The Next Wave - Seedance 2.0 Is Here… and It’s Better Than Sora & Veo

Hey, welcome to the Next Wave podcast. I'm back with returning guest and one of my best friends in the world, Joe Fear. He is our boots on the ground marketing business expert who's also paying attention to AI. We're going to do a very similar episode to what we did last week where we're going to break down the five most interesting things that happened in the world of AI last week. We're going to riff on it, we're going to talk about what we think is interesting about it, why we think you should be paying attention to about it, test some of these tools, you know, figure out life's greatest problems, solve the three body problem, figure out the mysteries of a life. By the end of this episode, we're going to have done all of that. So stick around and let's just, you know, get into it. Thanks for joining us, Joe.

Hey, thanks, Matt. It's great to be here again. Let's go.

Yeah, we got some really, really good feedback on the last episode that Joe and I did together and we figured, hey, let's do some more of these types of things. So here we are and let's just jump into the first story this week. So, Joe, I don't know if you came across this at all or saw it in your X feed, but There's a new AI video model this week called Seed Dance 2.0.

I did. I mean, like, I'm always seeing these videos. I'm like, I don't even know what to believe anymore. But now more than ever. Yeah, these videos, they're nuts. I know you try to get your hands on it really quickly. Right. But not so, not so able to do that here, at least in the US currently, right?

Yeah. No. The Chinese company behind this is a company called ByteDance. They're also the company that makes TikTok and Cap Cut. And I feel like they wanted to, they made it so that only non US countries got access to it early. So supposedly it's going to be accessible worldwide on February 24th is, is the, the thing that I've been hearing. I don't know where that number comes from or where that date comes from, but that's the date that's been circulating is February 24th is when we get access to it. But some people did find some loopholes and figure out how to get early access to it. Unfortunately, most of those loopholes have been closed. So we can't actually demo this on this episode.

Well, I mean, we could, we could show some really cool. I know on their website and Obviously we could probably stalk some people on X too to show what, what they made. Do you have any pulled up there, Matt, that we can just like get a little taste?

Well, they do have some on their website. I'm never a huge fan of just showing the ones they put on the website because you know there's going to be like the most cherry picked results. These are the ones that they were like, okay, that's really impressive. Let's show that one where maybe they had 15 results that they didn't put on the website because they're like, oh yeah, let's not do that one. But there's, there's a few that we can look at on their website here. Like, let's see if we go full screen on this one. I don't even know if this is going to be in English, but we'll find out.

Happy new year.

Of course. Yeah. So that's one of the videos.

It had some English in there.

I mean, yeah, it did say Happy New Year I think.

I mean immediately there's just a lot of detail and, and of course these are probably the best of the best videos that we're seeing on their website. What else do we have? But it just looks like it's a lot, like a lot more in depth. I mean, Matt, you like analyze the AI video way more than I do personally. But like, what's your first impression on seeing these ones?

So I mean, that video there, am I like overly impressed by it? Not. Not really. I feel like that's kind of level of like vo. The thing that I felt was most impressive by it is I feel like the lip syncing is getting better and better and better. I feel like the biggest giveaway when you're watching like a VO video or a SORA video is when you see their mouths talk and then the voice comes out of it. There's always this like uncanniness where you're like that didn't really feel like it lined up super well. And I feel like Seed Dance is finally starting to solve for that. Yeah, the other thing that I don't think we really saw in that specific video is that you can have it generate like multiple scenes with consistent characters. So it'll show like a person walking forward and then the camera will change to like the front view of them walking towards the camera and then it might change to like a side angle and it keeps a very consistent sort of environment and person as they're, they're walking around. And I think that's also something that's really, really impressive about it.

Yeah.

But let me. Let me actually pull up some on Twitter here that I came across. Oh, this is one of the things that the AI models have always been really bad at is, like, showing gymnastics.

I like balance.

Good for you.

Look at those 10.

Okay, yeah, this is cool. But there's some fun, like, unhinged ones that I think are even more fun to show off.

Let's do it. It's even better.

We just dropped our AI video showdown guide. We took Runway cling Vio and Sora, gave them all the exact same prompts and put them head to head to see which one actually delivers. You get the full, honest breakdown, pricing comparisons, and five universal prompts. You can test yourself right now. Get it right now. Scan the QR code or click the link in the description. Now let's get back to the show. So one thing that I know a lot of people care about is this right here. They want to know what the Will Smith eating spaghetti test looks like. That's the good stuff, and that's a wrap. Mmm. That's the good stuff, and that's a wrap. But, like, did you notice the lip syncing? Like, it looked pretty damn real, didn't it?

I was gonna say he's getting better at this whole eating spaghetti thing. Like, everything's like, yeah, play one more time just to see the lips. And you just gotta see. Yeah,

That's the good stuff, dude. It's. And that's a rap. Like, it's. It's pretty spot on. And his voice sounds like his voice.

Absolutely. Yeah, it does. He's so clean. I don't know how he.

This is one of my more favorite ones that I saw. You should recognize these characters.

Absolutely. You know, that was completely uncalled for.

The voice is, like, spot on to me. That's what is so impressive.

I'm more impressed with that kick right there. I didn't know he had that kind of flexibility or height. You know, that's pretty good.

But no, here's like, a UGC video that somebody made that's, like, pretty mind blowing as well. Check this out.

I live in an apartment with no tap outside. I just unfold this stupid frog washer, fill the 12 liter tank, snap on the four in one nozzle, and spray.

It's cordless, compact, and powerful enough for

a full car wash. That's crazy.

And then she's showing some behind the scenes of how they made it. But, like, that's nuts. This is the future of product ads online. Like, people aren't going to go out and shoot stuff. They're just going to take pictures of the product they're trying to sell. Have I go and actually make a video of them using that product. And we're not going to be able to tell it's AI pretty quickly here. Like they're like this video specifically like, you know, it's called the stupid frog. I don't know if you can see that. So like that's kind of a dead giveaway. It's not like super real. But can we zoom in if you didn't notice that?

I mean I would not go back to. Can you show the prompt like I want to see is that at the end of it? Yeah, here we go. So this video is generated. Okay. So gave a picture of the product it looks like and then the prompt over there. I mean that's nuts. And yeah, this is made by ByteDance. So TikTok and Capcut. You know, it's like very like the lines are blurred, reality is gone.

But actually let me, let me show you something real quick too. That even hammers this home even more. Let me find it real quick. So yeah, this is Vidya. I'm not gonna pronounce her last name, but she's the VP of ads and commerce at Google and she posted this that said VEO is coming to Google Ads. Nanobanana Pro was a success and we're ready to bring that creativity to Google Ads. As an advertiser, it's like having a real time creative partner. Businesses will be able to produce studio quality work in minutes, accelerating creativity at scale. But if you look at the video here, you can see they're just dropping an image of a handbag in saying generate video clips and then they're actually making video ads of that handbag. Like with somebody using that bag and walking with it.

Yeah.

So like it's not theoretical that this is what ads are going to look like. Google is actively rolling this out to be how ads like move forward.

I mean Google would be the company to do this, right?

Yeah, yeah, yeah. So we're, we're, we're actively seeing that happen right now. This is, here's one of my favorite videos though that I came across. This is like a typical day at the Waffle House. Okay, here you go.

If this is your first shift at

Waffle House, you have to fight,

Dude. Yes. Like a Quentin Tarantino film or something like, but better. I don't know.

Yeah, I mean the thing is like, okay, so look at, look at the lady in the beginning and Then it's gonna change scenes. I don't know if that's supposed to be the same lady or not. Hard to say, but there's like a real sort of consistency throughout it where you can tell it's supposed to be like the same people.

Totally, man. And not only are the visuals awesome, it's like the sound. It's a full blown movie. Now that, this thing now, Matt, do you know if it's stitching scenes like

so it generates 15 seconds at a time. This is a 27 second video. So this has to be at least two generations. Yeah, I wouldn't be surprised if they did like you know, four or five generations and then just took like five seconds from each or something.

Yeah.

But yeah, it generates 15 seconds. So this is obviously more than one clip.

Got it.

Here's another one. This one's coming for your old career where you used to make videos for people.

Yeah, it's gone here.

Here you go. This, it's actually does like motion graphic style as well. So check this out. So yeah, like ex motion graphic explainer style videos. It's got that covered too.

Which is interesting because yeah, I, I do work with like did that work a long time ago, but also had a client just recently looking for this kind of stuff and I didn't have the solution literally a week ago. Yeah, that's now changed.

There are other tools that have come out recently that you can do this kind of stuff. Like Remotion is one of them where it's a skill that you can install in cloud code and then you can actually prompt it to go generate a motion graphic video for you.

Yeah.

But nowhere near this level. This, this level here is still like quite a level beyond what you can do with Remotion. Yeah.

And I'm imagining just with everything that when you start to dial in your prompts, I look at this, I'm like, okay, how about like a brand style like I'm thinking for the businesses, entrepreneurs listening, it's like okay. To get consistency that way. It's like, yeah, you could just crank out these amazing looking promo videos, explainers just like as you think, you know. Yeah, yeah, that's what I'm seeing. Everything is just like such in. In time thinking creation now. It's wild.

Here's. Here's what Lord of the Rings could have been. Ready?

Yeah.

Speed running. Lord of the rings in 15 seconds.

Mr. Frodo, why don't we just take the Eagles straight to mount Duel. That actually makes sense.

But again like to me the thing that's the most impressive out of all of this is how well the voices line up with the people.

Frodo, why don't we just take the eagle straight to Mount Duel?

Like that is the best lip syncing I've seen out of any of these models so far, personally.

Yeah, for sure. It's great. I want the model. He said 24th, right? February 24th. Is that what it's called?

Supposedly that's when everybody gets access to it. Yeah. So we'll see. It can do. Trademarked ip, of course, as we can see by Matt Vid Pro's video of SpongeBob torturing Squidward here.

It's working.

He's really feeling it now. And here's another one from the, the anime One Piece. Man. That's one of the things about the Chinese models versus like the US models. I don't want to say that like the. Over in China they don't care about IP and trademark and copyright, but they're definitely a lot more lenient on it. And I think it's a lot harder for like US based companies or companies outside of China to enforce copyright in China.

Yeah.

So you've got these models over in China, like you know, the seat dance models where they're not as concerned about copyright. Where when you look at Google and like their VO models and you look at OpenAI and Sora and you look at all these models that are sort of US based companies, if they let you generate stuff like this, they're going to get sued up the. You know, it's going to be bad.

I mean, just think about like what this does for innovation too. You know, it's like they could just kind of open the, the net wide and start to just test like crazy, throw it out there. And in this case it's a big tease to us wanting to use some of the, these models that are just. It's incredible. But yeah, it'll be, I guess on the IP side. Yeah, that's something that the US and a lot of the other, you know, countries of the world kind of have to tie our hands around.

But yeah, well, I mean also with like the, the training data as well. Right. Like the. When I first started talking about AI, you know, whatever, five years ago, it wasn't like this controversial topic. Nobody was really paying attention to like, oh, OpenAI is taking our data and training on it and Google, I mean, okay, maybe Google people knew people were taking our data but they didn't know they were training AI models on it. Right. We knew they were collecting our data, but we didn't really realize the sort of AI implic of it yet. And so it wasn't like these controversial topics. And now, you know, you fast forward to today and they're under a microscope now. They weren't under a microscope back then, but they're under a microscope now and they're. And people are saying, well, where are you getting your training data? And Now Google and OpenAI and companies like that have to be really, really careful with what they actually put into the training portion of the models. Well, if over in China they're just kind of like we're throwing it all in, well, they're going to have the better models as a result, you know,

and that's the thing. That's why I thought of innovation, just the speed of innovation, the fact that there's maybe lesser, no regulations on that, that training data. It's like, what? Who knows what's going to come.

All I can say about this, there's obviously pros and cons, right? Like being able to quickly make an explainer video for your company that shows off what it does. That's, you know, that's pretty cool. You know, maybe if you do have an E commerce business and you're selling physical products and you want to make a quick video of somebody using your product, okay, that's, that's starting to lean a little more into like a slightly deceptive side of like, okay, we're, we're seeing this in use, but that's not actually it in use. So how good is the quality going to be compared to what we see in the video? You know, you've got that kind of thing, but then you've also got like the whole deepfake side. We're getting to a world where we're not going to know anymore.

Like we're there.

I think, like, I think everybody now is going to need to default to. When I see a video, I have to assume it's fake until proven otherwise. And I think we've lived in a world where we have to assume everything is real unless, you know, there's some sort of sign that is fake. But those signs are going to get smaller further, far between, right? So like now we're going to live in this world where, okay, whenever I see a video I have to ask is this AI or assume it's AI before I assume it's real. Like, I think everybody needs to sort of flip their thinking because it's going to be so damn hard to tell.

We need to have like what X has if you, if you see those videos, at least I'm seeing it now, or it just looks too good to be true. You like, hey, at Grok, is this real? You just see that so often now. And who knows if Grok gets it right every single time either.

But it's going to be a cat and mouse game. It's going to be a cat and mouse game and it's never going to stop because the videos will get better. Tools will come up to detect AI better companies will get better at making undetectable AI, New tools will pop up that will, you know, it's just going to be this circular cycle where each side of the equation is trying to outrun the other side. Yeah, and, and that's just, I think the world we're going to be in now where the companies and people are always going to try to figure out new ways to fool the world.

Do you think China and maybe countries that have lesser restrictions on IP like this for training data and all that will always be a step ahead in things like video?

Hard to say. And the thing is, China doesn't not have copyright laws. In fact, they do have treaties with, you know, I think the EU and the US and stuff like that that prevent copyright stealing over there. So Even companies like ByteDance, they're probably going to eventually like, nerf this model and make it so you can't create, you know, ip. Eventually, where I think they might win is, is they're probably caring less about what goes into the training data, because in China, they're just going to consider any data as fair use for training. They may not consider any output as fair use. Right. Like, you're probably not going to always be able to create Super Mario video or a Disney video or, you know, whatever you want from the models. But when it comes to the training side of how they actually train these models, they're probably going to always be an open floodgate where you can train on whatever you want, where the US is constantly trying to, like, tighten down on what you can use as training data. So in that sense, I do think that China's probably going to stay ahead in, like, a lot of the visual models.

Yeah.

But Google's a wild card because Google owns YouTube, and whenever you put a video on YouTube, you're essentially giving YouTube the right to train on that, you know, so it's like they do have this massive corpus of data as well. There's like all of this video that they can train on. So I wouldn't count them out. I do think that It'll probably always sort of go back and forth between Chinese companies and Google specifically.

Yeah, that's true. So VO pretty soon is going to have this and maybe some new stuff.

Yeah, yeah. I would imagine we'll see like a VO 3.5 within the next, I don't know, two or three months. And by the end of the year, probably a VO4. And I think if a VO4 comes out, it's going to blow people's minds more than seed dance, because I think we're going to, we're going to see what, what the next evolution of these is. We're getting these models so good that we like, you can't really even tell anymore. Like you look at these videos and you're like, that looks like it came from an episode of spongebob or One Piece or that actually looked like a real Seinfeld clip until like, you know, he actually smashes through a wall or whatever. You really, really can't tell anymore. So I think the next frontier is how do we make it so it generates longer and longer and longer videos. I think that's where the frontier moves next. Good point. So like this one' 15 seconds. OpenAI claims that Sora can generate 1 minute long videos, but we've never gotten access to those features. But I think that's what you're going to see. You're going to see it where you're prompting in the, you know, by the end of the year, like 15, 20 minute videos just from a single prompt. Right? Like, I think that's where this is all sort of heading next is how do we extend it while keeping the consistency throughout the video? How do we extend these videos to be 20 minutes long, but the character you're seeing at minute one still looks like this character you're seeing at minute 13. That makes sense, you know, So I think that's, that's the frontier they're trying to figure out next.

And whenever I see this kind of stuff and with AI in general, I'm always thinking, okay, where's the, where's the human element? Like where, where are we still relevant, Matt. And I'm not always grasping at that, but a lot of this stuff we can just let go. I'm happy I'm not doing the video production stuff anymore. Well, at least personally I think of like, okay, humans can probably decide what actually goes live, right? Like have, have a little bit more of a tap and not say that's always going to be the case either.

But yeah, well, we'll talk a bit about Matt. Matt Schumer's video in a little bit or his blog post a little bit in this episode. But, you know, I do think humans still own taste, right? Humans still own, like, I know what looks good, what is interesting to people. I. I still think humans are better at storytelling. I don't know if you saw that, like, sort of viral video from the Joe Rogan show where they had mat. Ben Affleck on. And Ben Affleck and Matt Damon went on this big rant about how they don't think AI is going to replace actors because, you know, they started giving these examples of these scenes where, like, you really feel the emotions behind the actors. You see the actors actually breaking down in tears. And, you know, they're, They're. They're putting their brain in a space that like, an AI can't put their brain in that space. And they have this example of like, there was a movie called the Smashing Machine with the Rock. And there's a scene in the movie where, like, he's in his hospital bed and he's like, crying and he breaks down in tears. And it's a very emotional scene and like, you really, really feel the emotions because I guess, you know, the rock in that scene actually went to a dark place in his head and managed to like, sure, pull that out into real life. And they were saying that AI is still pretty far away from being able to do that. I think it's sooner than they think. But I do agree that that's going to be hard to replicate for a while.

Yeah, it's funny. I just think in my brain I'm like, okay, so method acting, like what you're kind of describing there. It's basically a Claude project where you fill it up with another enough of instructions in silo it to then not like, let's hope it. I don't know. I. I like your point of view better. Who knows where the.

I mean, I don't. I don't have a crystal ball. I don't know where any of this is all going. And a lot of the times my predictions are wrong, but usually when I'm wrong with my predictions, it's because the thing happened sooner than I thought it would, not because it was further out. Like, I. I've made predictions where I've said, like, I think we're going to be generating, you know, coherent un. Like videos that you can't tell aren't real. I think we're going to be doing that within the next two years. And then like six months later, we got VO3 and I was like, all right. I was way off on that timing,

you know, Known me for a long time. Matt, you've always made predictions, even before AI, and they're always pretty. They're within the flavor, you know, like the zip code. Yeah.

Well, I mean, credit to us. You and I were doing a podcast back in 2020, talking about AI and how AI was impacting marketing. And back when we were doing those videos, we were getting like 60 views on them and nobody was paying attention. So the world's changed a little bit, but we were making these predictions back then.

We were all for a rocker back then. No one wanted to talk with us anymore, but look at us. No, I'm just kidding. Yeah. Yeah.

So, yeah, that's. That's seed dance. I think. I think we should probably move on because we have a handful more topics that we're probably going to riff and rant on for, for just as much. But the next thing that really, really sort of blew my mind this week is GPT 5.3 Codex Spark. You can't even say it in one breath. I dare you to try.

GPT 3.5 Codex Spark. Nope. Maybe, I don't know. How many. How many syllables? Is that the name?

No. These companies really suck at naming this stuff. So, first off, GPT 5.3 Codex, you know, without the spark. GPT 5.3 Codex is OpenAI's newest state of the art AI coding model. The Codex means it's a coding model. It's not the model that's in ChatGPT yet. It's really only available to ultra plans that are. No, sorry, pro plans that are paying $200 a month. But it's a really, really good coding model. Like, I've. This is the model I've been using lately to do sort of like vibe coding stuff behind the scenes. And this model is really good. It's only been out for like a. I don't know, by the time this video is out, it's been out for like almost two weeks now. But this GPT 5.3 Codex Spark model, it's the same model, but it's being run on different hardware. So there's this company called Cerebras. I don't know if you've ever come across them or heard of them, but Cerebras is like a company that makes inference chips. Essentially, they're like these giant, giant disk chips. And an inference chip is a chip designed to process the information at the time you prompt it with something. Right. So you look at Nvidia. Nvidia is obviously at the center of this whole AI bubble. Nvidia, you know, they have their, their H1 hundreds and all of their, their big hardware that's going into the data centers. And most of what Nvidia's chips are for have been more for that training side of AI. Yeah, when I mean, they're using them for inference as well. But they're really sort of optimized for training. So a lot of the big models that we've been seeing come out in the AI world have been trained on Nvidia chips. And that's why there's been so much demand for Nvidia is there's, you know, they want to make these models bigger, they want to put more data into them, so they need more, more hardware to train on. Well, there's two companies that, that came out recently. One called, well, not recently, they've been around for a while, but they've been sort of making headlines recently. One is Grok. Not Grok from Elon, that's spelled G R O K. I'm talking about Grok spelled G R O Q, which is a company that actually makes inference chips. And then there's also Cerebras. These two companies make chips that are designed for the other end. Not necessarily the training, but the side where when you enter a prompt, it gives you the response back. And they're designed to do that part really, really, really fast and efficiently. And so what this GPT 5.3 Codex Spark is, is the first model that is being used with the Cerebras chips. So it makes these models insanely, insanely fast. So here, I'll just show you the demo that they put up on their website first, because this will give you like a nice little side by side comparison. So yeah, the left side is the original GPT 5.3 Codex. The right side is this new Spark model. All right, so I'll go ahead and hit play on this. They're both giving the same prompt. Build a simple HTML Snake game. All right, so the one on the right is this new fast model. So check this out. It's working, it's working. 2 seconds, 3 seconds, 4 seconds in 5 seconds and it's done. And they're open and playing the game and it's already, it's already going. That was real time. That was not sped up. That was. It took five seconds to make that Snake game. The one on the left, still going, still going. 20 seconds, 20 seconds in, it's still going. They're already Playing Snake over here on the right. Like, they've been playing for a little while.

70 score already.

Yeah, they've already scored 80. They've already beaten their highest score. This one's still coding. Still coding over here on the left. Almost done. Maybe, I don't know. The video is 47 seconds, so I don't know. Almost done. All right, all right. And now they're finally playing at 43 seconds in.

And the right side looks better anyway. I mean, I know.

And the right side looks better. They're using the same model technically, so, you know, the model isn't really the reason.

But what I found interesting there is, like, when the code part came on, like when it was actually writing the code after understood what it was being asked to do, the code was just like, done. It wasn't like, oh, let's write it out and like go line by line. It was just done in a snapshot.

Yeah, yeah. Here, let me. Let me actually share my screen here and I'll. I'll fire up Codex on my computer.

Yeah.

And think of something you want it to build real quick.

Let's just think of like, you built your website recently, so some type of landing page website, maybe something for this podcast.

Okay, so here's, here's Codex. So if I come down here, let me just make sure it's full screen. So if I come down here, you can see I've got it set on 5.3 Codex Spark. Now, this is only available on the pro plan, so you're only going to see this if you're on their $200 a month plan. But let's go ahead and say, please design a beautiful website for the Next Wave podcast featuring Matt Wolf and Joe Fear. Make it look epic. Okay, let's just give it a simple prompt, see what it does, and just watch how fast this takes. It does take a minute to think. So it's thinking. This is all real time here. We got the front end design workflow, replacing index HTML. All right. Not quite as fast as making a Snake game, but it's done.

It's done.

It took 18 seconds. 18 seconds. So let's see what it made for us in 18 seconds.

I'm judging the epicness of this.

Let's see, here's the website it made.

Dude,

18 seconds.

Wow.

So we've got the Next Wave podcast, Matt Wolf and Joe Fear. We've got. Let's see, does the about page. No, it's not even a. It's all a single page. It just scrolls for us. If you Click on these things, but you can see 150. I don't know if this is. I don't think this is accurate. Seven days a week on air. So it just sort of put filler data in, right? It didn't. Didn't, like, do the research, but look. The destination for creators, operators and curious minds. Hosted by Matt Wolf and Joe Fear, we explore what's coming next and how it will reshape culture, technology, and the way we build. Listen to the podcast, meet the hosts. Here's the captains, Matt Wolf and Joe Fear. Latest episodes. Again, all filler data. Here's where you can listen. Spotify, Apple, and then even put like an optin down at the bottom. And again, it took 18 seconds.

18 seconds and you gave it no context, so we gotta give it a pass there. But I mean, like, that's an epic website. I can't even think of all that stuff in, I don't know, less than five minutes, ten minutes maybe. It's. Yeah, yeah.

So I actually had it build a game for me yesterday, actually. I think I might have just overwrote it, though. So what I'll do is I'll create a new project here. That's what I should have done the first time, but I'll create a new project and we'll call it See what's an app that we can build real quick.

Oh, man. What's an app? What's an app? What's something about. We love baseball. We love the Padres. Is there something around that? Sure.

Well, I'll just call it Padres. So we have, like, a blank slate here.

Cool. Do we want to know when their home games are so we know when to go and the best ones to go to?

Build me a calendar app that I can view in the browser that shows me when all of the San Diego Padres home games are. All right, I'm going to hit submit. We'll let it go.

We'll probably have to use this, by the way, Matt. So save this file and.

Almost done. Almost done. All right. 14 seconds later.

Okay.

Took 14 seconds. Let's go ahead and open this one up. Padres home game calendar. So it's even got all the way back to previous seasons.

They did all the research for us. I wonder how accurate.

Yeah, I don't know. Is this accurate?

That's what I'm going to pull up, actually. Give me a sec on my side. I'll spot check. Where's our schedule? There's a schedule? Okay, so does it pull it. What's the earliest date? 26 March 26th. 26th. Come on. Overlays. Get out of here.

I don't know if it did the research. Let's see anything about doing this. What's that? Cincinnati.

I don't think it's correct.

Okay, so it didn't do the research for us. It just built the interface for us.

Yeah, yeah. We'd have to feed it some data then.

Yeah. So if we gave it some data, then, like. Okay, so instead of 14 seconds, I mean, it's taking like, 25 seconds.

So basically, what then you would layer in? I don't know. Maybe you give it the website. I'm sure you could just give them.

Yeah. Let's see. Padre Schedule 20, 26. All right, I'll just pull it from mlb.com here. Source here, let me copy and paste this into Codex. Here's the actual schedule. Please update the dates. I don't know if it actually will read websites or not. I think it will.

I'm curious about.

So I'm just going to give it the actual schedule. Let's see what it does with it. All right. Searched web. So that is. It's reading it. We can see right now that it is reading it.

Cool. This is probably where it's going to take the most time is searching the web. Because the web.

Yeah, yeah.

I can make the app and pull the exact date that says.

Yep. Yeah. The research part is where it's taking the time when it actually knows what it's coding. It's like instant.

So the coding is the. I'm just thinking of, like. And obviously this is probably not perfect application for everybody, but just imagine it's like, okay, you just created an interface. Now you're populating the actual data, and then just use it on iterate. On top of it.

Yeah. So it took 52 seconds, which is forever.

What do we got?

Who's got that kind of time? And then. All right, so it looks like it's got even spring training in here because it's going all the way back to February now.

Yeah. Okay.

What was the first game you said?

Hold on. Where. Where are we at? Which data are you looking at?

Well, you just had one. You said they're playing.

Yeah. Actually, let's look at the. The first game. Seattle. It should be Seattle 20th. It says here the 20.

When is it?

It says the 20th. Of February.

Oh, of February.

Seattle. But this is the.

Okay, so this. What's the game? This one starts on February 22nd.

Yeah. Yeah. That's accurate. Yep.

So the Dodgers on the 22nd, the brewers on the 23rd, the Angels on the 25th, and the Mariners on the 28th.

Yeah. And everything is an Eastern time zone, it looks like. So that would be.

Yeah, we didn't. I mean, we didn't give it those details, but. No, I mean, obviously, some of those kinds of things, you prompt back and forth. Like, if I said change times to Eastern time, it'll probably take like five seconds. So let's see for sure. I mean, Western time, change the times to Pacific time zone. Okay. Seven seconds later. Let's refresh it here. Yeah. Three, ten Pacific.

There you go. Accurate.

Boom. It took seven seconds to make that update.

We'll have to. We'll have to keep dialing this one in, Matt, to make it work for us. You're like, all right, only give me the games I want to go to that are playing with this. I don't know.

What if I said narrow it down to the games that won't suck?

That's really what want to focus on.

Remove all the games that don't suck.

Well, it's worth actually going to. That's good.

Oh, let's see. I've already limited it to the Padres home game, so I need details. Remove non regular se season games. I don't know. It doesn't know how to know what game, what teams suck. That's okay, but you get the idea. Yeah, it's really, really fast. We didn't have any prompt that we've just tested take longer than a minute. The one that took the longest was going and actually pulling in the real schedule, and that took 52 seconds.

And that's the actual research going out to the web, looking at sites, going back and forth. So. Yeah, that's cool. That's cool.

So that's what made this Spark version so impressive to me, was that this is like, basically the best coding model we've got available right now. The GPT 5.3 Codex. You know, you can argue Opus 4.6 is pretty comparable. They're pretty close to each other as far as capabilities. But now we've got GPT 5.3 that can do coding in seconds instead of, you know, minutes.

Matt, are you finding yourself since. Because I have not personally used codecs like this, but I've been all over Claude and Claude code and all that stuff. Would it replace Claude code for you? Because I know you've been putting out all sorts of stuff out with that,

so I like a little more control. What I've. My impression of Codex is it's for people who just want to let AI, take the wheel. Right? Like, so I actually tend to use cursor because I do have more control. You can see, like, I can see all of the files. I can click into any individual file. I can actually tweak code myself if I'm like, oh, this isn't working right. Let me actually look at the code and see if I can find what's going wrong with it. I also, you know, you can do this in both Codex and Cursor, where you run parallel agents, where I have multiple agents all sort of running at the same time doing different things. But to me, there's, you know, it's busier. There's a lot more going on inside of this. But I also feel like I have a little more control. But I have been finding myself using Codex a little bit more since Spark came out this past week because, like, it's just insane to test.

Yeah, this is cool.

And like, if they ever roll Codex Spark into Cursor, I mean, it's like, game over. Like, I, I, I don't know. Cursor, for me is my comfort zone, I guess, where when I use codecs, I just feel like I have a little less control. Now you can open the terminal. There's a button to open the terminal down here if you want to type your own terminal code. And they do have an area where you can actually look at the code yourself by clicking on this button up here. And maybe eventually it'll load. So, Yeah, I mean, to me, I just feel like I have less control when I'm in Codex. But using Codecs, Spark, you have to use it in Codex. Right now, it's pretty much the only way to use it.

Yeah, I'm sure that'll change. I mean, it seems like they, they start to distribute to place, like, cursor pretty fast. But. Yeah, yeah, but I like the idea. I mean, like, I just keep thinking of, like, from thought to creation. Like, everything's just like that time frame, that time to something. Like, is nothing anymore.

I don't know.

Everything is just so fast. We just showed with video. Now we're coding with any ideas you could think of.

So when you, when you think about it, GPT 5.3. Codex is like the, the best coding model out there right now. And it used to, like, you'd ask it to do something and it might think for a few minutes before it finally did the thing for you. Well, now it's still thinking the exact same amount. Right? Like, the, the thinking that goes into it is the same level of thinking that was going into it when it was doing it for five minutes. It's just doing that same amount of thinking now in seconds instead of minutes. So now when you let it think for, let's say, five minutes, think about how much more it can do in those five minutes at this new speed. Right. That's where like, really the next frontier of this stuff is and why this is really exciting is not that I can do stuff now in 50 seconds or less. It's just think about how much more it's going to be able to do in five minutes than it was doing in five minutes before.

For sure. I'm thinking of, like, to get the most out of Codex or probably any app. I mean, and I'm curious, I'll point it to you. Is like the bottleneck I feel like is more just coming up with the brief and the idea and getting, you know, giving it all of that prior to, you know, doing the. I know cloud code has the whole planning thing, so you can really have it question you back and forth and have this robust plan.

I believe you can tell this to plan as well. I mean, if it doesn't have a plan mode, I'm looking at Codex right now. I know it's not on my screen anymore, but, like, it doesn't specifically have, like a plan mode, but if you prompt it, like, hey, help me plan this. Don't write any code yet. Like, if you put that in the prompt, it's not going to go and write the code. It'll just build out a plan for you. So it's just like a, you know, a prompting thing. I think again, it all comes down to, like, taste. AI doesn't. Doesn't know what to make. Like, you can't just say, go build me an app that's going to make me a million dollars. It's like, it doesn't know what to go make. So you need. You need taste. You like, you're going to probably be better at figuring out user experience, getting in there, testing stuff, going, okay, this button feels like it should be over here. And when I click this, it doesn't do what I expect. There's still, like, user experience stuff that I don't feel like the AI has a grasp for yet. So letting it build a first draft, getting in there, playing around with it and going, okay, this is what I do like about it. This is what I don't like about it. And sort of iterating is where I still feel like the human touch is really, really necessary.

Got it?

Yep.

So I'm seeing a line from the video side. Any of the creative work all the way to here. It's like the human touch have the taste for whatever you're trying to build or the person you're building it for.

Yeah, yeah.

But I'm sure it's going to change pretty soon. Just keep watching.

Yeah, yeah. I mean the Matt Schumer article, he was trying to make arguments. We'll get into it a little bit, but he was trying to make arguments that they're getting better at. Better at like actually the taste portion as well.

Exactly.

So the other thing that came out and to be, to be honest I, I don't have as much to say about this one, but I do think it's really cool. So Google released a new model this week with their the so they updated their Gemini DeepThink model and this Deep Think model is like for academic research and like trying to make scientific discoveries and things like that. So this is the newest model that they just put out. And if we take a look at these benchmarks here, I'll try to explain some of these benchmarks a little bit as we talk through them. So ARC AGI is obviously reasoning and knowledge and this one is you know, a, a pretty difficult benchmark to pass. But as you can see like they compare it to GPT 5.2 which isn't necessarily OpenAI's most state of the art model because they do have GPT 5.3 now as well. But they did compare it to Claudopus 4.6 and Gemini 3 Pro and you can see that when it comes to reasoning and knowledge, this one is pretty far ahead of the rest.

Oh yeah.

Humanity's last exam is considered to be one of the hardest tests to pass and a 48.4% makes this model the state of the art in that one as well. So this is like insane.

Just love the name Humanity's last exam.

Yeah, that's a little dystopic. MMMU Pro Multimodal understanding and reasoning. So like you know, giving it different modalities of like media and stuff. How good does it do? It's actually pretty on par with their previous Gemini model. Beats out Opus and GPT5 and then how does it do on coding and algorithms? You can see it's pretty, pretty ahead of those right when we get down here. So beyond mathematics and competitive coding, Gemini 3 DeepThink now also excels across broad scientific domains such as chemistry and physics. Our updated DeepThink mode demonstrates gold medal level results on the written sections of the International Physics Olympiad and Chemistry Olympiad. It also demonstrates proficiency in advanced theoretical physics, achieving a score of 50.5% on the CMT benchmark, which is condensed matter theory. Oh, okay. And I mean, so they're 50.5% compared to Claude Opus. It was. Claude was 17%. Like, major leap.

That's wild. Okay, I want to test this, but we don't know how.

No, I mean, I know how to test it. I just don't know what to give it. So this one. This is really interesting to me too. So they actually told it to make something that was 3D printable. With the updated DeepThink, you can turn a sketch into a 3D printable reality. DeepThink analyzes the drawing, models the complex shape, and generates a file to create the physical object with 3D printing. So you could give it an image and tell it to turn into a 3D printable file. So one of the things we can do is we can try to give it an image and see if it'll turn it into a 3D print file.

I think that'd be interesting. And then fire up your 3D printer behind you, and then let's see what happens.

Yeah, and then three hours later, we'll see if it worked. 3D printing is not fast. No, I. Well, you know what I can do? Let's start by creating the image with Nano Banana.

Now we're talking.

All right, so let's use Nano Banana here. What should we have it create?

I know you have your go to wolf howling at the moon, but you've probably overplayed that one.

I know. I've already 3D printed it. You might even see it on my shelf somewhere. Like, I've actually 3D printed an AI generated wolf. It's on this shelf, right? It looks good there somewhere.

I still have baseball on the mind for whatever reason, so maybe it's the logo.

This is 3D printed, too. This wolf Halligate the moon.

Oh, geez. Don't you make yourself. Now you're gonna be able to sketch out anything you want, and this is gonna get dangerous.

Yeah, yeah.

In a banana.

Oh, my gosh.

Sometimes I get a.

Create an image of.

I'm holding a squishy brain right now. So maybe. Maybe there's a interesting image of a brain we can do.

Create an image of a brain. Make it scientifically accurate. No background. Okay, so let's do that. Let's have it generate an image of the brain. But it did the fake PNG thing. Just make it a white background.

Happy that we all struggle with the same thing when we just do simple things. Come on.

All right, so I'm going to go ahead and copy this image. I'll paste it back down into our prompt box here. Let me switch this to the thinking model. Let me switch this to Deep Think. Okay, so it wants me to start a new chat. So I'll start a new chat. I'll paste the image in here. We'll use Deep Think. No, I want fast. I want Good. All right, turn this image into a ready to print 3D printable STL file. This could take a minute, though, so we'll see what it does.

Hey, this is interesting.

I'm on it. All right, why don't we get a new chat going here? Let me turn on. We got thinking on. Let's turn on. So this is the one with thinking and deep think. Okay. Find a solution to the three body problem. So we'll get that going. All right, so the other thing that's been all over the news lately. All over the AI news. Sorry, probably not all over the news news, but all over the AI news lately. Is that ChatGPT starting to roll out Ads. I've tried really hard to get it to show me ads. Like, I actually have a second account that's on the free plan just so I can, like, test stuff, and I cannot get it to show me ads. So I don't think it's rolled out to everybody yet. They're just sort of like, testing it in certain areas. Yeah, they did announce some of the companies that are advertising, like, I know. I think, I think DoorDash was one of them, and I think Target was one of them, and I know Albertsons is one of them. So I was like, oh, Adobe is another one. So I was, I went and asked it, like, what's the best PDF reader out there? You know, trying to get it to like, give me an Adobe ad, or like, hey, I want burritos. Where should I go to order those burritos and have them delivered to my door? Like, I was giving it softballs to try to give. Get it to give me ads, and it wouldn't give me ads. Yeah, I think it knew I was testing it.

I don't know. Like, I, I have not. I. I haven't seen anybody even report that they've seen them.

But no, I.

It's interesting. I mean, it.

We.

We knew this was coming. It was interesting. Also, I know you did a whole video on this, so, you know, deep dive on that too, but anthropic, you know, Claude thrown Out their super bowl commercials I thought were freaking hilarious. And. Yeah, and they don't totally paint the picture correctly, of course, but what. But yeah, it's fun. You know, you got people talking about it.

Right? Yeah, yeah. Well, so in the super bowl ads, they basically paint it as like somebody asks a question, the AI sort of responds to the question and then in the middle of their response they start pitching them an ad. Right. Like as if like the ad was part of the response, which in reality is not really what they're doing. In fact, here's like a screenshot of what the ads look like. So you know, they, I don't remember what the question they asked was. I think it's up here. Yeah. Ideas for my potluck or whatever. Right. And so it gives them their answer. And then below that it's got the, you know, the heirloom groceries sponsored. Right. So they make it very clear. It's got like a different background. It's got the text that says sponsor. It doesn't look like a response. The idea is that it's very clearly labeled as an ad and people aren't going to get confused on what was the chat GPT response and what was the ad.

Yeah. And I don't know. It's always, it's interesting and I know there's, you know, of course there's this, you know, there's always different thoughts of like, where does this go? You know, this is the beginning of way more to come. And just looking at like what Google has done over the course of time.

Yeah, yeah. In fact I can actually pull that up because I like I, like you said, I did do a video about it so I have some resources I can share. Let me find it.

Do you think other models will be quick to adapt this type of monetization like Claude Jim and I think the

other models are going to sit around and wait to see how it plays out for OpenAI is what I think is going to happen. Now Google doesn't need to. Google is able to sort of subsidize its chat because it makes all of its money from search ads, YouTube ads, cloud services. Like Google has so many other ways of making money that they can roll out AI chatbots and not put ads in in them anytime soon because all of their other business models sort of subsidize that.

Absolutely.

OpenAI, their business model has essentially become chat GPT. Yeah. Right. So they don't really have a whole lot of other revenue drivers outside of enterprises and coders using their API and people using Chat GPT. Like, that's how they make their money.

It's just wild. Yeah. Like, people not so much in the AI bubble, but it's called the Laymans or the normal folk. Yeah, like, you know, I talk about AI and everybody goes straight to Chat GPT, you know, and, and a lot of people are on the free plan. Some, I think, are now upgrading to the $8 a month plan, which I believe those are the only two plans this is showing on the ads. So I, I just start to think of like, okay, well, people are relying on those as their thinking partners. So, like, like, is this going to sway the responses? I don't, I think it says like, you know, it's not going to change how, how it responds to people based on, you know, if there's monetization shown right below there. But I'm just. Yeah, it's interesting to think about it. It's like, okay, so is the AI going to not give you the full picture until, you know, unless you're paying more for the, the tier of, of that model?

Yeah, I, I don't know where they're going to take it. You know, like, they claim they'll never go down that path. However, they have. There are talks about OpenAI ipoing, right. Like the talks are Q4 of 2026. If I had to guess, it probably won't happen until 2027. But the talks right now are, they, they want to IPO by the end of this year. Now, once a company IPOs, they now have obligations to shareholders to increase revenue.

For sure.

OpenAI is a company that loses money every month, right? Like they lose billions of dollars a quarter right now it costs them more to run their company by a wide margin than what their company makes, which is why they need to do ads. Now. When this company IPOs, and you have stockholders and investors and people that want to see returns on their investment, what happens when ads aren't doing the trick? Right? Like when I look at these ads right here, I'm never clicking on one, right? Like, I know you just gave me the answer I wanted to hear. And then you put an ad below it that is clearly labeled as an ad. Why would I click on that? Like I said, I'm probably not going to. So some people will, right? Like, they'll get some clicks on them, but they're also going to be the most expensive advertising platform out there. They want to charge $60 cpm per thousand views, right? So they're a very, very, very expensive ad platform. And sponsors are probably not going to get nearly the clicks they think they're going to get on these. And so what is, what is OpenAI's options there? They either go, okay, this ad thing doesn't work and try to find a new model, or they go, how do we optimize this to make sure that more people are clicking on the damn ads? Well, they'll follow in the Google footsteps, most likely, if that's the path they go down. And so if I share this here, we can see this is the Google path, right? This is what Google Ads looked like back in 2000-2010. Very clearly labeled as ads. They even had a yellow background says sponsored links. They had the sidebar ads. You knew those were ads, right? This looks very similar to this right here. Like, this is just a sort of modern version of that. It's got a different color background. It says sponsored above it. Very clear Google. Yeah, that's how it started for Google too. Right. So then we get to 2011, 2017, they've got this like yellow ad thing and this green text here, we can tell it's still ads. You scroll a little further, still says ads. 2018, 2024, starting to blur a little bit. It's starting to blur a little bit. You still, it still says ad here, but the ad is in, you know, black text and starting to blend in a little bit more with the ad. Fast forward to where we are today. This is what they look like. They put sponsored results at the top with a line, and then everything below that just looks like every other search. Right. They've made it so you can barely tell, unless you're really paying attention, what an ad is and what isn't. And, you know, there's people behind the scenes at Google testing and optimizing, going, what do we got to do to get more people clicking on these ads so we can drive more ad revenue so we can make our shareholders happy. That's where, Chad, I think inevitably leads to. If they go down this model, like, I don't think they have a choice if nobody's ever clicking on these ads. I think that they have to eventually figure out, how do we optimize to get more people to click? And the solution to that is make it feel more native to the rest of the sort of use case of the platform.

So, like baked into the responses potentially.

So maybe baked into the responses, maybe it's not even baked into the responses. Maybe it's outside. But maybe they remove the gray background. Maybe they remove the line that breaks it up between the actual response and the actual ad, they just slowly, slowly start to shift it to look more and more like it is part of the response. Yeah.

Because, you know, I believe this is what we have a. I think there's a toggle in the back. Right. You can turn off or on if you want your information to be used for essentially matching the right ads to you. Right.

Personalization.

So we have that ability. But still. Yeah, I see what you're saying. I think it's like the Google approach, you know, slowly start stripping away some of the obvious things and see where it goes from there. Because yeah, certainly they're gonna just. It's all about clicks, right? Like click.

Yeah. Well, they switch this in the background. All that's really gonna do is make it so that the ads are less based on previous conversations. Right. But like the ad will still show based on what you just asked it, you know.

True, true, true, true.

Right. It's just not gonna look at like, oh, like two weeks ago Joe was looking at Cialis. Let's put a Cialis ad in front of him today. You know, of course.

Nice.

So like it's, it's going to use the history of what you were chatting with less if you toggle that. That box, right?

Yeah, for sure. I don't know, man. This is like, it's crazy because, you know, this is just, it'll be interesting how it plays out. Like you said, Chat GPT is like kind of in a pinch, it seems like in terms of money, you know, revenue. I mean, do you, do you foresee any other revenue stream? Codex, I guess with the integration, but I don't think that's gonna do it for them. Right?

Yeah. I mean you're gonna have more and more enterprises using the APIs behind the scenes. You're gonna have more and more coders paying the higher tier plans to get access to more tokens to do more coding and stuff like that. So they do have other business models, but they've had those business models. Those exist. They're already making money off those business models and it's not covering their costs.

Yeah.

So it's. I have very mixed appeal, mixed opinions. Right. Like, I do see the appeal of putting ads in there because like if you do have ads, then you can give better and better models to people on the free tier. On the. They also have an $8 tier. A lot of people don't know this. They have an $8 tier called Chat GPT. Go.

What's the, what's the advantage of free to the $8.

The $8 model, I think just has better newer models in it.

Okay.

And not quite as much usage as the 20amonth one. Right. So like the 20amonth does have usage limits. A, you're not getting the most of the art model. You can't even use 5.3 in if you're on the 20amonth plan. You need to be on the pro plan to use the newest new model. The 20amonth plan, you can use pretty much their best models, but not the like top, top. Like this is our experimental. Like we're just kind of feeling this model out. But you do have usage limits. You will find a line. If you're sitting there prompting every day, constantly non stop, you find a line where it's like, all right, you've, you know you're wearing me out. You can't use me this much. Right. The $8 model, those limits get hit a little bit faster.

Right.

But also the $8 model, they're going to put ads in that model too. So free and $8, you're going to have ads in both of those. And so it does democratize the ability for better models. Like ads do allow it to give you like better and better and better models in the lower and lower tier plans and make it accessible to any. And that I think is really, really valuable. Like I think absolutely. I don't think only the people with money should have access to the best options to use AI. Like I don't like, we don't want to like increase the sort of wealth and knowledge gap to only the people that already have money. And so I see the value of ads. I do, I really, really think that there is value to doing it. But I also see the slippery slope that Google fell down that I could also see OpenAI following in those same footsteps.

Facebook as well. They're not innocent, you know, they turn into.

We don't want to go there with Facebook.

But yeah, no, I totally agree with you. It's like we don't want a knowledge gap because like everyone should have. It's democratization of this information. Right. And like having a thinking partner and it just opens up so many more possibilities. And I think, Matt, I don't know if this cues us up nicely for that article. We're going to talk about with like basically the biggest shifts that are happening right now that, that those people don't know about.

Yeah. Well, before we actually get to that, what you got? Let's see if we solved the three body problem.

Yes. Let's check in with our Deep thinking partner.

All right, so I asked it to find a solution to three body problem. To give you a direct answer, we first have to clarify what kind of solution you're looking for. Blah, blah, blah.

Oh boy.

So it gave us a bunch of numbers in math. Let's see. Run your own solution. No, that's what I asked you to do. If you want to actually watch a solution to the three body problem play out right now. Here's a working Python script. It uses numerical integration to calculate and graph the exact figures. And then we've got a bunch of math. Let's see. Create a visualization of this that I can run in my browser.

Nice.

Let's go ahead and do that.

Is there some.

And then let me jump over here to our 3D print thing.

Oh, it's pulling from the web.

Different thing it pull in.

Others didn't do what we asked it to do.

Well, so I did this when nobody was looking. It basically gave me two options. Turn this exact image into a 3D base relief or option two, download a true 363D brain. I said option two, but I thought it would be based on that image. So I think, I think I did it wrong. Let me just say. No, let's do option one.

Yeah, that's good.

All right, while we're waiting on that, check back again. Let's talk about, let's talk about Matt Schumer's article. I did a whole 30 minute rant about this article where I broke it down and. And he's sort of citing a bunch of sources but doesn't link to any of his sources. So I went and hunted down all the sources and actually put the stuff on the screen and showed off what he was talking about. And so I have like a deep rant on this.

Yeah.

So I'm curious on your thoughts and then I can further rant at you

if you want to rant at me. I loved your rant. Other people have done the rants as well. Some similar. But yeah, yours is good. I highly recommend everybody go down. I thought the COVID comparison, he said it was a little, it's a little far stretch there. But the point is like there's, there's so much that general population. Yeah. Just really people who are not tracking everything on X or maybe watching your channel or this show have no clue. It's possible right now, like I think that's the biggest thing. Walking away and like, yeah, everybody's sharing this on, on X. Basically saying, you know, share this to your loved ones. Like, start thinking about, like, what are your plans? Kind of thing.

And yeah, well, I, I feel like anybody that's like, actively tuning into this show or actively watching my YouTube channel, this article is not necessarily for them.

No, right.

They're the people that are paying attention. They're the ones that are tapped in that are going, oh, man, I'm, I'm seeing this all bubble up.

Yeah.

The whole idea of this article was, hey, go and share this with other people who aren't paying attention. And here, oh, well, let's do like a quick tldr. Because, I mean, this is a monster of an article here.

It's a hell of an article, but it's worth going through. But yeah, the tldr, he also wrote it with AI.

He admitted to that on the TBPN podcast I saw.

They interviewed him, like, the day after or whatever.

Yeah, yeah. So basically, he's basically saying, look, I, I, I'm a coder, I have a software company. I'm noticing that I don't have to write code anymore. This is just going and writing code for me. So if it can happen to me, it can happen to you. This isn't theoretical. This is happening right now. And then he goes on and says, all the people that are saying, but I tried AI and it wasn't good. His argument is, well, if you tried AI two years ago and went, it hallucinates. It didn't. Like, I didn't think it was that smart. Well, two years ago is like a decade ago. In the world of AI, the stuff that we were using six months ago isn't even nearly as good. We even demoed it. In this episode, we have AI solving the three body problem, writing us us 3D print files and creating apps for us in 17 seconds. It's not the same thing as it was even six months ago.

Not at all.

Right. It's not the same thing. It was three weeks ago. It's so anybody not paying attention, like, it's moving faster than you realize.

If you've gone on vacation, you're already way behind. So come back.

He goes into, how fast is this actually moving? In 2022, it was bad at math. In 2023, it can pass the bar exam. In 2024, it could write working software. In 2025, the best engineers in the world said they're not even coding anymore, they're just giving it to AI. And in February 2026, the new models arrived that are like the different era. And that's Opus 4.6 and GPT 5.3 Codex, those are the models he's referring to. And then AI is now building the next AI. And this is the part, I think to me that is the most sort of mind boggling for people. And it was something I was talking about like three years ago of like, this is accelerating. And here's why this is accelerating is because they're using AI now to write the new AI models. So GPT 5.3 in the article, in the launch article where they put this out here, I'll even pull up the launch article.

It stuck out to me too.

So in this article about GPT 5.3. So this is the GPT 5.3 Codex launch article. GPT 5.3 Codex is our first model that was instrumental in creating itself. The Codex team used early versions to debug its own training, manage its own deployment, and diagnose test results and evaluations. Our team was blown away by how much Codex was able to accelerate its own development. Right. And the Anthropic dude, Dario, he said the same thing about Claude. He's basically said, Claude, a lot of people at Anthropic are now using AI to improve their models. So we're now in this loop where we use AI to help us write new large language models that are smarter and better than the previous large language models. And then these new models then go and help write the code for the next large language model, which is even smarter. And now this smarter model is writing code for the next smarter model. So like this flywheel that was spinning kind of slowly for the last like, like three years is now like picking up speed. And now these AI models are writing the code that's writing the next AI model, which is a better a AI model that's writing the code for the next AI model, which is now another better AI model which is writing. And it's just like this flywheel is going now and these models are improving at a rate of like just insane exponential speed now.

Yeah, it, I mean, it just said it right there. I don't know. Is this AGI, Matt? I don't know. Are we there? No, I never know there. It just says it's trending or something on your. Saw that on X actually on the right side, AGI is here.

But I mean, Padres is also trending. It's the, the trending is always based on what you tend to look at. So like your trending is going to look different than mine. So mine's always going to be crap like Padres and AGI because those are the two things I engage with.

Yeah. But yeah, yeah, yeah, it's interesting. Like going back to, I don't know, back to kind of like the thread I've been bringing about. Like the human, human part of this all is almost like the human setting, the constraints maybe on these. What these new models can or cannot do, it seems like probably is where the role of the coders or engineers are working more with these days. But having these models just literally build themselves is the future.

Yeah, I mean, it's still kind of the taste thing, right?

Sure.

It's still. The engineers still need to tell it where to go next. The engineers still need to tell it, here's what this model was good at, here's what it's not good at. Like, I don't think these models are self reflective enough to go like, here's what we screwed up in the last model, so let's go make it better in this model. I think it's still the engineers going in and saying like, hey, this is what you screwed up in the last model, here's what you need to fix in the next one. So there's still like that element. But then the point that he's also making in this article is these models are still, are starting to feel more and more like they have a better sense of taste. Right. They have a better sense of, you know, what needs to be done next without the user prompting it of what needs to be done next. Right. They're getting a better idea of like vibes and what the visual should look like. And they double check themselves on mistakes and then go back and fix their own mistakes once they've realized they've made those mistakes. And so we're starting to see like sparks of that, the thing that we might call taste. The things, the, the vibrant intuitions that humans have that the LLMs don't necessarily have. We're starting to see sparks of those intuition pop up.

Yeah. Line to blurring even more. Well, this is interesting. All right, continue here on the tldr. What do we have here?

So he goes on to say, what does this mean for your job? And you know, basically saying like, look, if it can figure out coding and it can figure out all of this stuff, it's coming for your finance job, it's coming for your, your medical analysis job is coming for your customer service job. Like, basically these models learn to get really good at code first because code was the, the, the tool that opens up everything else that's after it. Right. Like the next frontier of insane large language models that can do anything you want isn't just shoving more training data into these models and trying to say, hey, get smarter. The next sort of frontier of these large language models is getting them to be better and better at using tools on your behalf. And if it doesn't have the tools to use, creating its own tools. So you tell it, hey, go solve this problem. And it goes and tries to solve it, and it says, hey, I don't know how to solve this, but I can make a tool that knows how to solve this or a tool that helps me solve this. And then it goes and builds that tool and then uses that tool to help get to the problem. Like, AGI doesn't come from large language models alone. AGI comes from large language models that can use tools, but also create the tools that it then goes and uses.

Right, right. It's resourceful to figure out what the gaps are, connect all the dots and give you the thing, whatever that is. Yeah.

It figures out its own bottlenecks and then creates tools to solve those bottlenecks. And so anybody that's in any industry going, this isn't going to impact me. Well, it probably is. I think the safest industries right now are skills. Like, I like, like real world, like outside, away from a computer skill.

Be a plumber kind of thing. Or electrician.

Plumbers, electricians, barbers, you know. Yeah, those kind of people. Like, do you. Do you want a robot with scissors to your head anytime soon? I don't.

Yeah, if you want that shave on your neck with a robot, I'm good.

Yeah. You know, athletic stuff. Like, I. I mean, a part of me does want to watch robots play baseball against each other, but I also think that would only be novel for a little while and then we'd get bored of it because they all have the same level of skill. Right.

Be perfect.

Yeah, yeah, yeah. So it's like, the idea of it sounds cool, but, like, people would get over it. We want to see humans do a lot of those feats. I still think it's going to be a ways off before people really want to go and watch a movie they know was generated by AI. I still think we'd like to see actors doing their thing really, really well. I do think that gap will close over time, but I think it's still a little further off than most people realize where we're gonna go. Oh, man, I really want to go see that new Star wars movie that George Lucas typed with one prompt. Like, no, I. I want to know that a human wrote it. Like, the Fact that I know that somebody put that much effort into making this thing is part of what makes it special in my mind.

The credits will just say prompted by.

It's horrible. Yeah, yeah, yeah, yeah.

Well, and that's the point of this article, right? Is it's. It's to basically shake people and say, here's what's happening. If you didn't know, now you know. Is there anything more in here? I think he kind of covered the.

No, I mean, that. That's pretty much the tldr. We kind of went over it. There are some analogies that I think are not great. Right. Like he says, this is like Covid where we were seeing it overseas for like a couple months, and then. And we're like, oh, that's over there. That's not going to affect us. And then the next thing we know, it's here and we're shutting down and the world changed forever. And he sort of compared that to this.

Yeah.

And also said that, like, CO was on an exponential growth curve, but no, it wasn't. You can't have an exponential growth curve when there's only 8 billion people. Like, at some point the curve flattens out because unlimited people. You can't give CO to more people than exist.

Yeah, yeah.

So, like, it's not. It's not a very accurate comparison in. In that way. Also, AI is not sneaking up on anybody. Like, if you turn on your tv, you can't not hear about it. Right. So it's like, it's not. It's not creeping up on us. And people are going, oh, that's over there. It's never going to impact me. People know it's coming. Yeah, right. I do think there's a lot of people that are thinking, oh, it's not going to impact me. I do think there's a lot of people that are putting their heads in the sand and going, oh, I hate AI. It has no usefulness. I'm never going to use it. It's stealing and. And screw it. I'm not going to use it out of principle. I think there's a lot of people doing that.

Yes.

And I think those types of people are going to be in for a rough ride. But it's not sneaking up on anybody. It's not spreading like a COVID virus where one day you don't have it and then today you do. But I get what you're trying to say, like, this is coming. It's coming faster than the majority of the world realizes. And what you need to be Doing is paying attention and not putting your head in the sand.

Right?

Like that's the tldr of it all.

There you go. Just become aware and watch my video on this. If you really aren't sure, become aware

like the AIs are doing right now. They're becoming aware. Okay. Sorry. Creepy.

I think. Do you want to do it? Should we check in on our experiments and.

Yeah. Which one do you want to know first?

The brain. I feel like that's the easier one.

Okay, let's see. So what do we got here? Go to image to s. It's telling me what do to do.

Oh, we don't want to do. We just want to be given the thing and go print.

Okay. Okay. I'm. You know what? I'm done with this.

We're trying too hard.

We're trying too hard. Demo basically showed them giving an image and saying, turn this into a 3D printable file. I did the same thing and it's giving me step by step instructions on how to turn it into a 3D printable file. It's telling me to go use a background remover and then upload it to image to stl.

Oh, wow. Come on, Gemini.

Get your step. Like, I knew I could have done that. I've done that kind of thing before. I wanted you just do the whole thing for me. All right. So. And then our three body problem. It did write some sort of visualizer. Let's see.

Should we throw that in Codex and I don't know.

Let's see. Copy the code below. Go online sandbox to like code pen and paste it directly. Oh, okay. Let's do that. That I'll open up CodePen. Ooh, there's our visualizer. Look. It solved the three body problem.

We did it.

We solved it. It's officially solved. You're welcome, science.

We're done here. This is great.

Where do I apply for the Nobel Prize?

I know, right form somewhere.

Where do we go?

Can we all get credit?

It is a pretty cool visualization though.

I like it. Do the colors mean something? Let's see. Like, does it?

I. I don't think so. I think it's just trying to show you that like, if you've got three planets.

Yeah, yeah.

If. If they move in this pattern, then they never perfect have too negative of an impact on each other, I guess.

Yeah. It's like perfect balance or. Yeah.

So. So there you go, man.

All right.

I'll write up a white paper. I'll have AI write up a white paper for me and Then I'll send it to archive.org and then. Then I'll send it in for the Nobel Prize. And we did it.

Just put this episode as a credit to show our work.

We did it. We did it. Go team. Here's a complete standalone interactive visualization of the three body problem. The stable figure 8 orbit that you can run directly in your browser. Okay, this is just telling me how to run it.

Yeah, it's not explaining it. I hear.

Okay, what's happening under the hood? The math engine. To prevent the orbit from instantly degrading into chaotic mess, the script moves the three bodies forward using an incredibly tiny slice of time. The simulation loop. To keep it from looking slow to the human eye, the script secretly runs the complex calculus function 15 times per frame before finally updating the visible image on your screen. Explain this. Like 5. Imagine you and your best friend are holding hands and spinning around in a circle. Okay.

It's pretty easy, right?

Got it. You just go round and round and you always know exactly where your friend is. In space, this is what happens when two things like Earth and the moon pull each other with gravity. The three body problem. Now imagine a third friend runs up, grabs your hands and tries to join the spin. Suddenly, everything gets crazy. You're pulling them, they're pulling you, and the third friend is pulling both of you at the same time. Gets yanked too hard, someone slips, and all the all tumble into different wild directions. Because everyone is constantly pulling everyone else it complete. It's completely impossible to guess exactly. Okay, so the figure eight. The magic trick. But what if you and your two friends practiced a very special dance for a really long time? You'll discover a secret trick. If you all weigh exactly the same, start the exact right spot and run at the exact perfect speed. You don't crash. Instead, you beautifully chase each other on an invisible track shaped like a giant number eight. This is what the computer animation I gave you is showing. It is a rare magic trick in space where three planets perfectly balance the gravity so nobody trips how the computer draws it. Because there's no magic math rule. To jump straight to the end of the dance, the computer has to draw the animation using tiny baby steps, etc. So I don't really think it solved any problem. I think it it found like a known solution of like, if everything was moving the same speed and everything had the exact same force of gravity and ever and all of things aligned perfectly, then you can get this figure eight dance thing. But in reality, isn't the three body problem much more Complicated because there are different speeds and different forces of gravity and different, like, factors. I don't really feel like it solved anything. It's just saying, like, there's this one sort of set of. There's this one scenario where if all of these being true happen, you can get this figure eight to happen.

I think we still apply for the award, but.

Okay, yeah, I'll submit it.

You just gotta submit and see what they say. All right.

I got my buddy Demis. He won a. He won a Nobel Prize. So I'll just send it to him and be like, hey, can you forward this for me?

Give me it, man. Help me out.

Hook it up. Hook it up with a Nobel Prize. Yeah. So I think I. I think we covered a lot of ground on that one. This was a different style episode. We kind of did this style of episode last week with me and Maria where we picked five interesting things and just sort of of ranted on them. Let us know what you think in the comments. Do you like finding out what happened last week and hearing our little rants and rabbit holes about what we really think and what it really means for everybody? Because they're sure as heck fun to

make Anyway, there's always news. And, Matt, you're on the. You're on the leading edge of it. It's fun to debate and go deeper on here.

So I hope to do it again next week. So if you like this, an episode like this, let us know in the comments what you think about this kind of thing. And if you haven't already, make sure you like this episode and subscribe to this channel. Wherever you listen to podcast, YouTube is probably the best place because then you get all the visuals. But subscribe to this podcast wherever you listen to podcast. And thanks so much for tuning in. This was a fun one. Thank you.

One.