talk about. So let's get right into it. So first things first. OpenAI has a new domain for chat GPT simply chat.com. obviously getting this URL is something that costs quite a lot of money. OpenAI has not told us how much they paid to get chat.com but the estimates are somewhere in the range of $15 million, which is absurd, but nevertheless, if you want to go to ChatGPT, all you have to do is go to chat.com and it will redirect you right to your best pal Chat GPT. Not really a huge piece of AI news by any means, but you can see the community is pretty funny memeing about it. Drop the GPT, it's just cleaner. Looks like somewhere during 2023 last year, $15.5 million for the domainchat.com man. Anyways, moving on. This is probably the biggest piece of AI news that I think was just swept under the rug. No one really noticed that this happened, but Microsoft announced their own agent workflow. They call it Magentic one. It can browse the web, access files, write and execute code, as well as support a human in the entire loop of the process. Hussian, who is a researcher at Microsoft, presented this as his favorite task executed by the workflow, ordering a shawarma sandwich. So let's take a peek at what this actually looks like. We start off with an initial orchestrator initial plan. We are working to address the following user request. Can you order a chicken shawarma? From Tenure Seattle to answer this request, we have assembled the following team. We've got a web surfer who can perform web searches, open pages, interact with content, clicking links, scrolling the viewports, et cetera, can summarize web pages or answer questions based on the content on the page. The web surfer can also be asked to go into sleep mode and wait for pages to load in cases where the pages seem to be taking a while. There's also a coder, a helpful and general purpose AI assistant that has strong language skills, Python skills and Linux command lines skills. Then finally the executor, an agent for executing code and handles local files. Given fax in the command, the request is about ordering a chicken shawarma. It's kind of doing an over analysis obviously, because this isn't a very complicated task. The location mentioned is, you know, somewhere in Seattle. Fax to look up current menu availability of the chicken shawarma, the ordering options available, contact details or website facts to derive none specifically, but checking the type of ordering process. If they offer online ordering, for example, based on available information, it then goes on to do educated guesses as well. This restaurant in Seattle likely has a website or a phone number for placing orders. And chicken shawarma is a pretty common dish at Middle Eastern restaurants. And if we scroll down a little bit, it then just gives a very basic rundown, five bullet points of what it plans to do to complete the user's request. So that's great. It is not hard for an AI agent or an orchestrator to come up with a plan, but I gotta say, I like this level of detail. It's already very promising. The hard part comes in satisfying the overall request, right? Getting these agents to work together to order a sandwich. Progress then starts to be made. Web surfer is going off. It seems to initially pump out an error, but the orchestrator comes back to the web surfer and says please retry searching for the menu. The web surfer then goes on after encountering an error and seems to do a pretty decent job at finding the restaurant in question. The orchestrator then goes back and says, please visit the website. Basically click on that link and check how to place an order. The web surfer then responds, clicks on that website and brings up the page. The orchestrator then says, navigate to the menu section as you would and check for the availability of the sandwich. Also explore the order online section. The web surfer then clicks on menu and just sends that information directly back to the orchestrator. As you can see, there is a very stark contrast between the Orchestrator agent here and the actual web surfer agent. The web surfer agent seems to be a lot more simple. It seems to be good at clicking on the right things that it's asked to click on and then sending that information back. But it can't autonomously click multiple things at once and jump around the website quickly. The Orchestrator does do a great job though at keeping it in line and keeping it on task. Anyways, moving along, it tells it to click the shawarma section of the menu, to which it goes ahead and does. The orchestrator then says click order online to find out how to place an order. And actually the web surfer makes a little bit of an error here. It seems to have misclicked or missed interpreted the instructions and clicks on a photo of some food. A lot of older agentic workflows would get caught up right here and be stumped by this, trying to figure out how to get out of that. The Orchestrator then says no, no, no, click order online section to provide details. The web surfer was smart enough to navigate back to access the online ordering section properly and then retries and is actually able to bring up the map for ordering online. Finally we figured it out and it is able to answer the question, yes, you can order the chicken shawarma from this website in Seattle and it can give you exactly the options for online pickup or delivery, which is pretty cool. Obviously we could also prompt this thing with some additional information to say, hey, charge this card and you know, have it sent to this address for this time, etc. And it would probably be able to do that. This is a very impressive agentic workflow. I am really, really happy to see that the Orchestrator has a really good idea of the context of the situation as a whole and pull back in on that web surfing agent when it actually makes a mistake. Now we're going to take a deeper dive and a deeper look into this for sure because they have a blog post, but they've got a full report, they've got code for this as well, and it is technically open source. At any rate, guys, here's a nice high level overview of this thing and a nice flowchart example so you can really see how it works. We have a basic task in this case running some python code, and it's actually quite a lot more complex than that. This is a pretty complicated task. Anyways, it goes to the Orchestrator, creates a specific dynamic plan which we just saw in that example, and we could see from the flowchart. This is basically how it works. The Orchestrator will then divvy up the plan to all of the sub agents. Web surfer, for example, navigates the URL, extracts the C code and then sends the C code onto the coder. But honestly, I wouldn't be surprised if instead the web surfer just went back to the Orchestrator and then the Orchestrator went on to send that code to the coder. That's definitely more likely, but at the same time that that's happening, the Orchestrator also sent off to the file surfer, accesses an image, extracts code, goes back to the Orchestrator, Orchestrator then sends that code over to the coder to analyze the Python code. Finally, once we get our analysis of both of these codes, I assume these coder sub agents send Back to the Orchestrator, and then finally the Orchestrator goes to the computer terminal and that computer terminal agent executes the code and you get a complete task. Pretty incredible. Oh, and after doing some digging on the GitHub folks, it looks like it's actually under the MIT license, which is definitely a good one. It looks like it's actually just fully open source, so scratch what I said earlier. Might have to put a note on screen about that. But yeah, MIT license, which is good to see. All right, guys, let's take a peek at some of these examples. Find and export missing citations of a paper. That's a pretty complex task. So this is the name of the specific file that it's editing. It has that pointed out. It's going to look for academic papers to cite that are relevant to the topics covered in the PDF. Methods or tools to find new research papers it already knows, like Google Scholar. It's kind of crazy. Is this how fast it works in real time? That's. That's a big question that I have. Because if it's this quick, that would be very impressive. Now it's using the web surfer to find the papers themselves. Seems to be using Bing, unfortunately, it's now clicking around. Specifically, it found a paper on a popular website here, AR xiv, of course, from Cornell University. It's summarizing the contents of the paper at this point, and it comes back and does that well, and it formatted it and put it into a text file. The coder agent did that. And Executor is now coming back to run a piece of code, run the following Python code to write the summaries to a text file. Wow. So it seems to just be working all in the computer itself and just using commands in the computer to do everything, which is pretty crazy. And it seems to be running, doing all of this on Unix, by the way, so this isn't running on Windows or Mac. Well, there the summaries are complete. We've also got describing trends in the S&P 500, which is really interesting. This one's pretty short, so it seems to be a quick and easy task. Orchestrator goes off and sends instructions over to the coder, actually, which executes Python scripts to fetch and display latest trends in the S&P 500. And there it goes, it executes that. And there we go. We have a final overview from the Orchestrator. Pretty darn cool. We already saw the sandwich task. Next up we have counting the number of members of MSR hacks, which I have no idea. I have no idea what that is, but I guess we're about to find out. Okay, so we give it the initial user prompt. It does a whole orchestration as we saw before. It is now going to the Web Surfer to find some information on the MSR team or whatever, the hacks team. It seems to be Microsoft Research actually able to pull up a website on that specific team, check the people section to count the number of people I assume, and it's able to click on that, then provide the image and count and list the numbers. And I guess Web Surfer can actually do that. It scrolled down one page on the browser, which is really impressive. Actually, this Web Surfer seems to be really impressive. The fact that it can know to scroll down and find more names and more people listed. I've seen a lot of agents struggle in that regard. Specifically the new Claudia agent does not know how to scroll on web pages. And yeah, it has seven total members listed on the page. Easy enough. So, yeah, an impressive agentic workflow coming from Microsoft. I think this really is significant overall though, because we're seeing a shift from LLMs that can perform only like one task to agentic workflows that can perform a variety of tasks. You can see this thing is decent on the benchmarks. However, it is nowhere near the accuracy level of a human. It's still around like 40%, 30 to 40% accuracy in the tasks that it is given. And you'll also notice right here that GPT4.001 preview are the models that they were testing in this thing. So it seems to be able to just be swapped out for different models. We could try Sonnet 3.5. We could, you know, llama 3.5. Any, any model we want to swap in there seems to work. But as I said earlier, yeah, this is all available on GitHub under the MIT license. So you developers should definitely go check this one out and implement it and see if you can maybe even improve it. It's great for the community as a whole and I'm really sad to see that this kind of got swept under the rug. You know, not too much traction here, especially for the AI space. So I'm happy to include it in today's video. I want to see what the open source community can do with this. Before we move on to the next piece of news, I do want to talk about my predictions for next year in regards to AI agents in 2023. I did say that 2024 was going to be the year of AI agents and it doesn't seem to be the case, obviously we're getting pretty close to the end of the year and yes, we do have some specific agent things going on like the Claude agent and this new Microsoft Magentic one, but definitely not the year of AI agents or anything like that. It was more like the year of AI video generation, I think. I do think though that we're going to see an increase with this agentix stuff in 2025 especially because now we have OpenAI's O1 model, which isn't done or anything. We have the O1 preview and Zero1 mini, but it looks like we're going to get access to the full blown 01 this year, which already does self reflection at a high level, enabling it to be a much better model, do things like coding very successfully and models like this can be plugged into workflows, open source workflows like Magentic one and increase the success rate of it. We will see how things go. But I am really hoping in 2025 to see just a killer app agent, something that can click around your screen autonomously and quickly and get things done with a high success rate. A high enough success rate to the point where you can actually ask it to do things and not have to worry about it completely bombing or failing a task. That's my hope. That's my prediction for 2025. We'll see if it ends up being true. Anyways, let's move on to the next piece of news. Suno AI has just given us a massive teaser at their new V4 model which is apparently coming soon. And if you didn't know, there's basically two big competitors in the AI music generation space. We have Sudo AI and then we have Udio or Yudio. We've had sudo AI v3.5 for quite some time now and it's decent, but I think it lacks in the voices particularly. Udio typically has very clean sounding voices, but Suno is able to produce wide varieties of music itself as well as being very, very quick to generate full length songs. So here's the little teaser for Suno V4.

Yeah, so no, yeah. Feels like I've been waiting for a lifetime. Yeah, looking forward to today. I bet you thought you never hear me so clearly. Welcome to.

Dang, it's like risen us up, man. No, I mean in reality this little preview is probably cherry picked, but it is super impressive. The voices are clear. Everything is clear actually. Oh man, I'm excited to get access to this thing and in regards to that, my good friends at SUNO AI actually reached out to me and they said hey, would you like to get to access early access to V4? And I said sure, why not? Of course I'd like to try it out. So you guys can bet that a full video on this V4 model whenever it releases is going to be coming very quick from me. So excited to give it a shot because it sounds pretty game changing. I will say though, we're still waiting to hear on the 11 Labs music that was announced seemingly months ago. They haven't released it yet and their demos sounded super promising. Pretty competitive with what we just saw, so we'll see. Obviously everyone's pretty darn hyped for Suno V4. It's been a while since we had a big jump up crease in the AI music generation space. People are pointing out that the vocal quality is really really good. The servers are going to be burning probably the voices sound extremely human. People also ask when they can prompt for a duration, for example 10 second intro bumper. I don't know, more complex stuff like that. I just don't see it being Sunno's priority. They really seem to be just give us like a text prompt and we'll give you a really good and human sounding song in like a few minutes. McKay, Wrigley, you have my attention. Lots of people are overall excited though. Obviously SUNO has been definitely a community favorite without a doubt, but so is UDIO AI and 11 Labs. Next up we have a quick Flux update from Black Forest Labs. Obviously if you guys don't know, Flux is the model that took the AI image generation world by storm. It was open source in a decent way. It was great with things like text adherence and prompt coherency and obviously its ability to be fine tuned and trained is pretty much second to none. Anyways, they're adding new high resolution capabilities to Flux 1.1 Pro, extending its functionality to support four times higher image resolutions up to 4 megapixels. So this is 4,256 pixels by 4,256 pixels. I think that's like off the top of the dome. While maintaining an impressive generation time of only 10 seconds per sample. I don't know how they're doing that. Maybe it's some upscaling or if it's just raw on the model I'm blown away. But Black Forest Labs is impressive. You can see the examples they provide are obviously all very impressive. Now playing Flux Ultra. I really like this one with the pool in this old broken down place with the light Shining through. That's cool. Higher resolution, no compromise in speed. This Ultra mode option enables image generation at four times the resolution without sacrificing prompt adherence. Unlike other high resolution models that experience significant slowdowns at higher resolutions. And you can see the performance benchmarks in comparison to ideogram V2 with an up sample or mystic is just insane. 10 seconds and you get a very, very good ELO score. Mystic is maybe twice as long to get a generation and then another twice as long for ideogram V2 with an sample. They have a brand new RAW Mode Coming to 1.1 Pro as well, which captures the genuine feel of candid photography. Toggle this feature to generate images with less synthetic, more natural aesthetic compared to other text to image models. RAW mode significantly increases diversity in human subjects and enhances the realism of nature photography. And I mean all of these seem to be pretty darn good. This would be for more specific prompting, more specific styles. This obviously looks like a flash photography photo from the early 2000s. Let's say this looks like a photo taken on an iPhone. You know, this looks like an old 90s TV show sample, something like that. Anyways, this is all available via their API. I wish they would release this open source though, man. I. I would really push Black Forest Labs, at least when they release like a new foundation image generation model to open source all of this old flux stuff when it becomes old. Because that would be pretty awesome. Anyways, yeah, four megapixels. It looks pretty darn good. The detail is definitely quite a bit higher than your standard AI generated image. And finally, moving on to our final piece of news of the day, they have Xportrait 2. This is by ByteDance, the same company that owns TikTok, if I'm not wrong. Next level AI lip syncing. It not only delivers incredibly realistic lip movements, but also captures natural facial expressions that perfectly match the emotions. And, and you'll notice something here. We just saw a very similar technology come out of Runway called Act one. So is this better? Well, let's check out the rest of El.Cine's thread here to see some examples. Down here we've got the driving video, obviously a scene from a real actor in a movie. And then we've got, you know, probably an AI generated image of a woman. And it's able to apply it shockingly well, like pretty scarily well.

It didn't take me no 18 years to realize the soil was hard and rocky.

Dang man, that's intense. It actually does a really good job at transferring that over. This could very well be better than act one. I wouldn't be surprised. Looks like another example from the same exact.

I got a life, too. I gave 18 years of my life to stand in the same spot as you. Don't you think I ever wanted other things? Don't you think I had dreams and hopes? What about my life? What about me? I've had a life.

Wow. It's pretty good. I feel like the performance by the actress almost doesn't fit this woman, though. I guess that's obvious. They're trying to show the transplant here because she, you know, she's wearing headphones and stuff like that. But yeah, if you were to use this, obviously with the intention of, oh, I'm going to perform as this specific character. I can already see you can fine tune your own acting to more fit that specific character. But yeah, it does a good job with, like, turning the head and sort of making that headset feel pretty real in 3D, I have to say. And obviously the facial expressions, the eye movements, the eyebrows, the mouth, it's all very scary realistic. Let's be real. This is terrifying. We've got another one.

Confidence and trust in me and that.

I have signed a letter of agreement, a contract in which I have accepted that responsibility. Wow. This one's a little bit more choppy, but you can see that some more jerky motions and reactions from the original clip can lead to some pretty jarring changes. So that's obviously going to be a limitation here. The original actor is obviously acting crazy and that doesn't perfectly translate onto a still picture of a head. But again, those eyebrows are pretty crazy, man. It does a. It does a good job. I don't think there's any doubting here we have a less realistic character. This is a obviously more artistic representation, but. But it seems to do a pretty darn good job. Look, she looks around and turns her head. It seems to be able to do a wide var of swinging head movements, which is something that runways act one. Definitely struggles with. She turns her head all the way. Wow. And it's good. It almost does, like, physics with the hair, if I'm not mistaken. There's definitely a little bit of morphing in the hair as she turns her head, but it's not too bad at all, man. Definitely just from deriving based off of one initial image. At the end of the day, this is really darn cool. And this is something that I also wanted to mention here. It can do tongues, man. You can stick your tongue out and wave your tongue around and it will pick it up and translate it onto the character pretty darn well. I mean, yeah, it's. It really is so cool, right? There's so much possibility here. I was really excited when act one came out and this is honestly looking to be even a little bit more capable than Act 1. Hopefully they open source this technology so we can have things like an upgraded version of Act 1 that is based upon this. It seems to be able to be applied to any face. One thing I'm wondering is how it does with non human faces which we don't see in these examples but so far so good. This is darn impressive coming from ByteDance and folks, that is going to be it for my AI News roundup of the day. Let me know if you have seen any of these floating around this week. A lot of this was surprising to me. Impressive technologies, impressive agent technologies, impressive what is essentially deep fake technologies and some cool stuff from Flux. A little bit of a slower week for sure, but definitely worth checking out. Thank you so much guys for watching today's video and I will see you in the next one. Hopefully it's going to be a Suno V4 overview and super excited to try that one. I think that one might have actually been my most exciting little tidbit. They didn't give us much of a preview, but darn it looks good. See you guys later.