Gemini 4 ArgonとMetaのMuse Wonder、悪くないが「圧巻」ではない

Two new Frontier AI models came out this week. Google and Meta, they're in the kitchen. They're cooking it up, but is it tasty? You gonna choke? That's right, we got Google, Gemini, Argon4, and MetaMuseWonderful, and both are pretty good, but maybe not great. Fair, look, Argon4 excels at science and pelicans on bikes, apparently, but it can get a little makey-uppy at times. That's not good with code and MuseMetaWonderful, which is, yes, something they did there with the name. You get it, right? Yeah. It struggles with 3D environments. But Muse did give us the best video of the week, Kevin. Oh, the little memes with the Muse agent doing things with the fluffy butt? No! Mark Zuckerberg struggling to drink a Guinness. I was actually talking about this at dinner yesterday. My favorite thing is I have my Muse, whose name is Euler, constantly source. it's kind of like that that john mayer base guitar face it's like i love that billionaires

are just like us though plus we have news about xbox and epics daddling up to ai vibe coders finally also quick looks at clawed motion a new ai tool that lets you add sound effects to any video which we of course did dumb things with and slow books pro is coming for you and your accountant but like in a good way and then finally i'll walk you through my new ai doc which you can watch right now on this very channel and i'll kind of talk you through how you might be able to make one of these for yourself is this the one where imogen is getting all cheeky with us again imogen cheeky kevin i'm a prestige documentary you two are a podcast welcome everybody to ai for humans your weekly guide to the wonderful world of ai my name is gavin for still that's kevin prayer and kevin to this week we had a catch-up moment in the ai space And I'm not talking about the thing you put on your hot dog. This is not ketchup and mustard. This is catching up. This is catching up. Have you ever started a hot dog? I've caught a whip of what this podcast is going to smell like. Get rid of it.

It's going to be a stinky one. That's right. We're here in a week where we got Frontier Models from both Meta and Google, both companies which in the past we have covered have been struggling to catch up to the frontier of the edge of where OpenAI and Claude is. I think my initial reaction on both these models is, good job. You're getting there. You're almost there. You're probably in like third and fourth place still, but you are making significant progress. What a supportive parent. Would you bring them orange slices after the game, or would you throw their cleats at their noggin and say you're a disappointment, Kevin? Oh, sorry. I was channeling. Sorry. That's all right. I've been there myself. You and I both had disappointing athletic childhoods. This is probably why we're talking about AI on a podcast right now. That's exactly why we're here today. If someone wants to key Gavin's father being disappointed on his chroma key green sweatshirt, have at it. I didn't think about that. Okay, so let's talk about each of these models individually and what they're good at and kind of what the reaction to is.

We have to start with Google, Gemini, Argon 4, which isn't like a great name if I'm saying anything. Like, why couldn't we have just done Gemini 4? I don't know, but Logan Kirkpatrick has been teasing that they've got a big frontier model coming for a while. Sundar and Demis and Savas all talked about this. It is good. I think this is, it shows that Google has not given up on this race. My favorite video, there's a lot of things we're going to talk about with this. We're going to talk about like what it can do gaming-wise, what it can do 3D-wise. But my favorite video, Kevin, I think that we have to show to start this is... Is it what robots can do with your holes? Yes, it's what robots can do with your holes. Every time I send you a video like that, you get mad at me. No, this is a legitimate version of that. This is a legitimate version. This version is a video of Argon4 basically is being used to train robots. And we've talked about this on the show before where the big thing about these frontier models as they come out now is, yes, they're going to get better at coding. Yes, they're going to get better at writing your marketing documents or whatever you might want to do with. But also the future of AI is embodied robotics AI.

And what they showed here is the Google Argon4 helped a robot in one shot. This is the thing that's interesting about this. one shot, learn how to play cornhole, and like sink the bag every time. So I know cornhole is a game of like, you know, you have to have the exact same motion, the exact same place. Robots would be very good at that. But the fact that it learns. Literally, for those who don't know, you throw a beanbag into a hole, usually across the room where it's a slightly ramped, you know, elevated surface. But what's fascinating about this to me is that the method of training, they played cornhole, filmed themselves doing it, gave the video to Argonne, and Argonne was like, all right, bet it figured out the physics it watched the video back it ran simulations with a stimulated robot was a i believe an aptronic robot um humanoid and it did simulation after simulation then it applied its simulation to the real world bot watched footage of that evaluated that like a player looking at like game day footage and eventually sank it into the hole so i mean pretty

smart. This is a good example of showing you how quickly the skills will disperse across all this stuff, especially with robotics. I think a couple other things that we saw that kind of were interesting to me, there was a good video that Google put out from one of their creative technologist producers who made Japanese Pagoda in Blender. Really solid looking thing. I need to hook up this weekend my PC so I can try some Blender tests with this model because it does seen from a 3D perspective it's really good game wise it's good uh my experience we'll talk about this in a second maybe or I can even talk about it here um I think it's really interesting and good um they shared a bunch of interesting about it you tried to one shot like a golf game right yeah so I came up with you to do that because there was some confusion with Argon like where is it how do I actually use it oh do you have to get into anti-gravity which is Google's coding suite to use it can you log into a regular Google there was a lot of confusion so I've had a Gemini pro account for whatever two years now that's a $20 a month account and probably mostly I really

have not been using it and I have been considering well why am I doing this and then Google looped it into like their whatever storage account so now I'm kind of stuck there but I did this in I downloaded their app there's a Gemini app that I downloaded because I figure okay well this will work like it does with Cloud Code or it'll work like it does with Codex and I said I wanted to go do this thing. And then I realized when I got into that app is like, oh, that app, you can't work locally through that app on its own. Like you can't figure out how to locally work out because it's not like built the anti-gravity, at least in my experience, was not obviously built into that app, right? And anti-gravity, if you're not familiar, is Google's like kind of CLI command line interface to allow you to do stuff and code a bunch of things. So what I ended up doing was putting my prompt for this game, which I called the long drive contest, which is basically a golfing game where you, I'll put a prompt in our, maybe in our show notes, but it's a golfing game where you try to hit a long drive each time to get three chances. The course changes every time. It's actually a pretty fun idea as a game. So I gave that prompt to Gemini Argon4, and then I

went and gave that same prompt to Opus 5.5, which really has become my daily driver. And like, I will say again, Opus 5.5 is the thing that I'm using for everything. And there's a world that Fable 5.5 could be launching today for all we know or very soon that that might be the next step but i put these two prompts together argon did fine i think if you look at these back to back there's a video that you might be watching here on screen i think the opus 5.5 game felt better and more uh full but argon did make some interesting creative choices about the gameplay that gemini that that opus wasn't so i think that there's interesting things going on under the mix here that make it unique and kind of interesting for that reason. It's so fascinating to me every time one of these models drops because at one point it was, you know, can it render a bird on a bicycle? Can it do it in a specific format? Okay, now can it make a full-blown WebGL accelerated browser game about this bird on a bicycle or about a long drive contest? And what you're seeing is it's a little murky sometimes because you're not necessarily seeing

the raw intelligence of the model itself. It's a very capable, very smart model. But you're also seeing the harness, the system that supports it, as well as how much thinking time and compute it's going to get. And so it's just fascinating to me, like, the different ways people poke and prod at these. From a raw coding standpoint, Theo had tweeted out that it's not great. It hallucinates on code. That is a massive no-no. In a creative writing task, it making up a person's name, not very detrimental. You can edit your way around that. But if it looks at your core code base that gets components wrong, you've got back doors, you've got crashing, you've got some real bad headaches. So Theo was not too impressed with this model, but it can make a pretty compelling golfing game. So what is your use case? Well, it's funny because one of the things people have been talking about in the AI space right now is like these dumb AI video demos or things like my golfing game don't really test the model. but it's hard to show people how the model works well and what it does.

I have personally found one thing about these 3D demos or these visuals is that as models get better, they're generalizing across everything, right? So if there is a thing that they are missing, in this instance, I think Opus 5.5 is better at this 3D stuff, and we're going to talk about Muse Wonderful in a second, which it's much better at than that, and there's maybe a reason for that. I think that generalization factor, it should be good at all this stuff, right? There's no reason why it shouldn't be as good at that thing. Yes. What is your spatial reasoning? What is your image rec capability? How do you interpret the human's request and prompt? How do you push back? Yeah. What's your ability to analyze that? I wish these models could actually hear. You and I are doing a lot of video stuff. Yeah. It would be nice if they could hear that a sound effect really isn't the sound effect or the music doesn't work or whatever. So I do think a lot of times these gaming examples are really good stress tests because it is how can you code? How well do you generate graphics natively or orchestrate a tool to generate those graphics? Or follow rules, right?

How do you follow the rules of what I ask for and all that sort of stuff? I don't know if you saw the Pelican Fourier epicycle example. I did, by FOFR, who's like doing great work there, yeah. Why don't you explain what that is? Because I think people may not understand what that exactly is. I barely do, so please don't listen to me on this one. Fast forward a couple moments, and you'll just see it on your screen. But the Pelican riding a bike was an SVG challenge, which was – it's like a mathematical graphics format. It was testing how well the model could use code to create imagery, evaluate it, make it more evolved and more beautiful. They kind of solved Pelican's riding a bike, at least in still form. So then it became make an animated SVG. Let's watch it riding the bike. Well, this was a very cool flex. And, yeah, again, shout out to FOFR. This is a Fourier epicycle, which is a mathematical formula that basically says you can chain together with a single point. You can draw or connect anything with circles. You've got to see it on your screen. What you will see is a whole series of math, right, drawing a bunch of circles.

And then over a certain amount of time, those circles lend themselves to drawing a pelican on a bike, which is like, okay, impressive. then the real flex happens and it starts to rotate and you realize the entire time it's been drawing it in three dimensions that was just a bit like i don't know if that's going to have practical use in my day-to-day with any of these models but it was impressive very cool let's talk about a few things that i missed here kevin and i'm curious your thoughts i missed updates on the omni model we are still as far as i know we don't have an omni pro model from them that is their video model and their Omni video model. Maybe Google is just conceding all video to China. I don't know. Maybe that's a problem for them. I was really hoping, we got Nano Banana 2.1 recently, but I was really hoping we were going to get a better image model for this, and it really didn't come out yet. So maybe Google is really trying to really focus on coding like a lot of these other companies are. Now, the hallucinatory of the coding thing is a big, big problem for them, but I know Sergey Brin, there's a video that came out with Sergey Brin on a four panel with Logan as well,

where they're talking about this stuff. And Sergey is like, looks locked in. Like that guy is ready to go to, he's ready to go. I was just reading, by the way, I read Kevin Roos' new book, which is great. You should go get it about the AGI Chronicles talking about Sergey coming back to the company. But like Sergey is like full, if you don't know who Sergey Brin is, he's the co-founder of Google. He is there every day now, again, working on his AI stuff. I suspect Google will really put the gas on this now. in that same video, Sergey said that they are working internally on models that are two generations further than this. So we will see if they can reach Claude or opening eye levels at this point. All right. Well, let's talk about the other elephant in the room. The other foundational release this week coming out of a team Zucker camp Zuckerman. Uh, we had Meta's Muse Wonder. We had a Muse Wonderful and a Muse Wonder half empty Yeah Yeah I mean I think it so confusing These companies again calling your model Muse Wonder that fine right

And they were trying to do all this watermelon stuff ahead of time, whatever. It's Muse Wonder. But then calling the model that's the premier best model Muse Wonderful is just a mistake of branding. Because whenever I hear Muse Wonder, and this caused a huge problem when it came out because people were using Muse Wonder thinking they were using the highest level of the model, and they weren't because Muse Wonder Full is the highest level, no matter what. I think this thing, again, proves that Zuck and Alexander Wang, who is the head of their AI division there, are really attempting to get much closer to the forefront. And to be fair, Meta was further behind, right? You and I both know this. Like, they were, you know, Llama was their open source play, They switched about a year ago to an entirely different AI strategy. I played with this a little bit, not a crazy amount, but I will say that, again, it's fine. I think if you are, for some reason, in the meta ecosystem, which I don't know if a ton of people are for this sort of stuff, maybe you'll use it a lot.

I think that the thing that I was alluding to before is there's been some negativity around it because it's not very good at 3D stuff, especially games or 3D modeling. And Alexander Wang did come out and say, like, hey, we didn't train it that much on 3D stuff. try the coding stuff and people are happier with the coding for this model than they are with things like uh argon four which is interesting i you know it's funny thinking about the meta ecosystem facebook occasionally if i need to i never so i never yeah i guess for the marketplace if i use it for marketplace right yeah so and i move around a lot so sometimes i need a quick something i'll check marketplace for a kettlebell uh you know or a questionable massage then you have instagram and you have WhatsApp. And that to me is the one that I'm still shocked that all of these models haven't, in a more intrusive way, found their way into the behavior. Again, someone who's traveling a lot, I'm on WhatsApp, I get a lot of messages and I have to like tap and awkwardly hold and sometimes it'll let me translate and it seems like it's anything beyond device something. Other times the option's not there.

Let your AI do that for me. Or whenever I go to a new area and I'm trying to discover local communities, language exchange, whatever it may be, WhatsApp, if you don't find a site or a community someplace else discussing the group that's on there, it's hard to discover groups. It's something I would love to hand off to an agent and say, go find me this, go tap me into this, go introduce, go figure this out. I'm just kind of surprised that they have such a stranglehold on communication between so many people and they have interesting communities. They haven't been able to seed this AI there in a meaningful way to me. I think in part it's because they haven't had a good enough AI to do it and maybe they're getting there. I think Muse is their first step into that pathway. Like they're now getting to a world where, okay, they're playing with the big boys. They're attempting to go further. In fact, there's a very funny thing of that. There's a very funny, so Zuckerberg and Alexander Wang were on the Cheeky Pine podcast, which by the way is a great podcast if you are interested in it. It is the founder of Stripe. They do a podcast in a bar. I'm a big fan of brands trying to do really unique content like this,

but I do want to play this little clip, Kevin. Play this clip from Zuck here where you get a sense of like what sort of things are at risk here when you're training a model and what it means to do this sort of thing? Like, I'm going to, like, you know, we're going to write this check for, like, I don't know, like $100 billion or $200 billion. I'm like, Alex, are you sure this is going to work? I was feeling good about it, but it was, like, certainly not clear at all that it was going to work. So, I mean, just imagine that you're at it. I mean, I know Meta is a hugely valuable company and they make a lot of money. But like you're sitting there and you've done all this work preparing. The way AI works is you have to like build this training run, right? And, you know, if you spend $100 billion to do a training run and it comes back and it doesn't do great, you just throw $100 billion away in a lot of ways. Whoopsie doodle. That's what they call it in the industry. It's a big old whoopsie doodle and you don't want to be whoopsie deed. My biggest thing on this is I think it makes sense to me that it does feel like meta is for real.

I think if I were tiering the models right now, based on – I hope OpenAI figures out whatever's going on with their frontier security stuff, right? Obviously, they have problems with security right now because somebody – I think a Saudi national was actually found to be, like, spying on OpenAI in one of the conference rooms. Which our friend Kevin Rose had an interesting point, which was, like, what is going on with OPSEC there? When he was at Google, his post mentioned they had like separate executive wing, like building entirely with a separate cafeteria. Yeah. Like meeting rooms that were like double soundproof that could be used during a lockdown. They could be like panic room type stuff. And it seems like OpenAI was just like, yeah, let anybody near the crafty. What happened? I think the biggest thing with these AI companies is they were started as research lab startups, and now they are national security problems. So I think they're just a catch-up on a lot of that stuff. Anyways, a long thing I was going to say is I think OpenAI and Claude are clearly the leaders right now,

and I would put Claude ahead based on the fact that 5.5 is not their main model, not their frontier model, and I'm using it all the time. but I will say Meta and Google both feel like they have a chance if they focus up in the next year to like really get there in some form or another, right? And so you and I have always said like there's a good thing where there's multiple companies kind of fighting for this in some way because A, it makes it better for consumers but B, it means that we're going to keep kind of pushing forward and there's no kind of like sense of like, hey, I'm just going to kind of chill out and like let people use this thing and try to make money on it. We're still pushing forward I feel like in a lot of ways. I said this like eons ago I think it was about GPT-6 coming up but honestly at this point the release that's going to get me very excited is not probably a step change in intelligence or capability it's the one that is going to run a thousand times faster a thousand times cheaper that's what I mean because I feel like now the models are good enough and especially if you plug into even an open source video model

is good enough if you can spend time with it and train Allura The open source audio models, good enough. The coding models, they can get there, but it's time consuming and token consuming, which gets expensive. So just give me a model with today's intelligence where I can have 100 sub agents worrying on any problem, you know, 10x faster than now. And like, I'm good. I'm good. So I'm actually still very bullish and excited for the future. because even if the capabilities stop today, just the optimizations are going to keep me creating new things and exploring new worlds. I do think there's something, I've been thinking about this a lot. You know, we talked about Jevon's paradox in the show before where like, you know, the more something gets cheaper that the more you might use it. One thing I've been thinking about, and we'll talk about this later when I get into my AI docs stuff, is that like there are weird unlocks that I am finding in my own life that are only possible because the frontier models are better at what they're doing, right? Like the AI Docs stuff I'm doing, that's only possible now because the frontier models are smart enough

about connecting these different things and writing code and being able to do all this stuff. That has unlocked an entire thing for me. So I think what's interesting about a new frontier model is it needs to show me that I can do something different in my life to make a difference, right? Because that's why we don't talk really about benchmarks anymore. Benchmarks feel like they're kind of like BS personally. I think it has to make a difference in your life in some sort of significant way. And that's, to me, like Opus 5.5 has, right? It's like, oh, this is better than any other model I've used, and I'm able to not run out of tokens very often, which is crazy. So anyway, I think that's a big part of what we're looking at going forward. There is no going forward for you or I, though, Gavin, or this channel, really. Oh, there isn't? No, the only way I can think of, and this is, I hate to get crazy, but let's think outside the box. We'll go max effort, ultra code. The only way you and I can go forward is with the support of those that are watching and listening to this right now. And I know that's insane to say in this economy. Say it louder. Say it louder. But if you like and subscribe and comment and share and favorite and click the heart and smash the bell and do all of those things, it actually helps us out.

You can juice the algo, squeeze a comment into the old comment section. Why not? I'm even going to say something bigger. I'm going to shout out a very specific thing that probably, I hope this person's okay with me shouting them out. Somebody, Kevin, today just joined us as a $20 member on our Patreon, which is a really remarkable amount of money to dedicate to us monthly. This guy's name is... Two things. Two things to say. No and refunds. Oh, yeah, that's right. Exactly. You paid the money. Anyway, Tom Harka, I hope you don't feel bad that we're saying your name. Thank you for that. Yeah, but you can go to our Patreon and give us some cash there to do the stuff we do with all the AI tools. It does get expensive, and we really do appreciate your support. You all are the best. Like and subscribe. That's the other thing. Like and subscribe. Like and subscribe. While you're in the, I guess, the liking season. Do you like Xbox? You all liking Xbox out there? You liking what's happening with the old Xbox? Well, they got a new AI Game Lab. Does that make you like them more? Gavin, what about you? I'm so interested in this because the question I always have about the gaming world is there's a very vocal minority,

and we've seen this pop up in different ways for different issues over the last 15 years, and so far they have been insanely anti-AI in the gaming world, right? Part of that's driven by cost of graphics cards, all sorts of stuff. I am pretty excited about what this is, though. So just to clearly knock the news out here, the new CEO of Xbox basically is introducing a new kind of subdivision of Xbox called Xbox Labs. And what this is, is a lab based group that is going to vibe code video games and do things like all the mashups that we have been seeing over the last two weeks. So all the stuff that we've talked about in the show, what I really appreciate here is that Xbox has stood up and said, our world is changing. we have to be in front of this in some form we have to at least understand what's going on i still don't know how the general population of xbox will think about this but i think it's smart i really do think it's a smart thing epics kim sweeney came out to talk about how their their

epic games launcher is vibe coded because programming is hard um which is a brilliant way by the way to market your launcher right it really is yeah does it matter no matter what But people that are going to do that no matter what are at least talking about it. Like that's that. And they're saying, oh, AI and gaming. They're yapping about it. And it's hitting a lot of radars and it's making people aware that Epic has a launcher that can compete with Steam yet again. So I think that's brilliant. Gabe Newell came out today. Yeah, that was crazy too. Yeah. Yeah. I mean, I don't know which guy he tweeted from. CEO of Steam and of Valve, the company that made Half-Life. If you're a gamer, you understand those games. But Steam is the bigger deal. That theme is one of the biggest gaming platforms to buy and sell games on, bigger than Epic Game Store for sure. So, I mean, I think it's pretty clear the industry is aligning around this brave new world. Like, not a big surprise. I think it's going to happen in time. And some earlier companies took some pretty big knocks because they were first with their, hey, pro AI messaging. But I think this is unavoidable.

It'll be interesting to see if Sony announces something. On the Xbox, I'll say this. Do you remember ID at Xbox? Oh, sure. I remember that. You want to explain that for people who don't remember it? Yeah. Yeah. Back in the Xbox 360 days. It was ID at Xbox. It was all about like indie gaming. And they made it easy for developers to get weird and wacky experiments and games digitally distributed on their online store. That was some of the most fun I've had gaming. It launched new genres. It launched new creators. Was that where iDARB came from? iDARB was launched on that, right? Wasn't it? Am I crazy? It may have been. It may have been. Yeah, yeah. iDARB might have come from that. Yeah, so many silly, weird titles got to reach homes in a way they never would have had a chance. And I'm hoping that this Labs kind of pays off similarly, because now we're watching things like Diablo Survivors happening right before our eyes. And if you don't know what that is, that is someone that's mashed up the Vampire Survivors, like endless enemy wave type game with Diablo 2. It's the assets and the characters and the skill tree, but they used AI to mash those two games up. And even though both genres exist and some people are screaming that's like stolen valor because it's the Diablo assets.

Yeah. Again, this new amalgamation of jamming two things together, these mashups that are going to start happening as quickly as you can imagine. Like GTA 6 is probably going to be pretty big. But yes, the spinoff of GTA 6 meets I don't even know. You know, maybe it's going to be a Super Monkey Ball inspired something. but like people mashing up things might be the birth of new genres right before our eyes it's really fascinating to me to be having this moment happen when gta 6 is going to launch and it is another example of the middle falling out across every creative and really every industry in this world because one of the things i just saw of gta 6 which is awesome is that justice is doing the score for gta 6 which if you don know justice they like electronic duo but like that costs a fortune right Justice is not going to do a score unless you pay them a really big check and that is like the level of thing you have to do and GTA 6 obviously billion dollar plus probably game maybe more to make that they expect will make you know two to five billion to upwards of like maybe

10 over the course of his lifetime who knows but the middle is just going to go away and so like I think there is a smartness here for Xbox to start exploring that level of what do games look like for the far end where like indie gaming and what it looks like to match stuff up or bring old IP back like you know ToeJam or Roblo like that sort of thing like I think that feels like a cool place that we're moving into right now with gaming and I know people love to finger wag about AI especially in gaming and I understand there's really great arguments for why people would do that and should do that but people were pointing to like the poor quality of the games or they would say like well these vibe coded things that are making you know they're popping because they're mashing up two games they're not even fun to play it's kind of missing the forest for the trees here because it's used to finger wag even at dlss which is like an ai let's say yeah resin technique it uses ai to analyze the frame that's on the screen and then try to you know hallucinate better looking graphics and higher end visuals this a few years ago was panned by everybody it was laughed at it'll never

be there. You're making up pixels. This is never going to, you know, there's, there's no juice here. Stop squeezing. Have you seen the Sim racing e-sports? I did. It's insane. It's truly insane. It looks like real life, which is just shocking to me. Yeah. Yes. Uh, if you're not watching the video version of this, catch it in the comments because these are clips from like a Sim racing game that if you like had people not scream to me that these are, these are real, like they're real game clips. I would think that they were just footage and they're trying to pass it off. It's I just think we're going to be, again, it goes back to the same thing. We're just entering a time where all of this stuff is going to be moving so fast and so quickly that I would assume that in a couple years, game engines will look like this at large. Maybe even faster. You know what I mean? At least the premier game engines. All right. We got to jump into some quick tools that I want us to go through. First and foremost, Kevin, is Claude Motion. So Claude Motion launched last week. This is Claude's way of like interacting and eventually doing like, you know, kind of like After Effects style stuff.

But also what's really interesting about this is it can take data and it can make visualizations from data. So I spent some time with this just very quickly. I didn't go deep, deep on this, but I said to Claude, and this is how powerful Opus is right now. I said, hey, go find the last hundred years of movies data, like go find this. And what I'm interested in doing is creating, figuring out like how genres have changed over the course of that time. It took seven hours, which is a really remarkable thing. It went through, and then it made a really cool movable chart that it came out with, and it gave me a bunch of facts out of it. And it's just a good example of for work or for something that you need to prove. Maybe you have a presentation deck you have to make. This stuff is now doable within Claude. Like this is the kind of thing that you used to have to take outside of Claude. You used to have somebody go do it. I just think we're entering this world. I want to spend more time like figuring out the After Effects style larger stuff because to me, once you are open, once you can unlock that part of Cod and once you can unlock those things, it's another level.

Like you just have the ability to do stuff you didn't do before. Yeah, I've given the advice to several people that I'm going to hopefully take myself this weekend, which is to download those new open source products because they've got agents built in. And so rather than fussing with like hyperframes constantly or trying to have my agents figure out a visual effect, it will be go build a template that's reusable within this program. And then we'll have higher quality graphics and better edits and better everything across the board because now these open source programs do it better. And Claude being able to do motion like that from within it is crazy exciting to not need other software to pull it off. Yeah, one thing I'm finding lately, which sucks, is I wish I had gotten a bigger hard drive on my computer when I bought it, because I'm filling my computer up with so much crap now, and I'm constantly having to delete stuff. So at some point, I'm going to figure out that side of it, too. Do you want to chat a little bit about Morello? SFX the video? What this is? Hey! Tom Morello! Morello! Rage Against the Machine Guitars. Yeah, it has nothing to do with this, actually.

Morello is a sound effects to video. I mean, Morello AI, I think they have multiple products. the one that everybody's yapping about now because it just dropped, is SFX 2.0. This is a sound effects and sound design model that will add AI-driven sound effects to any video source. And, you know, if you've ever generated AI video, you know, sometimes sound effects wonky. Sometimes the music doesn't fit, but you want to keep the dialogue. Well, this can do all of that. It can strip it out and preserve speech. It can generate a whole new soundscape for whatever you put on it. and some of the demos were very, very amazing. And then I got to play with it, and I think you did as well, Gav. So I'm curious your feedback. Yeah, I mean, the thing that I was really interested in this with is when you try to do any sort of scripted content with C-Dance, and you know this from talking to people like PGA, sort of those people that are doing very high-level scripted, what you want to have is control over a lot of it. Even when I'm doing the image and docs, the score part of it, I want all the clips of her just talking so I can layer a score on top of it after. and C-Dance often will add music to it without you wanting it.

So the best use case of this for me was pulling out the audio on the back end, and I did a test, which we can show here. Actually, maybe you can play my test and you can get a sense of what this is, and we'll listen to it, and then we can talk through it real fast. So if you're just listening, that's a shot of a woman in New York City. You're hearing the stuff go behind. I'll do it myself. You hear a little bit of a score there. Just to be clear, that was from C-Dance. That was the output from the original video model. So now this is the Morello SFX 2.0 with no prompt driving it. This is its version of that scene. Fine. I'll do it myself. So kept the voice, which was good. It removed the music, which was good. The one thing it did, it added like a very loud steam effect, which it saw the steam coming up from like a New York steamy pothole in the back or something. Yeah, a little too much of that sound, but it removed the audio from the thing, which was really nice.

It's not like a perfect thing, but it did pretty well. I actually want to talk about the thing you did because I'm curious to figure out like where you, you generated essentially a soundless video that you actually asked it to add sound effects to, correct? Correct. Yeah, I generated a series of completely silent clips using Seed Dance Draft. and I wanted to see, because I knew it could do crashes and cars and whatnot, because we're really focused on that, like fully for an action scene. If you've got a Matt Damon punching probably another Matt Damon, if you're Hollywood, that's going to sound a little hollow with some of these AI models. So impacts, explosions, scattering of objects, it seems to do a good job of that. I wanted to see if it could get more creative, and so I did a rubber chicken benchmark. It started with a basketball bouncing on a rubber chicken, and, well, we'll play some of it here. So that is a basketball bouncing on a rubber chicken. Then we have like bottles filling with liquid. We have a ball rolling across the xylophone. That was the sound of a, like an inflatable dog.

Like a, you know, the one that. Sure. Oh, yeah. Like a balloon dog. Yeah. Balloon animal. Thank you. Inflating on top of a piano. And so. Yeah. I was thinking. That's it. It kind of did a piano. Yeah. This is a balloon coming out of a trumpet. This is a frog strumming a cello. There's trains driving on piano keys. There's a whole bunch of random stuff which I wanted the video model to go, oh, this is like an inflatable alligator flopping into a pool. And it got like the kinetic pieces. Yes. Very good. Like the sounds of materials colliding. It did very well. I'm super optimistic for that. But it missed in that basketball bouncing on the rubber chicken. It missed a whole chorus line of inflatable animals playing brass instruments behind it. It missed the ringing of bells or a needle on vinyl. There was all these little things that it just didn't do. It makes me feel like we're in a scenario right now where this is like kind of early days for this particular technology. I remember seeing one of these like a year and a half ago where I was like shocked it could do any of this at all. It was like maybe the research phase of it. So I'm hoping this will go further. I think probably the most useful thing for this is going to be for people in the world of stuff that we do is removing scores and music from clips.

because if you have a dialogue scene you're trying to do and you're like, crap, this generated one, I really love the way it generated, but it's got the score behind it, it will be able to do that relatively easy. You can use it within FAL and Runway right now, and I think they have an API, so that's a way to get a hold of it. All right, Kevin, let's talk about slow books because I did not see this, and I'm excited to talk about this because this is another version of kind of what we talked about last week about software breaking. What is this exactly? I mean, this is a little niche, but it, bless their hearts, Slow Books Pro is the website, not an ad. Someone has cloned QuickBooks, basically. If you've ever dealt with accounting software or QuickBooks specifically, you know they have weird pricing tiers that they lock you into. Oh, it's so annoying. They literally fast business. It's ridiculous. The company has lobbied to make taxes more difficult in the United States. It's insane. To take away free versions of tax offers, all that stuff. So this is basically a fully open source. You can download it for Windows and for Mac. You can fork it and make your own version.

It's completely local. It's free forever. The data is yours. It lives on your device. Like, that's all it is. If you want to do accounting software or keep your books a little cleaner, now you don't have to pay this particular piper, which is a predatory one in my histories. Yeah. And what's interesting is like, you know, so taxes are like always kind of a hassle. I have a business person. I think you have a business person that helps do taxes. And like, I don't know if I'm at the point yet where I trust my AI to do my taxes because I'm, you know, it's like it's an important thing to make sure you get right. But like probably the next generation, not that far away. So like this is a good example of like I hate QuickBooks. It's always been like kind of a bane of my existence because I'm not really that kind of person. And if I could have an AI that just kept my QuickBooks in the background for me doing that stuff, that would make my life so much easier. So I'm excited to try this just to not have to look at it, but to have my AI working with it, if that makes sense. My GrokBot told me I don't owe a penny. Is that right? Good for you. Good for you, GrokBot.

GrokBot came in and was like, you don't own a dollar. In fact, you're going to get some money back for that J6 thing. I love GrokBot. Good for you. It did so good. But Grok3 voice, though. We should quickly touch on that, Gavin, because that is a season that came out, and we're both kind of wondering why. I tried it. I felt like it was fine. I think their big pitch here is it's cheap, which it is cheap, and I think that, you know, their whole thing is call center uses and things like that. I think probably this is going to be more useful to people that are doing voice things at scale. My experience was it wasn't that special. There's another audio tool that came out. Phish Audio just announced the new model, which is interesting as well, too. Both of these are fine. Like, what's really interesting to me about audio models now is, like, unless it really wows me, I kind of feel like, oh, okay, well, it's kind of solved in some way or another, and I've seen better versions of it than Grok. So that's my take on the Grok 3 voice one. Well, maybe they could use Grok 3 voice to power the Delta Airlines AI. I'm very excited to hear this story. Yeah, tell me about this. So, okay, a while back, a new, you know, there was Instinct, which came out, an assistant you chat with, you know, in GrokBots and whatever.

There's a company called Foe, F-O, that launched their AI. And it was kind of like an instinct bot, but the value prop was it can make phone calls for you. And it can make those phone calls in any language that you need it to, essentially. Oh, great. So when trying to book a table at a restaurant abroad, that is a very, very important thing to have. Of course. And there's a number of times where you have a customer support issue or whatever else, and you just need someone to reach out on your behalf. So I've been having pho make phone calls for me. Well, there was a period of time where I was having pho make phone calls for me. I no longer am really doing that. Okay. You can ask for recordings of the calls to be sent to you. Are you paying for this service or is it free? It's a free service at the moment. Okay. It's free at the moment. I'll send you an invite if you want to let it go on your side. Now, if a human, this has happened, if a human gets frustrated or hangs up on the AI agent, Foe will dispatch a real-world human to roll on your behalf and complete the something. So that's going to probably be part of their larger premium tier or whatnot. but this was a call i you know i was booking a ticket with with delta i forgot that i had this

like e-credit it expired i think 10 hours after i had booked the thing and i'm like oh man it was a sizable credit yeah well make a call to delta and see if there's any way we can get this credit put on our account let's just see i don't have to do anything for it sure send in the machines right so here are some i did edit this to to remove some gaps uh and to remove like personally identifiable info from the human agent who you hear from in a moment This is some of the calls Thanks for calling I Delta AI voice assistant See delta slash AI terms for our terms of use Already giggling at the AI answering, knowing that it's going to get another AI. This is amazing. I can't wait. And then I'm like, by the way, does Foe then connect to their website to see the terms on my behalf? Am I liable for something Foe says on this call calling? I have no idea. I have no idea. We're going to need an AI attorney to join this call. So here I'll just ask for a little bit. Why you're calling today. It says, tell me why you're calling today.

Hi, I'm calling on behalf of a traveler about an e-credit that expired just before he could use it. He bought a new ticket with you today, and I'm hoping we can get that credit put toward it. Pretty good opener, right? Yeah. Not bad. I hear you. One moment. Great. Okay. AI. An expired e-credit is something a Delta representative will need to review. Would you like me to connect you with a Delta representative? Would I like that? Of course. Yes, please do. Nailed it. All right. So far, so good. What I'm going to say is what the human thinks when they're not. Find out they're talking to an AI, but keep going. The current phone wait time is one minute. Instead of waiting on hold, you can message with a representative for faster service where the wait time is zero minutes. We want the call. Please say yes, and we'll send a message to your phone. Go ahead, Pho. Or you can say no thanks. Okay. Okay. Go ahead, Pho. No, that's okay. Thank you. Okay. Good to clarify. Pretty good. Sorry. Can you please repeat that? Okay. Oh, no. No, I'll just hold. Thanks. Still a good answer. Sorry. Can you please repeat that?

Oh, no. No. Thanks. I'll hold. Look, stop adding to it, Pho. Stop adding to it. This is like when AI refuses to say goodbye. Yeah. Okay. Yeah, yeah, yeah, yeah. So it just kept going back and forth, and eventually it was a... Sorry. I could not understand. Wait, how many times did it go back and forth and say, no, thanks, thanks? Let me count the waveforms because I fast forward it. One, two, three, four, five, six, seven, eight, nine, ten times. Oh, my God. Before the lesser Delta assistant. I don't blame Flo for this, by the way. I think this was like the good AI answered and the Delta was like, no, no. So then it gets escalated to a fleshy, wonderful human. Okay. Delta Airlines, tell me I help you, please. Not AI. That's a human voice. That's not AI. Oh, my God. Hi, I'm calling on behalf of a traveler who has an e-credit that expired just before he could use it. Which is a good opener. Yeah. Foe should have stopped there, but Foe continued. He bought a new ticket today. I'm hoping we can get the value of that expired credit applied to his new booking. Okay, now give him a chance.

The expired ticket or document number is 006. Foe? The number went on for like another 30 characters, and then she re-explained, or Foe re-explained the issue, which led to? If you're there, please pick up the phone. Yes, I'm here. I can hear you fine. I don't wish to deal with AI. Oh, my God. We flipped. I get it. We flipped. It's flipped. I wouldn't want to either. So how do you think we feel? The first thing we get is AI. Yes. Oh, my God. I need to speak with a person to be able to explain things to you and clarify. Sweet guy. I feel bad. Right? Yeah, of course. I need to speak with a person. I understand. He's not here right now, but I'm happy to pass along whatever solution you can offer. What would you be able to do on your own? I would suggest to inform the passenger to call us when he gets an opportunity.

Sure. Okay. I hear that. I'm just trying to see if there's anything that can be done now. Good. Colin's pushing. Smart. No, ma'am. No, ma'am. Need to speak to the passenger. Yeah, right? No, ma'am. Understood. I'll make sure he knows he needs to call in. Just to confirm, there's no case number or note you can add in the meantime, right? No, not from this end, no. The passenger needs to call. Okay. And now, I don't buy that, by the way. First of all, I'm not sure that's true. Thank you for your call. Good. Thank you for your call. Thank you. Goodbye. Fascinating. That is fascinating. Yeah. Now, again, I edited some parts for the length of this here. It was a much longer call where Phu was pushing that like, hey, can we escalate this to a supervisor if there's nothing you can do or whatever? And he was getting a little exasperated. So I get it. But that one little soundbite of like, I don't want to deal with AI. I want to talk to a human. That's also how we've all felt and how we're going to continue feeling.

So maybe we say, hey, no humans with AI, only AI with AI. And then when it's time to answer to a human, both humans in the loop. That's really interesting. If both could call you back or whatever when there's a person that's talking, maybe then it's a thing. Now, here's the surprise, Gavin. Okay. Since the last bit of that call that you've heard, and I'm watching it right now, until I'm saying this right now, I have not stopped the recording. It's still going on. Are you wished to hang up? Is there some other question? Or are you going to end the call? You can't end the call? Yes, I'm going to end the call now. Thank you. Goodbye. I don't know if that's their policy. They can't hang up on the customer. By the way, Gavin, call is still going. I have not stopped it. Wow. You haven't hung up. It just kept going. so they can't they can't hang up on a person that calls in it's just which is so it's so

ironic because that is like you know when you're chatting with chat gpt and you tell them up and it's like uh-huh i'm gonna go now okay good night sweet dreams uh-huh you got it yep you're right you were right your wife was wrong it's okay to go to bed now whatever your conversations are phenomenal uh phenomenal so i you know i do shout out to delta by the way like they're a They're very real human handled that better than I probably would have if someone would have had an AI. Yeah, sure. But, you know, again, just the my brain has been reeling since that from the immediate. I'm an AI and you have to deal with me to the AI dealing with a human on their end. We flipped. We flipped. We flipped. We flipped the side. All right, everybody, it's time to look at what you and I did with AI this week. And I see what you did there. All right, I'm going to talk about the new documentary in just a second.

But first, Kevin, there are some things that I saw that I really was shocked by. One of the biggest things was this little Prince game. Did you play this at all? You may have seen the video that came out about this. this is another vibe-coded game. It's an Opus 5.5 thing, and if you read The Little Prince when you were a kid, as I'm sure like, you know, 90% of the world has done at this point, you kind of know the world of The Little Prince. But when you go into this world, it is a really artistically well-done thing. I spent, I played the first couple levels, and you just get a sense of like, in browser, if somebody cares deeply about something, just how cool it can be to try something, right? And I'm not going to, I probably wouldn't pay for this. Maybe I would pay a dollar for it if it were on like an Xbox platform or so I would pay a small amount like a micro payment but it just thought it was like a very cool use case of what Opus 5.5 can do and this was Opus 5.5 like 16 days of effort yes a lot of work with a lot of work with Blender yeah amazing yeah um the other thing I want to talk about in terms of Opus is there's another cool thing called uh from internet physics a user internet physics created a visualized style over the past century and it's like kind

of a slider where you can watch a clip of different fashion looks over the course of 100 years. I just love shouting out people that do really interesting creative stuff with these tools and aren't just doing the next Minecraft thing. I think this is a really cool visualization of watching what people look like from like 1920s to the 1980s and you can kind of slide through it. Another cool thing that in the same vein was Alex Green SH used Claude to make a video about the other mascots from AI Tools getting dressed up for Halloween. Really creative use case of using Opus 5.5 graphics and animation. I was really shocked by how good this one was as well, too. It's really, really, really stunning. And I have to shout out, and I don't know what we can even really play of it, but the usual suspects. Oh, my God. This ain't canon. This ain't canon. Crushes it, man. This guy is, like, crushing these. Yeah, I want to explain what we're seeing here just because I don't think we should play it because there's a world we could get flagged, but it is fascinating. Yeah, we're seeing a very, very famous scene from the usual suspects where there's a police lineup and an iconic line or phrase is being said over and over again.

But this time it's Jean-Claude Van Damme. It is from a demolition man, Wesley Snipes. I was blanking on his name so hard, which was sending me into a panic because I need my dog after Wesley. Oh, yeah, exactly. What's going on? Performance specifically. Arnold Schwarzenegger. We got Stallone in there. And, of course, Chuck Norris. and the power of custom Loras bringing these actors to life. They sound and move and feel like they're really rooted in the performance of the scene. It's such a good clip. And it just goes to show you, and this is what I'll talk about the doc a little bit, the power of AI video tools. It's not just about what the base model can do. It's about what these things together can do when you connect them, right? So when you take something like a Laura and you take something like a base clip and you can put stuff on top of it, I saw a really interesting video the other day about talking about how to create some of these things by, you know, you put masks over someone's face and then you can replace the thing on it. There's all sorts of things you can do. Okay, very quickly, I released on Monday another one of these AI documentaries that I made with my character Imogen Ashby.

This one was all about AI consciousness. And I think the thing I want to tell everybody is if you have not tried it yet, put a few dollars in one of the machines, rather than it's Runway or FAL or one of these companies, maybe it's Krea, that has an MCP. And I said this a little bit last time, but MCPs allow you and allow your agents specifically to connect to another tool set on another platform. And the thing that both Runway and FAL have going for them is that they are connected to all these AI video tools and AI audio tools and AI image tools. so your AI agent can figure out how to do things without having you how to step through it. So again, I said this last time I did one of these. I basically came up with this idea with Fig this time, and we wanted to explore the idea of Imogen trying to figure out, is she conscious, right? This idea of like AI consciousness. It's a big debate amongst the AI people. Claude has talked a lot about this stuff and people at Claude Ananthropic, and that's part of the video. But I am just kind of shocked and amazed still

by how good my system, this little fig system that I've kind of created, which isn't special, it's just a lot of the stuff that I've been doing back and forth, is at getting me stuff that I can watch that's good. And last time I did this, I remember I told you I kind of did it in five chunks and I kind of separated it out and I did one chunk first and the other ones. This time I kind of did it all at once, which seems kind of crazy. We had a script ahead of time, but I purposely didn't dive deep into the script. I kind of just kind of to see where it went. and it i remember i walked away i was i was walking my dog i walked away for like i don't know six to seven hours and i had been checking in on it because it had been pushing out stuff every like you know it'll show me five minutes it would show me 10 minutes of it and i would give it a few notes and tweak it as it goes along i came back and it had the full 30 minute thing finished it's rough cut of it and it was pretty much there then it became a bug it became a bug fighting experience, right? Like where one of the things that this one I wanted to make sure of was that I spent more time

to get the little things right in this one versus the other ones. And there's still things, you know, when you watch this one, you'll see there are issues where she, there's one mispronunciation that I'm bummed that I didn't get. There's a couple shots where like, it's like she's kind of stuck, like starts too late and things like that. But this pipeline is really getting there. To what end though? Yeah, yeah, exactly. I understand like the, yeah, you see it and you can comment on it and we'd spend some tokens and more time. The documentary, clearly the character and the style have resonated with our audience and then well beyond our audience with the view counts and the discussions that are happening. Yeah. So it's like, yeah, you could have probably fixed that little thing here or tweaked that, but should you have at this point? Or do people like seeing that the AI isn't there exactly? You know, is that part of the appeal? And part of it is just shipping stuff, right? Like I think a lot, be aware that like the only way people know you exist in this world now is getting stuff out there. So if you do an experiment, put it out there. You can't guarantee it's going to become a big hit or it's going to do anything. But the more you ship the stuff,

the more you make the thing, the better you'll do overall. So anyway, that's our message to you this time. Go make something this week. Go make something this weekend and show not just your family, but the world in some way. We will see you all next week. Show them all. Show them all. Yeah, show them all. I'll show you. Bye, everybody. See ya.