Claude Cowork, PixVerse Real-Time, Apple's Creator Bundle

0:00All right, it's time for our AI Roundup. Let's talk about Claude Coworks, real-time AI video, and a couple other things. Ready, Addy? It's been way too long, Joey, time for the AI Roundup. Let's get into it.

0:14All right, welcome back, everyone, to Denoised Roundup time. Hi, I've been Addy. Hey, doing good. How are you? I'm good. Things are picking back up this week. Now that the holidays are over with some new updates. Seems like the AI researchers do take a break. I guess they do. Go on vacation. Or they just know that everyone else is on a break. and don't want to hear about it. They pretend to be on a break. So I think the coolest one this week is from Claude. Now, you know that I've been a big fan of Claude and Claude Code. I've talked about Claude Code for a long time, but it is a little bit technical.

0:45You've got to get into Terminal to set it up. But Anthropic just released a preview of what they call Claude Cowork, and it's basically in the Claude app that you would put on your desktop computer. But it's kind of Claude Code for everyone else. It can control your computer. It can manage files. that can access files, build some apps. So it's a lighter weight version of Cloud Code, but built for non-coding people. So marketing, writing, whatever, other things where you wanted to access files and do things on your computer. So it's sort of like the next step of having a smart AI agent, but that can also do stuff on your computer.

1:21Oh, that's really cool. Have you given it a spin or where is it at? I messed with it a bit. So it's in a research preview. You can only use it right now if you have a Claude max plan, which is the$100 and up a month plan. I mean, I'm sure, what's it called? OpenAI would do this too. And then eventually it would open up to everyone else. It'll come down, yeah. So yeah, this is kind of what it looks like. It's the Claude app on my computer. And then there's a new tab up top next to chat that says co-work. And it has, when you say new task, it sort of has some suggested ideas. So create a file, crunch data, make a prototype, draft message.

2:00I'm just thinking like if you're a creative person, right? Like whether you're an artist or, you know, even like a editor colorist, whatever, what have you, VFX artists, like how much of this is usable for automation? Like you have a bunch of really unorganized file name structures and can this go in? Exactly. That's actually a pretty good use because you can, a good use is like organized files and you can, there's another button. And it looks like a normal Claude chat experience, but on the bottom, there's a button where you can say work in folder. So you can point it to a specific folder and then give it instructions like organize these files.

2:37And then it would like look at them and see if there's any duplicates, big files, maybe you can give it instructions on how to name and rename files. If it's similar to Claude code, which I'm sure it is, it could also look at images and files and like actually make context aware decisions of what is in the image. It's a VLM built in. Yeah, I mean, Cloud's a multimodal model, so it can look at images and be like, oh, it's a dog. If you're like, put all the dog pictures in one folder and all the cat pictures in another folder, I'm pretty sure it could. And go change the metadata tag to dogs or something.

3:11Yeah, it could do that, because I've done that with Cloud Code, so I would be pretty sure it could do that. Maybe there's some guardrails right now as they have it in beta, but I'm sure that it'll do that eventually. Because even when I first tested this a few days ago, It seems like they've opened up some more connections because another thing I learned recently that's pretty cool is there's a Claude for Chrome plugin. And so that's another backing up. We have talked about how the different AI companies have sort of launched their own versions of browsers for this sort of agentic web interface where what's it called?

3:44The browser company has Dia, Perplexity launched Comet, OpenAI launched one, which I don't even remember the name of that one anymore. And so they went the route of like, we're just going to build a whole brand new browser and we're going to put the AI agent in there that can help you, you know, do stuff, possibly take control. Like Dia will let you will take control of your browser and do things. You know, if you're like, go book me a meeting, it'll like kind of do it. Claude took a different approach and they basically they're like, we're not going to make our own browser. We're just going to make a Chrome plug in so Claude can run inside Chrome and then it can kind of do things.

4:13And initially it was like a little sidebar thing where you could do things in Chrome and talk to Claude. But now what I just I learned this this week, Claude the app on your computer, if you're just in the regular chat thing, it can now connect to Chrome in it could connect to Claude Chrome. So if you're in the Claude just a regular chat and you're like, oh, hey, like research this thing or go fill out this form or I need to access this Google spreadsheet and fill out this data. It will like spin up Chrome sort of like OpenAI's agent thing that we talked about a while ago. And then it'll just but this is running on your computer.

4:45So it has your logins saved. You're not sending them to a browser. So it's just controlling your local computer version of Chrome and doing whatever it needs to do. So it sort of has turbocharged this agentic feature where it can go access the web right from cloud. It's the window to the world for your local desktop, right? You're just hooking up the power of the internet to this local desktop computer. Yeah, and instead of the way that OpenAI and Manus handles this, they sort of spin up a virtual machine and there's a virtual browser. but with Cloud it's just running on your computer and it just accesses your regular Chrome browser so you just log into your stuff.

5:21Also I think it's really smart for them to offer a plug-in to existing Chrome because how many times do you think about switching your browser to a whole new different browser and you're like, I've been using Chrome for a long time I'm just going to use that? Yeah, because these other browsers they don't have the robustness of the years that Chrome has had already and the people are used to. And the Chrome browser is what underpins basically the last 15 years of the internet revolution. All of the amazing apps that run on Chrome seamlessly is because Chrome is able to adapt to so many different circumstances and web engines and so on.

6:04These browsers have a lot of catching up to do. Yeah, people have plugins they like. And even though most of these other browsers are built on some version of Chromium, because that's also kind of like the webOS, the plugins don't always play nicely. So you might have a handful of plugins that you use a lot, and then you try to switch another browser. And technically, you can install them, but they don't behave the way they're supposed to. And then you have this dilemma where it's like, well, do I just give up this plugin or do I jump ship and not use this browser? Anyways, it's nice when things are already built into Chrome because everyone already, for the most part, has Chrome.

6:34Yeah, I mean, one of the items that I'm thinking of is like, I need to do some research on a thing that I'm doing. And there's a lot of research material available on YouTube, but I don't have enough time to watch like 15 hours of content. Yeah. To like get this agent on a browser to be on my YouTube account, and then just consume all of it and then come back with like a summary of like 100 videos or whatnot. I don't know if it would watch them all, but what it could probably do is spin up like a little app that would extract the transcripts and then, you know, maybe give you a summary that way.

7:07I will say because Google's been integrating Gemini into like everything. And so Gemini is also in Chrome as well, but it's also in YouTube. Yeah. OK. And Notebook LN does a good job with YouTube sort of captions. But YouTube itself, they have that Gemini sidebar, which I've used a lot, too, where you can be like, hey, summarize this video or give me the main points. And then it'll also give you a timestamp. So then you just click on it and then it'll take you to that thing. So I've used that for like, you know, an hour long interview. But it's like, I know they said this one thing that I really wanted to track down again.

7:39And it just tells you where the time code is and you click and you go. Yeah. I mean, that's that's I hope nobody uses our videos in that way. I guess it's fine if they do. But if we have a good guys, watch it. Especially these roundups. if we mention a product that someone's researching that product and they're like, I just want to find out what these guys say about that specific thing. That's handy. I'm going to throw some keywords out there. You throw it off. No, it's confined to us. Quinn, 2512. Quinn, 2511. Z-Image. Z-Image. Z-I-T. Alibaba. Cloud Cowork. I think this is just the start of unlocking, once people figure it out, a lot of really powerful features and making it much more user user-friendly and not having to deal with terminal.

8:22It's very cool. It's definitely a step in the right direction. Yeah, I've still been living in terminal. You have gone ham with Cloud Code. I had to upgrade my max plan. I was on the$100 plan and then I kept hitting the limits and then I'm like, I guess I've got to do it. To our viewers, I get these random texts from Joey like, I've just figured out this thing in Cloud Code. The other thing with the web browser is if you're in Cloud Code and you're trying to build a little app, because it has web browser control, it can build the app and then it can go launch the app and test the app, and then if the app gets stuck, it'll solve it.

9:00So you could just basically let it run off and it'll start building the app and debugging it and then solving the problem. So I don't have to go mess with the app and then figure out what the issues are and then tell it. It'll do it on a loop until it like... Debugging and iterating on its own. Wow. Yeah. That is like the promise of AI, right? Like doing a lot of that boring stuff for you. Yeah, and this is stuff I have absolutely no idea how to do normally. I wouldn't even know how to troubleshoot web developer code. Yeah. All right, next up, what do we got? So this is Pixverse. They launched a real-time AI video generation model.

9:35Oh, real-time is real-time. We're talking real, real time. I know we've gotten some comments of like, we should demo this stuff as we talk about it. And I agree with that. And I was going to research and demo this one, but But it is on waitlist only right now, so I could not access it. It's still in vaporware land. I mean, it's nice that we mention it here first. We sort of kind of, I don't know, figure out what the implications are. And then by the time it gets into our hands, we're ready to test it. So in this demo, it's like a video generating sort of like a World War II in the snow. But it's got like new text prompts happening.

10:06And as the types of text prompts, the video like instantly changes. Oh, wow. That's cool. So, I mean, that's what they're saying, like real-time. This is getting into gaming territory. That's true. That's true. So, yeah, I mean, I think at this school, this is where we're, you know, I think this is definitely back on our predictions of just like some forms of real-time video generation. The only one we've really seen so far is CREA. CREA. Yeah. I mean, that was a little bit different. That one's sort of like moving the shapes and the sort of shapes react and manipulate the video. This one seems more giving it text prompts and stuff happening as you're prompting it.

10:42Ultimately, I think where this stuff is going to win is when this can run on a local iPhone, a local device that's super light. And you can run very simple mobile games with just AI. But instead of the traditional CG limitations of having very photorealistic stuff not be possible, in this instance, with something like this, you can have a highly photorealistic mobile game on your device running in real time. Like that is more for gaming rather than like usefulness for video production or previs or anything like that. I think more definitely more on the gaming side because that's where the revenue is.

11:23That's where the money is. And so that's where a lot of development will happen. As I always say, as much as I love our market film and TV, the money and the slice of the pie is just so small. Like that'll be the last thing that would they get to. Yeah. Yeah, when you're talking about phones, I was thinking we were going with that was like having this run on a phone or device and you could film with a real actor or some people and it's instantly changing, warping them into the fictional world. But with them still driving their performances. Yeah, again, I don't think it will be for traditional storytelling.

11:57Like, for example, like a 90 minute film or a 60 minute episode of where you can kind of choose your own adventure. I think that stuff is so niche. Where you would see an application like this is, let's say, a big brand like Coca-Cola. They unveil this metaverse-like thing where you can go in and experience things only in that world. Instead of trying to build a metaverse with traditional CG, they're doing this with something like a real-time AI solution. Yeah, like a side experience thing. The quality isn't there, but I could see this being useful for like previs or idea generating and the planning stages for film production.

12:36But yeah, I see what you're saying with gaming. And then we talked about this on the last episode or two episodes ago. Like it all has to tie back into a role model that is persistent. Like you can't, it just can't be a 10 second thing. You got to come back to the things that you walked past. Right. And the thing you did there is still there when you did it. Yeah. All right. Let's see. Next one. This one's up your alley. We got a new flux model. Yes. Black Forest Labs. They're back at it again with a new model called Flux 2 Klein, as in Calvin Klein. No relation. So this one, I just saw initial sort of glimpses of it.

13:12Obviously, quality is bonkers. Everything that Black Forest Labs releases, quality-wise, it's bonkers. They don't do low-quality models, but they're unconventional in the way that they architect things and then the way they think about the market. So this, to me, feels like very much a one-to-one with where Zimage Turbo is today, which is that it's a fairly lightweight model, 13 gigabytes of VRAM, which is nothing. A lot of the older GPUs, I don't know, I think a 1080p or 1080ti could probably run this from three, four years ago. And on top of that, it's open source, and it's really strong with not just text-to-image, but image-to-image.

13:55And so we're seeing a lot of image models really now do a lot of development and releases on image-to-image workflows because it's not just about generating once and then using it. It's about generating once, figuring out what else you need to do, putting it through a model like this, modifying it ever so slightly, and then keep going, keep iterating the way real artists do in any traditional tool, and then finally getting to the output that you want. And this is built, run local, modified and comfy. So really like a Z image competitor from Black Forest. Yeah. And to note here, one thing to note, and this goes back to us being big fans of comfy is, and we have a comfy update later in the roundup, this is utilizing a lot of the comfy native nodes.

14:40And so this is just going to be easier for everybody to develop on and work on. If you're going to build something on top of this, like if you have a really cool style transfer or some crazy, you know, uncanny thing that you want to build, you're going to use native comfy nodes, which is probably already existing in your install. And you don't have to like hunt and find and download that custom node that this thing uses. Yeah, and that's also easy for sharing workflows and templates because it's like I've gotten templates and then it's like it used a version of a custom node that's no longer supported or it's out of date and you got to try to track it down.

15:13Yeah, the more you could just use native built and stuff, I feel like it's the better because it's just less room for error or stuff getting out of date or deprecating. Yep, absolutely. And on the image to image, like what I call image modification, quote unquote, like that stuff, some of the examples look as good as Nano Banana Pro. I mean, you make of it what you will. I know you've been messing with Zimage a lot. Are you going to swap this in with some of your workflows and kind of see how this compares with Zimage? Yeah, absolutely. I mean, you know exactly how my brain works. I'm always trying to get that extra 2-3 % increase in quality or increase in speed or whatnot so I'm always playing around with different tools so right now, yeah, you're right I'm heavily down the Z image path and we have a couple of episodes for you on that stuff I'm going to take some of that and convert it over to Flux Calvin Klein I'm sorry, I just can't help it it's just so easy to make fun of it You're actually going to meet them and be like I love your Calvin Klein model and they're going to be like, what are you talking about?

16:15Do you guys sell underwear or AI models? You're talking about real models, right? Not AI models. Right. All right, we've got to report back once you test, do a comparison with this in Zimage. You got it. All right, also from our other favorite open source world, terrible names again, but there's another update with Quen Image Edit 25.11, multiple angles, Laura. So basically it is an add-on for Quinn Image Edit, but it specifically is designed to give it a source image and you can get output images of different angles, camera perspectives of that source image. And it gives you the control over sort of like what angle and rotation you're trying to get with that, you know, quote, virtual camera control.

17:07I love it. I love it. This is one of the hardest problems to solve in image and video generation. Even though Google Nano Banana Vio is somewhat decent at this problem, we do need a model that is specifically designed to tackle this. Yeah, it's decent, but as we covered in that Nano Banana demo, it's okay if you want to talk about moving the camera in a prompt in Nano Banana, but if you're reverse angle or other angles with specificity, it doesn't do a great job at that. And this has actual, maybe I'm thinking of another thing, but this has 96 camera pose support. So four different elevations, eight azimuths, and three distances.

17:50Oh, you said azimuths. You lost me at azimuths. I'm sorry. What is this, Apollo 13? Why are we talking about azimuths? That's the thing. That's the other angle thing. Listen, nerd talk, while we're doing the nerd talk, when it comes to changing camera angles, the issue is it's not deterministic. This is on the realm of becoming deterministic, like truly having a repeatable, consistent way for you to change camera angles. Wait, so you're saying good or bad? What are you saying? I lost you at deterministic, didn't I? Yeah, so like a computer graphics workflow where you have a virtual camera and you're moving that camera in three-dimensional space is 100 % deterministic.

18:34Like you can accurately predict where that camera is going to go, the output you're going to get, the camera angle you're going to get. You can repeat that five years from now and get the same thing. AI models are not that yet. So this is pushing into the land of deterministic. So good. Two thumbs up. Yeah, I think this is like, I mean, it does one thing, but this is something that's really needed for the stuff we do, where it's like, here's my image, I need to move my thing this way. I'll take it. Like I'll do all my finessing in ZIT or like I can jump from model to model. That's not a problem.

19:11Yeah. And if it does it accurately and well, look at this, got this dog nailed. It does look a little soft, even in the demo. Like, so we would need to run it through an upscaler and a detail enhancer and all that stuff. Yeah. Or I mean, it's still, it's a good, this could be a combo with like still an out of banana or other workflow where it's like, if I shift my camera and then it gets good enough to give me the angle and the information that I need, then you could still run it again and upscale or enhance the details. But at least you have that base image of the angle that you do want. Very cool.

19:40And I know we're not covering this, but I believe Quen also has a version of this called 2512. Yeah, so the big release is the Quen 2512 image model. Quen is already like a sweetheart of the open source world. And then the offshoot of the 2512 model is this camera changing 2511. I just remember also seeing a headline too. I think we had this in the newsletter the other week. But it was basically that Alibaba's Quinn family of models hit 700 million downloads to lead the global open source AI adoption. It is the most popular open source model. Wow, that's a lot of downloads. Because if you think about how many people are actually messing around with AI using Confi and so on, not that many.

20:27So this is probably literally everybody that's messing around with Confi UI. Yeah, well, because basically you're like, oh, hey, it's like the closest thing you can get to like an Out of Banana Pro, but free and you can run on your system or run it on, you know, building it however you want. It could be a lot of device downloads. Maybe, you know, some of the stuff is already running on the cloud and things like that. So, like, you know, when you want to populate a server farm with CoinImage, you're downloading like 2000 at a time. Oh, I see. You're just like a bunch of instances already just downloaded.

20:59Yeah, because there can't be 700 million people messing around with AI. And that seems far too long. And this specific model, which I'm sure I don't know for certain. I'd be surprised if 700 million people know about Nada Banana. Like Netflix has, I think, something like 300 million subscribers, and Spotify has 700 million subscribers. Like, that's a lot. Yeah. Yeah, I don't think it's people. It's downloads. Yes. Yeah, so yeah, I remember seeing that. That was a good... Interesting to see how fast it's grown. I feel like because we've talked about what is the business model of free. And I think it's like I realize it's sort of like the Unreal Engine business model where it's like free to use, free to download, come mess around with it.

21:42But they know it's like once you scale up, you're going to need servers, you're going to need more infrastructure. And so, you know, if you're if Alibaba is an option, like that's something that could spin up and run it for you or their cloud compute service. so it's like hey come download and test it out for free and then once you need to scale like that's where they make the money yeah the way to clarify the way unreal used to work and i think they slightly changed this is all of the non-commercial use is free so like you and me use it to do this podcast whatever it's free and the minute you turn the unreal engine scene into a product that's out there making money so like uh let's say you build a game on it then they can come after 5 % of that game revenue.

22:23Yeah, so it'd be slightly different than that, but the idea of like, hey, we're going to give you free unrestricted use of this product. It's not. There's no features that are paywalled. Come use it for free, but once you are trying to push to another level, then we're going to make money off that. It's a really smart thing to do because it builds like this creative community, builds allegiance to your product, and your product then gets tested and hardened over the years. And then finally, the people that are amateurs and hobbyists, grow up to be professionals in real companies, and then they use that thing to build something incredible, and then all of a sudden now you're making money.

22:57It's a long play, but I think it's a smart one. I mean, yeah, we talked about this before. I think it's the same thing that Resolve did with, I mean, Blackmagic did with Resolve. Yeah, the studio versus the free version. Yeah, the free version is still really powerful. People learn how to use it. Then they become professional editors, and they're very comfortable on Resolve. And on top of that, what Blackmagic was really smart to do was to continue to invest in Resolve and just build the crap out of it. They never stopped innovating on Resolve. Yeah, especially after adding the editing features.

23:29Yeah, and the same thing that Epic Games does. Every three or four months, there's a new Unreal version, and it's better and better and better. They're not going to really stop that. I did not have it in the lineup, but now that we're talking about Resolve, it also reminded me of another update that was actually pretty significant, that Apple has basically launched their own... It's pretty much a competitor to Adobe Creative Cloud. It's Apple Creator Studio, and it's a monthly or yearly subscription. I think it's like 130 a year. And it bundles all of their creative products into one package. So Final Cut, Compressor, Motion, Logic, and then the keynote pages, that kind of stuff.

24:06I'm going to quote Omar from The Wire on this. You ready? Come at the king, you best not miss. Is that a message for Apple coming at Adobe? Yeah, if you're coming at Adobe Creative Cloud, which is the king of Creative Suite, and if you get it wrong, you're irrelevant the next day. I mean, I know there's just so much annoyance at Adobe for their pricing and the monthly subscription and the costs. And, like, people are annoyed. And I don't really know if they're going to switch to Apple because I feel like people who would have switched to Final Cut would have already done so. Like, Final Cut's been...