AI Video Levels Up, Amazon Blocks Muse, Paramount Stays in CA | Diet TBPN

0:01John Coogan:How many days till Christmas? TBPN's Road to Christmas started weeks ago. We got 94 days until Christmas. Double digits. We're so close. Every day, one day closer to Christmas. So get excited, start shopping for presents, start picking out that Christmas tree. How was your weekend? It was, couldn't have been better. Yeah, it was good. We were off on Friday. We went to Boston and back basically the same day. Sort of a brutal flight back and forth, a little tired. But I got completely one-shotted by the new AI video model. SeaDance 2.5. Gone are the days of face swaps. I remember when Dreambooth, this paper, came out from Google.

0:44John Coogan:And it allowed you to very realistically swap your face into a photo. Now, this is something you can just do with a single prompt today. But back then you had to go and take 10 or 20 different photos of yourself at different angles, different lighting, usually just dug through your camera roll, got them. Then you had to feed that into like a Python script that was either hosted on Colab. And then the Colab thing was weird because it was a luck of the draw what GPU you got. So you would like sign up and create an instance. And then sometimes it would just be like, you got a 1080 TI. And you'd be like, bah, like I can't do what I need to do because I need a lot of memory to do a particular thing.

1:20John Coogan:and then sometimes it would be like you got an a100 brother like lock in like it was just random provisioning and people would be like so excited when they got one or the other um but there was this musician who was like a guitarist who implemented the paper and then and then uh basically distilled it down and figured out how to run this is a crazy thing but this guy who's just like a hobbyist machine learning engineer figured out how to get dream booth running on like basically commodity gaming hardware. So then you could go and you could take your face, fine tune it, and then be able to generate new photos in the style of whatever.

1:56John Coogan:It was all built on that open source one, not mid journey. The other image generator, Ahmad Moustak was stable diffusion. It was like stable diffusion as the base. And then this dream booth thing on top. And it was super cool. Cause if you actually went through, if you jumped through all those hoops, you basically had access to the magic avatar thing like three months before everyone else and you could like send one to a friend and be like oh this is cool and then it was just completely commodified it became an app and then it just became available in every chat app very shortly after but the AI video has been moving a lot slower it's very compute intensive obviously since opening I had to shut down Sora and it just feels like none of the major American labs are as focused on this.

2:39John Coogan:So, uh, I got completely addicted to this, had a lot of fun doing it. I'll walk through everything that we did. You no longer need to go get a different app or sign up for anything. I was just texting with codex remotely. I had a PC that was running this. And then I would just be like, I didn't even say, take this video. I would just say like, go find some videos and put Jordy in them. And it would just do that. And I didn't even have to like upload photos of you. I would just say like, go find some reference photography for Jordy across the internet. And so far, it seems to have made a lot of sense not to focus on video, given that Codex and Claude can connect into all these other tools and models and just give you what you want as a user without incurring any of the costs of focusing to make great models or the actual costs of serving them and having to allocate GPUs.

3:28John Coogan:I think people will be having fun. Oh, speaking of Codex, let me tell you about Codex. Codex is a powerful workspace for getting work done with AI agents. Whether you're writing code, analyzing data, creating content, creating some of the videos I will show you shortly, or automating business workflows. Kodos helps you move projects forward from start to finish. So the latest turn of AI video models, they're really, really good. Not just for the high fidelity stuff, which we'll show you some of, but even going back to like 90s VHS footage, it can edit anything and it doesn't look weird 90 % of the time.

3:56John Coogan:CDN's 2.5. The big update here is that it replaces the whole person, the whole object, as opposed to just a face swap. It just hits way harder, you'll see. And it matches all the subtle details of the footage, so things integrate really nicely. Dylan Patel posted a video of him as Drake in this cheetah print video over the weekend. And I think a lot of people fell for this. It tricked everyone. I think it tricked a lot of people because it's super high fidelity. Let's play Semi-Analysis latest post on X, Dylan Patel as Drake. And what's interesting about this is the video model is really good.

4:32John Coogan:but also this Drake video came out a day ago. And so it hasn't saturated the internet. Like a lot of people - Yeah, a lot of people hadn't seen it. This is the first, yeah. I saw this video before I saw the actual Drake music video. And of course the choreography is fantastic in the underlying video. And that drives the motion. He's using AI, he's using CDNs 2.5. Higgsfield wrapped CDNs 2.5 in a model that sort of a harness that sort of makes this a little bit easier. You can use that via the API. It's also available on fall. And there's a couple other ways to interface with it. But it's interesting because it's highly compute intensive.

5:10John Coogan:China's obviously way behind on compute. So what's going on with the strategy? But before we get to that, it's really fun. So we've got to play some of the more videos that we made. So I had Codex finding famous or funny videos to use as bass clips. I would just say I started with the rap videos, did a couple of those. I think there's one maybe in the timeline. Yeah, Noah in the X chat says no headphones, obviously fake. Oh, true. Yeah. That's a good point. Yeah, this one, because I saw this originally with Elon Musk and Sam Altman rapping, and I think we might mute it just because of swearing.

5:43John Coogan:But what's interesting here is that it does a pretty good job of getting us, but there's moments when I sort of blend into Elon Musk, which is funny because it's not like it's originally an Elon video, but clearly the underlying video that was driving the AI model was the Elon and Sam version. And then our heights kind of changed. Like I look a little bit more like Elon there, but also then I was able to just say like, okay, go make a bunch of tech videos, like just find iconic videos from tech. And it did it. And it found like sweet baby rays. I didn't tell it to go get that video, but it just did it, which was very cool.

6:20John Coogan:Some of them came out really well. Some of them didn't do too well. The one that I liked, I think the most, And it's a funny dynamic that we noticed where when you're looking at an AI video of yourself, you're thinking, that doesn't look like me. And then when other people watch it, they're like, wow, that looks exactly like Geordi. Yeah. So let's watch Geordi introduce the MacBook Air as Steve Jobs. I don't think – actually, Apple kind of bans everything. So we'll see if this one gets flagged. But we might be cooked on this particular one. But let's play. I said video is solved. Roll the dice.

6:51John Coogan:The post. Yeah, let's roll the dice on this one. Boom. So it didn't change the voice. So you still hear Steve Jobs' voice, which is also funny about the Dylan Patel thing because it's Drake's voice. So it's like, what did they think happened? He's like lip syncing to it or something. That didn't make any sense. But what's really cool about this is that it changes Geordi, the small Geordi and the big Geordi. It changes both of them in the same frame. And then even there, you only see a little bit of Steve Jobs' mouth and it was able to change that to Geordi. So it can represent what's going on outside of the frame.

7:24John Coogan:And historically, that's been very difficult for AI models. They want the face in the center of the frame. They want to be able to crop it and think reasonably about it, but it did a good job there. Jordi has Jensen taking the GPUs out of the oven. Added some mass for this one. Yeah. It's very funny when you use a different person with a different stature, different musculature. Very funny clip. Iconic when Jensen announced. Was this the H100 announcement, the A100 announcement. It was pretty early, but this is still such a funny gag that Jensen actually did this. I love it.

8:05He's really grunting.

8:07John Coogan:I mean, that's a heavy rack. It's the world's largest graphics card. Such a funny bit that he just did this in the kitchen. It's so good. I got something I got to show you. I got something I got to show you. Jensen's such a master performer. We saw that on stage at Salesforce, Agent Force last week. The key to.NET, the key to industry transformation, the key to success is developers, developers, developers, developers, developers, developers. Developers, developers, developers, developers, developers, developers, developers, developers. It really got like the level of sweat on your face. Pretty correct.

8:53John Coogan:But the face is jumping around. Not quite dialed. I think you'll have to perform that one IRL. Some productions will use this where it helps in urgent scenarios. I think some of the stuff is already actually being used in Hollywood, specifically for reshoots. About 10 years ago, Henry Cavill was in Superman. And he left Superman. They wrapped production. he went to film Mission Impossible. In Mission Impossible, he has a mustache. So he grew out a real mustache, was filming Mission Impossible. And they were like, we didn't get the shot for Superman. You got to come back and film another scene as Superman.

9:30John Coogan:And he was like, I can't shave my mustache. And they were like, okay, we'll fix it in post. And in the newsletter, there's a photo here that we can pull up of how it came away because they had to remove and reconstruct his lip and his mouth using 2017 technology and it was not good it was very very rough probably very tedious very expensive isn't a mustache like the easiest thing to apply oh oh like fake like you should have just shaved it and then put on a fake mustache yeah i don't know i think uh mission impossible they want the real thing but the superman on the right is henry cavill with a mustache that has been removed in post and it was uh yeah widely widely critiqued by the vfx community.

10:11John Coogan:So you can imagine for stuff like that, these tools will creep in very quickly. For most things, it's going to be, okay, well, if you needed someone to drive the video, why don't you just have the actor do that? Maybe for stunts and stuff, you'd do it. But for most performances, I think you'll still be doing it the old way, at least for the time being. What about some of the other more action-oriented? Yes. So this was a fun one. I had to go and pull Johnny Knoxville and Steve-O clips throughout their filmmaking and TV. Generational run. Yeah, their entire run. Make about 20 of them and then put it together into a trailer that we can watch now.

10:51And I really enjoyed this.

10:54John Coogan:We gotta do that in the office. I think Tyler was thinking about getting that.

11:06John Coogan:This one didn't seem that bad to me, but Nick said this was terrifying, getting in a cactus. I don't know. I've never really been poked by a cactus. The hands on the bikes. I don't know how the guy on the left wins. Of course, a lot of Steve-O in that one. A lot of Steve-O here too. But I think these work so well because a lot of the underlying footage looks like AI. It looks like this. The prompts. This looks like an AI prompt. Like drive a jet ski in a pool in a luxury house. This is real, though. This is real. We actually shot this. Yeah, this is real. No, the only thing that's different about this is the suits.

11:44John Coogan:We replaced the characters. But the underlying footage is all real from Johnny Knoxville and crew. The fire hose rodeo. This one, that looks fun. Just kidding. I'm just dying laughing. The old alligator in the house. The old alligator in the house trick. I think they chilled that guy down, so he was real slow. And then the bull run. But a very impressive demo because it's able to replace multiple people and do it very consistently, even through some really aggressive motion. So I thought that was pretty cool. So the whole pipeline here was I was running Codex on my PC, interacting with that from my phone, the ChatGPT app, And then I had Codex go and generate reference images with just ChatGPT images because it's pretty good at that.

12:38John Coogan:And then use those to swap through CDance, which was actually rendered via Higgsfield. Higgsfield, amazing in terms of just delivering the assets like as an API. Extremely confusing as a product because they basically have two ways to give them money. You can load up API credits and then you can like buy a plan. And so at one point I ran out of API credits and it was like time to upgrade to the plan. And I was like, okay, cool. I'm having a good time. I'll do more of this. Click the button. I'm like, okay, yeah, you loaded up a couple hundred bucks in here with like the ultimate plan or whatever.

13:11John Coogan:And then it's like, those credits are not available in the API. And I was like, great, thanks. And then, and then it was like, okay, time to add some credits to your plan. I was like, okay, this is what I need to do. I add credits to the plan. It's like, those credits are on your plan. Those aren't API credits. Those are plan credits. And so I have like so much random money stuck in various pieces of Higgs. Well, and I think that's been a massive complaint. Higgs field has faced a bunch of, I think, various allegations over the last year. I mean, the website looks like a gambling company. It'll be like countdown 10 minutes to claim 75 % discount on your next credit.

13:50John Coogan:It's like it's totally optimized for like brain rot as you go through the flow being like, oh, you got to buy. Oh, yeah. I had seen stuff of like not being able to actually cancel your plan. So now that you have a plan that you can't use for anything, it's over for you. It's over for me. It's over for you. You're a customer for life. My question's about CDNs 2.5. I was like, how is China winning? They have one-fifth the compute, one-fifth the compute. No rules. Is it that simple? And then also, why aren't they open sourcing it? They love open sourcing everything. They're like, we love open sourcing, except when we're on the frontier.

14:21Yeah, good point.

14:22John Coogan:And there's a weird dynamic there. So first off, did some digging, got some answers. Let's dig into it. At a high level, intellectual property rules are somewhat flexible over there. That's true. But the models are insanely compute intensive. Like it is, even if you're like, I don't care about Hollywood IP, come sue me in China, good luck. There's still a finite amount of compute. And so why aren't you on the RSI trajectory? Why aren't you working on coding models? Like, why don't you just take that compute? and put it on the next GLM run or the next Kimi run, right? Well, there are reasons for that.

14:59John Coogan:First is that ByteDance has a more natural flywheel here than the leaders in America. So Anthropic and OpenAI do not have good surfaces to deliver video, AI or not video. Like OpenAI tried with Sora, got a decent amount of people to use it, but it was not the obvious flywheel that ByteDance has. And so over in China, ByteDance is a big company leading, and they have TikTok doing, CapCut, ads, e-commerce, recommendation systems, tons of immediate benefits to the core business without needing to pull together a new demand surface for video. This sort of exists in America with Meta and Google slash YouTube, right?

15:37John Coogan:But AI slop is deeply unpopular in America, although I debate whether or not this is slop. This is sloppier than some of the videos we showed you. But both companies, DeepMind and MSL, they are on the RSI trajectory. They're focusing on agentic workflows, personal agents, knowledge retrieval, coding, recursive self-improvement, AGI. It feels like they are more AGI pill than ByteDance, which is like, look, we got a good thing going with Duyen and TikTok. We got a lot of video space. Let's dominate there because we have an opportunity. And also like the AI, the IP infringement thing is not actually dead to rights.

16:21John Coogan:They haven't necessarily been caught in any major way. It's just that when the model comes out, people are like Tom Cruise and it's like perfectly Tom Cruise. And so it's like, what's going on here? So some of that could be the outputs and the harness that actually like reviews outputs. There is an argument on fair use for training to include everything, especially when you think about TikTok. So people go to TikTok and they upload clips from movies. Some of those are fair use. Some of those are transformational. You can actually make a vibe reel from Top Gun that is transformational and passes the fair use test, which is that it's not a replacement.

16:57John Coogan:So you're not like, why would I watch Top Gun when I could watch 30 seconds of Top Gun edited like a crazy hyperlapse edit? So it's not competitive in that way. It is transformative and it provides some commentary layer on top. So that is actually okay. You can use clips from movies if you're commenting on the lighting and you're trying to explain the lighting. All that means is that some of that data goes onto TikTok and lives there forever. And Hollywood can't say anything about that. Same thing with YouTube. Same thing with Instagram. There are plenty of film criticism channels that post fairly long clips from movies completely acceptably.

17:34John Coogan:And they don't even need to share revenue because it is transformational. It is open. It is fair use. So they have a lot of data to train all this, but they're not open sourcing. They're not doing open weights. They're certainly not open sourcing the training data set. And it's a little bit of a narrative violation from the default Chinese AI strategy. There are a few reasons for this. First, no one wants to open source when they're in the lead. We've just seen this time and time again. And second, with coding models, open weights are a play to build a Chinese tech stack. And so you open source the coding model, build the developer ecosystem, and then maybe that helps Huawei stay competitive.

18:15John Coogan:Maybe that reduces the amount of dominance of the American technology stack. And so there's a little bit of public incentive to build up the indigenous supply chain that comes from that. Less so in video, where they are vertically integrated and have everything down to the distribution surface. So, in general, if you are a challenger that needs an ecosystem, you want to open up. We saw this with the early meta stuff. And then if you're an incumbent and you have distribution, you want to close down. So, there's some interesting knock-on effects, mainly around a speculative future value of video models and training humanoid robots.

18:54John Coogan:Like, China also seems to be doing really well in humanoids. If AI video becomes really important in sim-to-real translation and training these AI, training humanoid robots with simulated video, then this could be part of the bigger picture. But I think that's farther out. I think for the next month, we're just going to be talking about memes and funny videos. Anyway. That's right. Meta on a tear. Meta's up 27 % this month, according to Shield. Monat, almost entirely thanks to Muse. This makes sense, right? Is that? Almost in time. I mean, the semi-analysis breaks down. Basically, when you have a company that's spending hundreds of billions of dollars on something without much to show for it, and then you have something to show for it, then the market's going to react positively.

19:39Yeah. They're saying, so you're saying there's a fighting chance here. Totally, totally. And the feedback on the product is incredible. There's a lot of investors that have tried it, and they're like, hey, this is a good product.

19:50John Coogan:And interestingly, they are in the lead in a subcategory where they have never been the lead before. Like Google had Nano Banana, which like there was a moment where it was like, oh, Nano Banana is the best. Obviously, Anthropic had best coding models for a while. Like like crowns have changed, but Meta's never really held the best model in knowledge retrieval or when they did Vibes. That was not the best version of that. But this is at this moment in time, the best version of what they have. And it's building off of a model that's not on the frontier, but able to deliver an experience that puts them in first place for this subcategory, which is exciting.

20:28And can immediately be incredibly integrated across everything else that they do. Exactly. Shopping, messaging, all these different things. Yep. Shopify and Muse announced a partnership this morning. Oh, cool. Counter positioning a little bit against Amazon. Banhammer. Amazon came down with the banhammer. Buko Capital says, Amazon cuts off Muse. While I am bullish, Meta and Muse, I think many people are overlooking the digital knife fight that's about to occur. Nobody wants to get commoditized or layered here. Let the games begin. Amazon says it has cut off Meta's new Muse personal AI agent from shopping on Amazon.com on behalf of customers.

21:03After attempting unsuccessfully to get the Facebook parent company to voluntarily exclude the e-commerce site from the experience, the problem, Amazon says, is that it never agreed to any of it. Meta didn't tell Amazon that Muse would access its store. The agent itself doesn't identify itself when it browses, and it appears to capture and store customer credentials, which the company says create privacy and security risks. As of Sunday night, people trying to use Muse to shop on Amazon were seeing the pop-up. Continued access by an unauthorized AI agent violates Amazon's conditions of use, to which our customers have agreed.

21:35John Coogan:Interesting. I wonder how Walmart will respond. I think Walmart had always been more friendly with chat apps, specifically chat gpt and that's the kind of thing that you do when you're not when you're not the leader in the space you're more willing to yeah experiment i think they wound up rolling their own uh i think they have sparky and amazon has rufus and so there's two dog named uh ai agents that do the shopping and then and then most recently tyler's been dailying rufus by the way yeah just for everything for everything just uh psychology yeah getting close to solving another millennium. Yeah, yeah, yeah.

22:11Rufus, underrated.