#14 GPT-5, Genie 3, GPT-oss, Funeral for Claude

Welcome to Command AI, where we press command F on the internet AI news and somehow still can't find our sanity. I'm Samuel Gregory and I'm Kabza and we are here to decode the weekly chaos in AI, web design, and development. This week we have three big news. My favorite, Google's Genie 3 creates interactive worlds in real time. Think like gaming in real time. Open AI drops gamechanging GPT5 plus a free opensource model that runs on your laptop or even on your phone which is crazy. Uh and that might be a funeral for cloud and all of this we'll get into it uh in this week's episode. Amazing. I'm saying that website might be a better way to show this.

Let's in further you're cutting off a little bit. Sorry. Um your your internet connection is not really stable. It seems like it's all right. I was just loading a new website and that kicked that. Ah, here we go. Let's let's get let's get into this. Genie 3, a new frontier for world models. And it's exactly as you say. It looks insane. You can generate 3D worlds via prompt and and it's not just generating the 3D world and that's it. You can actually move through it using this kind of you know uh directional pad here. So this person is in a sci-fi world. He's walking through and you can look around. Um I have got the video cached here.

So I'm not going to play videos. Yeah, this is like insane and it's it's like the direction that we are going with this like they've obviously prompt take part in a show horse show event skiing the Alps you can control yourself helicopter pilot absolutely insane um real time interactivity as well and this is all procedurated and they they show you so this guy's painting This is the good one. And then you can prompt in real time. Look. So, um, give it a second here. Man and chicken soup. So, in real time, you can actually prompt. That's good. Oh, I want a dragon. Absolutely insane. Uh, ah man, this is so so cool. Let me let me let me run down the facts.

Right. So, every every frame is generated based on previous frames. It's re basically real time using input and using arrow keys. Um, you can obviously change the scenes on the fly with text prompts. We saw a few elements there. Useful for simulating environments uh for like AI robot training. I mean, you know, think of think of what this could actually do. So, I definitely think simulation training which is mental in VR. So no doubt this will be plugged into some sort of VR device or made into 3D and stuff like that because that can be really good for um obviously like military training and things like that but like actual treatment as well. So yeah um of like psychosis and all the rest of it.

So, but on the more lightart lighterarted side of things, creating real-time video games, movies and TV shows, like you you can become I mean it's limitless really the sort of sorts of things you can generate and um it's only it's bit sad it's only being used internally at Google. No public release announced. So, bit of a shame there that we can't play with it, but like it just goes to show where where we're going with this sort of stuff. And um this is like genuinely mindblowing. Yeah. Yeah. Real time as well. Like the computer like this is the thing uh Theo's mentioned it quite a lot. This is the unique position that uh Google have got themselves in is that they can actually um they can do so much because they own and operate the hardware as well as like they've got the full full stack, let's call it.

So they can do really heavily computational things like this. Um but it's interesting that they've just got no planned release date. Like what's the what do you think the strategy is there? Do you think it's did test people actually care about this? Like just imagine all the things that you can build with this that that is insane to me. For example, Google maps. Have you thought about it? We have like 3D worlds already with Google maps and Apple maps, but they are not very exact. They are not really detailed. They are not really nice. But we also have a lot of uh like images, what are they called like street view images from Google. Now imagine combining these things and with this generative uh AI to create a replication of our world with every single detail in it and now take that and use it for so many things.

For example, just the Google maps example is an obvious one, but also gaming. Have you thought about gaming in your neighborhood like Call of Duty style? I mean, I used to when I was like 16, I used to play some custom DLC. So, we would uh edit maps, add things to them, and I always dreamt of having my neighborhood maybe, you know, with a little bit of like extra fun stuff added to it, and then play with my friends that were close to me all we we were from the same neighborhood. Play Call of Duty there. Theoretically, you can always create your own game. But now imagine how easy this gets with this model. Uh so this is one of the very like for me very interesting use cases video games like Call of Duty in your own uh neighborhood with your own friends.

Another thing as you mentioned for robotics um they already have like some generative words for worlds for them but with this they robots could have almost a perfect replica of our world that would be also very interesting to see uh how they can use it to to learn to learn and I don't know how this will be for like self-driving cars probably could they could use it as well so it's not robots. Uh, self-driving cars also are type of robots. But yeah, I I think this is like super exciting. I'm just not sure why Google is not releasing it. Maybe because it's really costly. Like a video typically takes minutes to generate and now imagine this is kind of like a video in real time.

they are probably running this on huge clusters of like probably like even potentially thousands of GPUs at the same time because it's not like a video for which you can wait to be generated. This is real time. If you move, it has to move and that probably needs a lot of um GPU power. But with time, as we've seen in the past few years, we went from Will Smith spaghetti to now realtime worlds, it's just insane to me. Um, looking at like the history of how we got to Genie 3 as well. Um, game and gen. No idea what that is, but it's uh 320p um game specific. I don't really know what a lot of these these metrics mean, but like you know, basically they've gone from a few seconds and this is Genie2, the last one, which is 360p.

We're up to 720p. I think that's an important distinction to make there. We're still only 720p, but it's, you know, still pretty good. 3D environments to general. Again, I don't know what general is. Um limited keyboard mouse actions. And this is promptable promptable world again don't really understand that but you know you can navigate around it. We've gone from 10 to 20 seconds which is pretty good to multiple minutes. So that's fantastic. Unlimited, right? If you get to multiple minutes, you could I I don't see a reason for it being limited because it can just look at the last two seconds and based on that continue generating or maybe maybe like it's something with the context.

For example, in the website, they show a few trees that stay consistent in in the video or like what is it called like in the experience, they stay consistent. Maybe there is something about the memory of these elements staying consistent only through a few minutes. But if you don't care about the consistency of 2 minutes ago and if you like progressively just move forward but not backward, you could run it unlimited. Yeah. So if you look here, I mean once it's generated all this, it's not generating anything new, you know, it's it's in the it's in the memory presumably. I don't know how this technology works, but you know, again, this character is running forward.

it's already generated this stuff. But I guess there's perspectives and it changes perspectives. But yeah, without knowing the technology, I'm not too sure. But but you see here, I mean, here's here's an example. Keeps on. Here we go. So here's the difference between Genie 2 and then G3. So there's a lot more fidelity uh inside of Genie 3. It's this is everything's on max here from by the looks of things. All the there's no there's no nuance. There's no kind of subtlety in the in the image. So um and then obviously you can see here the interaction has been has ended. But yeah I mean theoretically the the user isn't moving constantly. So what's been generated there's no reason to you know stop generating anything more.

I could keep going sort of thing. But yeah um pretty pretty cool. Um and and the other thing is is is physics as well. Real world physics. So, there is an there is a cool one that I've one here where it's like it's going to it's going to bash through one of these lights. Now, look. Boom. Like, it's that real world physics that is really really um impressive as well. Uh so, yeah, this is just just man imagine video games in So, here's the question. Which one will come out first? GTA 6 with their amazing. We We've seen the visuals, right? Have you seen like they have the beer and it looks like the beer the fluid inside of the glass is moving in GTA 6 trailer and it's like amazing visuals.

But will we have that first or you know Google's what whatever game engine in real time that look at this like this to me does not look 3D generated. This actually looks photo realistic to me. I mean the guy is barely moving so this looks a bit naff but like this all just looks like video footage like it's actually a movie. So what is this? Is this this generated? So another thing that I'm re No, it's not 3D. It's they they literally just make video on the go. That's how I understand it. This is like video on demand and the interactivity is from what I understand it just generates more frames that are technically just a video.

It's not real 3D environments. I guess then the ones that look like video games are probably stylized to say like the robot one we were looking at like maybe that's stylized to look that's how I understand it. That's how I understand it that it's not a 3D map. It's only like just frames. What I'm what interests me a lot about like this turning it into video game or like other type of experiences is to not let the AI to create the entire experience but have some sort of like a wireframe of like a 3D wireframe of for example the character in a 3D world that a game developer can create but they don't have to develop all the you know nittygritty all the details like the the bottle with the liquid inside of it that they don't have to create that.

They all they have to create is for example the character just directing it and moving it through the story and these models then kind of like create all those like detailed layers on top of it if you like say the render if you wish and that's how I see it moving forward. I think that would be like huge for the gaming industry. We will have like photo realistic like true photo realistic. Yeah. It kind of flips it on its head a little bit that we can just use like photo realism in movies instead of having something that looks like a 3D video game. You know, it just I wonder I wonder I'd like to see some uh video game studio reactions to this and see what Have you seen the Have you played Call of Duty mo Modern Warfare?

Warfare one and two and three also was but like you know Captain Price. Oh no, not to that extent. No, I don't remember. I was called Captain Price. Like my gaming tag was for a long time Captain Price because I was just obsessed with it was just cool. A cool guy, a cool dude in the the in the um in the series. It's a guy with a like long mustache. I don't know what it's called. Like you when you have the mustache like all the way down here. Handlebar. Yeah, kind of like a handlebar mustache. So, Captain Price. So, uh I played Call of Duty. The point being I play Call of Duty again, the remastered version.

And what they did, Activision, they basically enhanced the visuals from, let's say, kind of like very like old style graphics to a better one. Now, I'm looking forward to play it again potentially with like actual like photo realism. That would be insane. There are like so many iconic scenes. people who played it, they I'm I'm sure they connect with me on this. Uh like imagine in the scene where you throw the the knife and you hit I I don't want to spoil it for the ones who haven't played it, but it's so old. You throw the knife in one and you shoot somebody with a knife essentially like you hit them in the eye. Um there there there are like there there is another scene that is breaking every man's heart is where they burn you.

So your someone in your team that you think uh they are on your side. They they betray you and they kill you and your friends and they burn you. And you see that from that perspective and just having all of like all of that Call of Duty in because it was like I I believe it was the peak of Call of Duty those um Duth 3 and yeah have it remastered in photo realism. It just excites me. I think you need to see someone mate because if you want to see that being shot in the eye with a dagger in photo realism. Yeah. I think you need to speak to someone. Yeah. Yeah. Yeah. I I need to go out and touch some grass, mate.

You need to do more than that. Touching grass ain't going to remove those dark thoughts. I know mentioned the dark ones, but uh it was so good like the storytelling was so good. This is a note on what I said previously that in a way I don't know how storytelling with from AI will be but so far from what we've seen also from uh VO3 it's called right the video generation uh the the that tool is just creating some frames but the interesting storytelling is still coming from humans right it's that well that's it I mean like it's all well and good got having these desires to like you Oh, the video game industry is going to die and this that and the other.

I mean, we've seen it in Vibe coding. Just because you can now technically and physically build a website doesn't mean it's going to be a good website that actually does anything because you don't hold all of the other elements that go into building a good website. You don't Oh, more direct example would be the Yeti vlogs that I talk about where just sometimes a little bit love on YouTube. Hey, leave us a comment. Where are where are you? um you know they it doesn't make up for bad storytelling when the joke is just rubb rubbish. Just because it looks cool and oh there's a yeti holding a a camera um 100% doesn't mean it's going to be very good.

So uh yeah so you can you can try try as you might. It doesn't mean it's going to be a good video game you know. Exactly. It's just going to like AI is going to enable more of us and it hasn't gotten to a point that can replace any of us. uh even despite these companies talking about AGI and coming up with new models, speaking of which we we should jump into it, none of these tools are where they these companies um claim to be close to AGI to take the creativity. So far we are the creatives and these are just tools that allow us to move faster and um yeah with less manpower do more.

Should we jump into it? Should we jump into OpenAI? Yeah, they came with two big news. And so first off the off the press is GPTOSS which is uh two state-of-the-art open weight models, right? Right. And you got 120 billion uh parameter model and a 20 billion parameter model. So some really exciting stuff here. um apparently deployable on consumer hardware and these are these are fitted with the Apache 2 license meaning that you can uh reuse them uh pre-train uh post-train them um do what you like with them put them anywhere and you know open AAI can't take a scent from you which is really really cool so as I say operatable on consumer hardware we'll get into a little theory that I have coming up in a bit.

Um but they are basically the the performance is on par with with 04 um 04 mini um and uh this is the 120 billion parameter one and the 20 billion parameter is kind of like 03 mini. So, we're still not we're not we're not talking like amazing amazing models here, but there's something you can download and run on most um hardware, particularly the 20 billion parameter model. I don't know the exact sort of size of that, but apparently you can run it on 16 gig of RAM. So, that's pretty interesting. Um, my theory is that this may be the under it could be what Apple intelligence becomes is my theory is my theory. Um, I'll get you get your response to that in a sec, but good for tool use, train of thought, reasoning, and and Healthbench as well.

Again thinking about the fact that these can run locally and it performs really well on healthbench meaning hospitals can um run this on on their you know I don't know how deeply ingrained AI is in hospitals but you imagine a power outage and you know you're you're very reliant on AI systems or having a you know um another perspective with an AI on someone's condition or whatever. The fact is this is a really nice place for it to health getting good results on health bench is a really nice place for it to to to be good at, right? Um it's got adjustable chain of thought which is really nice. So um you can actually say give us give me some quick results.

So don't don't your train of thought needs to be uh fairly small. Um or you can make it massive if you want some some deep thinking and all the rest of it. I guess guess more for the health bench sort of stuff which is really cool. Um, it leverages mixture of experts. Where have we? Let me let me start highlighting some stuff so we can um There we go. Mixture of experts, which as we learned last week is basically a way for the the um oversimplification on it's not quite routine, but you have designated experts that basically um encapsulate a certain portion of the data set. So 100 billion parameters activates 5.1 billion parameters per token.

So there's an expert dedicated to 5.1 billion parameters and it can it's like a specialist in that subject basically. So improve improves performance uh because again you're only activating 5.1 billion parameters on that 120 billion parameter model whereas the 20 billion parameter only activates 3.6 billion parameters. So you're getting that um increased performance, you're getting that um yeah, the responsivity and things like that. It's got uh 128 uh,000 contexts, which I was doing some research today and you know when when GBT first came out, GBT2 I think it was. I think it had only 6,000 as a context size, which is just crazy. Here we go. There's the brilliant parameters. So just I mean like let's be real Google talk about a million parameters but like I think context size is way overrated.

I I'm just I'm thankful for the 128 basically which is on a free open source model so or open weight model. Um and what's really really interesting is when we get to the post-training stuff, uh is it post-training? Uh basically they have um including a superi supervised fine tuning stage where basically they tried because obviously you can do what you want with this model. they've actually tuned it in such a way or or implemented such um features into the into the because remember it's not open source it's open way into the model where uh open AI have tried to fine-tune it to do malicious things and is actually really good at not doing those things detecting whe whether it's I don't think this is specifically what I'm talking about but you know that that's really cool they actually intentionally ally tried to make it do harmful things and it it wouldn't do it.

Um because that's the thing with this all this open stuff is that once you release it out there it's done then you you've released it. It's it's free for the world to do whatever they want with it. So really really cool and before we get into some metrics or whatever there's train of thought. Here we go. This is the fine tuning that I I told you about. So set state-of-the-art approaches for safety training during pre-training. We filtered out certain harmful data related chemical, biological, radiological and nuclear um data and during post training we use deliberate deliberative alignment and instruct hierarchy to teach the model to refuse unsafe prompts and defend against prompt injection. So um again just that uh safety thing.

Um, and there is a bounty. Um, just lastly, there is a $500,000 bounty to if uh Red Hat challenge basically. So, if you can crack this nut um and reveal vulnerabilities OpenAI are offering on it was a bit higher there. A bit higher like on the top. I've seen that. Yeah. Yeah. The first Yeah. Um the last paragraph. Got it. Boom. Boom. Boom. So, red teaming challenge. So, there you go. Get on it, boys. see if you can uh possibilities. Yeah. So, you get half a million dollars if you can break into this and then Mark Zuckerberg will probably hire you. Yeah. Yeah. For like a couple million dollars at least. Oh, you got 18 days.

You got 18 days. Go on, get on it. Um so, yeah, really, really exciting. really uh obviously open source and and open weight isn't a brand new thing but let's be clear like um open AI are good you know and they have they have a lot of the best interest of AI we'd like we'd like to think and what they say um they have humanity's best interest at heart whereas you don't know these open source and and whatever models you you we often don't know what they're you don't you don't know whether they're aligned in the same sort of ways. Again, so we're all taking it at face value, but um I've not had a chance to use it.

I I think that it was like an immediate update. I' I've seen mixed responses basically on how good and how fast this model is. I think I used it for a bit and it would it would take it would warm up. So, my first prompt was like whatever and I'll plug one of my videos uh here. I've just released a video where I get and in fact I think it was you that requested this video. how to get open router uh sorry um open web UI set up on your own domain name. Um so I've released that video or it's coming out or something like that. But the point is I got it up and running.

Um and it was a bit slow. Oh, no. It was it was on my local machine. It wasn't on my own domain and um yeah it was slow at first but then it kind of came up with some stuff and it I saw uh Theo's video on it like oh no that's GPG5 but loving tables but like I think that was G GPD5 so ignore that but like uh yeah I've had little I've had a few little bits and bobs um of usage with it but nothing too crazy. I don't know if you've dug into it at all or No. Do you test models? Uh the the next one that we are going to talk about in a minute.

Uh I tested that a little bit but honestly no. I'm I mean I use cursor. Uh I I stopped using it for a bit because I was busy with other stuff. Now I'm using it again and I test the different models that are available in cursor. But the new ones that come out, I don't have my own test even though I have something in mind. You've seen the the the slider thing that we have like the carousel. It's quite complex and I yet to see any model. I tested a few models uh to see if they can create that and it's too difficult. So my test maybe I can test models in a year time when the models are good enough to create something that complex.

But uh I want to mention a few other things. So one the Apache 2 um license license. So what we are saying what we are seeing is that OpenAI's new model is more open than what Mark Zuckerberg has been putting out there. Uh I don't know. I don't know what the license is on Llama. I thought Llama is like you can just do what you want with it. It's just it's just there. It's open source. I've heard I don't know I we need to check that probably but I've heard that with Llama you are you are you can use the models up to a set limit of users I believe like after I might be wrong about this but after like a million users or something like this you have to pay them or show that it's their model or something like this we can we can check that but Apache 2.0 Oh, that that sounds great.

I've seen a few I've seen a few memes or like a few jokes about OpenAI's new open model in relation to your AI girlfriend, you know. Yeah. You know, people now have apparently AI girlfriends. And if your AI girlfriend is not running locally on your machine, she's for the servers. It's bad. I know it's bad. You know, they say like she's for the street. Okay. Yeah. And she's for the servers. You know, she's not running on your machine locally. She's she's deployed to every she's deployed to everybody. You know the movie her, right? Yeah. There there is the scene at guys. Close your eyes if you haven't seen it. Close your eyes. Close your ears if you haven't seen it.

What am I saying? Close your ears if you haven't seen it. But there is that scene at the end where the the AI girlfriend is talking to I don't know like 6,000 men at the same time. Um, and the guy is heartbroken because of the exclusivity thing. And now your AI girlfriend can run on your machine. So she's not for the servers. So, she's not deployed to everybody. I don't know. I've seen jokes about it. Anyway, so if your app is over 700,000 uh million monthly active users, you must request and receive a special license from Meta. So, that's what you're talking about. Yes. So, Open AI's new model is more open-source than uh Meta's open-source AI models.

Yeah, if it makes sense. I think I think Meta are in a weird spot right now because they're sort of Yeah, like what is their USP? I don't I don't really understand like I think I think I think Wow. I think I know what their USP is. I think they it's why they've got a bad reputation. It's it's their hu it's their data they already have. It's the data the real personal data that they you they have on their users that is not public to the internet which is all open AI have. You know well they're probably starting to collect data usage from their users but Facebook has that and people freaking hate it. I really I mean I've built my website I really hate AI.

Um and I was reading some of the responses today because they are still coming in. If you want to rant about hating AI, head to I really hate ai.com and um someone ranted about Zuckerberg and just like everyone hates him. Like no matter what he does, everyone everyone hates him. Um he's the most untrustworthy person in the world trying to build like something that is so like he's I'm gonna call it now. I don't think he's gonna win. I don't think he's gonna make a dent. I think his reputation of his company, no matter what he does, everyone hates Facebook or No, everyone hates Facebook. That's the wrong way to look at it. Just because you hate something, whatever.

I think that just a reputation's been ruined that the trust is gone, you know, and it's like over multiple times it he he's done it again and again and again and he's just not trustworthy. uh when when it comes to your data, I believe the only thing that I I found to be like genuine about him is his interest for um like VR and even he even changed the name of the company to meta for metaverse and for a for VR. But VR might be something like 3D that never happened. you know, we had a a boom with 3D TVs and glasses and all of that. It never happened actually. And VR, I would be sad to see VR gone, but that's the only place I would see Facebook doing like a great job.

I mean, they bought Oculus. It's not even that it's their product, but he genuinely likes that. And for why he is doing AI, I don't even know. like I don't I don't feel like he's genuinely interested in AI the way that genuinely uh Sam Altman is interested in AI. Sam Altman is the AI guy, but Mark Zuckerberg is to me is not the AI guy who's like genuinely interested in it. Even like I was surprised to see Elon Musk um you know getting that good with Grock but because he neglected AI for a long time. He was never that huge AI fan. But don't be fooled. Don't be fooled. Right. I And I agree with you.

I don't think he's there. But CEOs have to act in the best interest of the company, right? And AI is a necessary trend that people need to leverage. Like no matter what you think, I'm sorry, but AI is just here to stay. Right? If you're tuning into this podcast, I'm preaching to the choir, so let's just not bang on about that. But the point is is CEOs need to adapt. And I'm learning in my company right now. I'm learning to adapt. And I'm I'm targeting an industry which I'm not that interested in. But I believe that it's going to I'm still going to be doing websites. I'm still going to be building apps, but it's in an industry where I feel like that's where the money is and that's where the business needs to lie.

And I'm developing a passion for it, you know. Um, I don't underestimate how cutthroat and how um I was going to say soulless, but I don't think I mean soulless, but how um yeah, don't don't believe for one second that Sam is just in love with AI just because it seems like he is or Elon Musk. Like you just you just you just answered your own question in like I didn't believe Musk was. I bet he isn't. I bet he's just chasing what's necessary for him to earn money. That's just a CEO's job. If a CEO's been if if a CEO is seen to act otherwise, then they get thrown off, you know, but this ain't a business podcast, so we'll uh we can sidestep that one.

Um not that I like it that much and I know you don't like it like looking just looking at the benchmarks. Benchmarks are only good until the next model comes around and naturally the next model is going to be better. But the I think the interesting thing about this model is that it doesn't aim to be the best. It's just it's it's just interesting to see how comparable it is to some of the other, you know, state-of-the-art models. And you've got the the the two open weight models here competing quite well. I mean, this is um this is without tools. Uh sorry, this is with tools, without tools, with tools, without tools. And it's compet it's still not you know it's competing very well with 03 and 04 mini here.

So it's just it's just by comparison really and it you know it it stands it stands its ground. We'll look at some charts later on because um there's been some interesting turn of events with the with the other the other model uh that open released. But um yeah the they're not there's no it just it's just competing very very well. And I actually like 04 mini quite a lot. I haven't I I don't use 03 or 03 mini too much. So, but 04 mini is pretty pretty good and it it shows in these results, but it can hold its own, you know, it can hold its own. So, I'm I'm it's pretty cool to see this out in the wild.

And um yeah, I I think the the the other thing to end on really is like I really think I really think um these open models, these locally runnable models, particularly the 20 billion model, like I've now downloaded that onto my machine, you know, and I've got it ready to go whenever I need it. And that you have to remember that is the world's information in your pocket. Like I know we've got the internet, but to have it in a in a in a system in a way where you can ask a very specific question about health. It's about, you know, we talk about health bench and stuff like that. To have that in your pocket is just next generation and yeah um I'm excited about it.

So how does it work when you run these models on your local machine and they need a tool call for example to use the brow? Do they use the browser? Do do they have something internal? How do they access the internet? Well, I mean just because it runs on your local machine doesn't mean it doesn't have it. What do you mean? Like if you if you don't have internet connection, it can't run tool calls. Tool calls are separate from the model, right? Yeah, that that's what I'm saying. Like how what does it happen? Essentially, if you don't have internet connection and you ask it a question, it just can't run the probably just gives you a warning.

I don't know. I'm not too sure. I don't know. I haven't I haven't got a a system set up to you know but that that is genuinely that is genuinely very interesting to what what you say like have it locally and it will just get better from here I I mean we've been saying this and we are continu we are going to continually say this it will just get better and we are exactly at that point we are seeing the next generation of the models speaking of which uh we should get into it after. Well, I mean, I I'd like to know what your thoughts are on this potentially being a foundation for Apple Intelligence.

Oh, that Yeah, let's talk about that. Um, yeah, it runs on 16 gig, which is exactly what Apple Intelligence supposed well is supposed to run on or what they're trying to get it to run on. Um, obviously they've had a massive focus in privacy specifically, you know, health as well. I mean, Apple are all about health right now with with watches and so forth. And, you know, don't get me wrong, these are all very uh good things, but like they seem very well aligned um with Apple's values and principles and it's free to use, you know. Yeah, that that sounds that sounds great. But when you when we when we heard about um OpenAI talking about running their model on on a on a hardware piece like on a smartphone or a laptop I was actually thinking about something else.

I was thinking about not Apple but Johnny IV and their own product that we haven't seen and we haven't heard anything about it 100%. They are like optimizing for that. Yeah, 100%. Yeah. Um, yeah, it would be stupid. I mean, these things need to be offline. Yeah. So, um, it makes sense that they would build an offline access, but yeah, I know you're totally right. Yeah, they're optimizing for their tool. Yeah, for sure. We didn't touch on the update on that, but they did they did actually release a really strange update which didn't make much sense. It sounded like that Johnny IV is kind of no longer part and let me get the let me get the website up.

Um what's it called? Open AI and love love from ah the name is on top of my I had to I had to IO um let me let me share this with you. We're going off piece now, boys and girls. But uh so it says update in 9th of July. So it was actually just a month ago, less than a month ago, has officially merged. So IO products has officially merged with OpenAI. This is the confusing bit. Johnny IV and Love from remain independent and have assumed deep insight and creative responsibilities across OpenAI. So it's like what do you mean remained independent? Wasn't Johnny IV part of IO? like what are they talking about here?

And you can see this is the update and then this was the original. This so there's only been one update since and I just I had to I mean I still don't know what it means. Let us know if you know what it means but and they have assumed deep design and creative responsibilities across OpenAI. It doesn't seem like integrated to me. It seems like they are just partners, but it looks independent. It looks to me as if like Johnny Iv has just sold the IP or he's sold his part in IO. Um because the team has officially merged with Open AI and then Johnny Iv's like okay but I want to I want to stay with love from you can hire us independently of whatever we do.

So, it's like he's No. God, this photo is just ridiculous. I wanted to say that. I wanted to say that. Oh my god, that's so Anyway, it it's it's going to be a meme for decades, this photo. It's already a meme. Like, people, it's just so cringe, man. It's so funny. Anyway, sorry. Um, yeah, it seems like Johnny Iv has let go of IO but has committed to it overseeing it as part of Love from. Yeah. So, yeah, but they've said that without saying that that Johnny Iv is no longer part of IO or at least he doesn't own IO or he's not dedicated or you know um committed to specifically to IO, but who knows?

Anyway, uh should we go to unless you got anything more you want to say? No, that was it. That that was it. Nice. All right. Well, let's go to a quick break and uh we'll we'll well we'll get into the latest and greatest. Support for this episode comes from Flowbase. If you are building in Web Flow, Framer, or Figma, Flowbase can help you build much faster. With over 4,000 components to choose from, you have a huge variety of wireframes and super clean, nicely designed sections that you can put together by just simply copy pasting them to your project. Over 15,000 icons as well, over a,000 illustrations. They also have a super helpful Web Flow app called Boosters.

With boosters, you can add things like sliders and mares and countups to your project with a few clicks and without any coding. So yeah, big thanks to Flowbase for supporting the show. If you want to try them out, go head to flowbase.co. And just for you guys, I want to show discount command flowbase all capitals for 20% off any plan. This is a limited offer, so get in there early. All right, that's it. Let's continue with the show. And here we have GPT5, the uh the big boy where everyone's excited about. What we were waiting for, which Sam Olman teased uh just a few months ago said is coming this summer and we finally got it.

Um and it's pretty good. I'm hearing I'm hearing a lot of still mixed but pretty good things about it. Um and it's pretty gosh darn cheap as well. And and that's actually for a reason that you don't think about, but we'll get into that. One $125 per million input tokens and 10 mill $10 per million output tokens, which puts it in line with Gemini 2.5. Um there are three models been released with a fourth on the way. You've got obviously GPT5. You got GBT5 Mini, which is 25 cents uh input, $2 out. uh GPT5 Nano, which is uh five cents per million input, 40 cents out, which again just two really fast performant models um and alternatives to the big big daddy.

Um 400,000 uh token context window uh input context window with 128,000 output. Token caching offers 90% uh discount on those things. So that's a pretty big savings to be honest. Um uh knowledge cutoff date is the September 30th of the year 2024. So still, you know, it's it's funny how behind we are in some ways, but GPT5 outperforms or matches models like Claude 4, Grock 4 at a fraction of the cost. And again, we'll get into that. One thing that's really really cool about um GBT 5 and very very very impressive is they've got hallucinations down to 1%. And you know, obviously, ideally, we want 0% hallucinations, but you ask a human, humans probably hallucinate more than 1%.

Do you know what I mean? Um, much more. Yeah. Yeah. And you got s psychop. I can't say it. Overaggreeable. Overaggreeableness. Sycophony. I should have really looked at how to say that. I can't say it. psychopy overaggreeableness, right? So, um we've seen this this glazing. So, in in simple words, it's going to agree with you less, which is a good thing. Um these AI models in general, whatever you tell them, they tend to agree with you even if you are dead wrong and they glaze you. Uh and it's better if they are less agreeable. I mean I prompt prompt all the time don't agree with me give me new perspective like be critical but now this model is going to do this out of the box.

Yeah. So you should be seeing this is available through the API but you should see this roll out um in your chatbt throughout the week. I mean do you have it? A lot of people have pretty much already got it to be honest. Um but yeah really cool. Let's let's go through the article and start picking stuff apart. This is this is it's it's cool because it's like it will it's not a traditional rooting system but it will figure out it's a unified system um that answers more questions and deeper reasoning model thinking model for harder problems and and a real-time router that quickly decides which is something you've been after for a while uh which used based on conversation type complexity and tool needs and your explicit intent.

For example, if you say think hard about this in the prompt, it will obviously think and do more things. So that real time routting is really cool. You just you just set it and forget it. A real good general model and apparently is good for like um coding as well and specifically UI. It's it's really quite a modern clean UI. Do you know it's just struck me actually. Maybe a lot of the UI sucks because the training data is only up until like you know 2023 or something like that whereas we've reestablished a bit of a trend over the last year which GPT5 might be up to date on anyway. That could be that could be but in general like um making models being good with backend uh code is difficult because there aren't as much backend code available.

It's you know literally it's hidden and what is open source is you know limited with with front end all frontend code is available but tons of it is trash I mean look at all websites made with WordPress the quality is trash like if you think of websites their design like the UI and also the code of that UI I would say mo like the overall internet is way below average. That's why Well, I don't I don't think your argument holds water. I tell you why. Why? Because because the front you say the you know the front end code is all available whatever. No, it's not. Because even WordPress websites are pre-built. The you're not looking at the PHP code.

HTML websites and things like that. It could be a Nex.js or React website. So you're not seeing the JavaScript that's used to render that page. Yes. Basic HTML. Yeah. And even even when you do look at the JavaScript, it's all minified. It's all whatever. So we're talking about your it's very equal back end and front end because at the end of the day, it's all trained on public um uh data, GitHub repos, right? So I would say it's pretty pretty pretty even in that respect. Um interesting. I can't answer the question as to why these vibe code things can't you know do the full stack development and why they mock a lot of data but you know if I could answer that I'd probably be earning a lot more money but even like the when I talk about UI I didn't mean I know JavaScript is technically more like front end even it's used like everywhere but just the HTML CSS that is out there is extremely low quality but HTML CSS structure is nothing like too complex, but still like AI hasn't been, you know, so far I haven't seen them being doing a great job at structuring all of that like a pro front-end developer can.

I mean, you know, at the end of the day, it it is what it is, but it's um yeah, the the the UI I mean, the UI is is pretty good. Can can you please like uh scroll up and in the video show they show one piece of a UI that it seems like that it's created by GPT5. It blew my mind when I saw it. I was like is it's like a diagonal slider because I want to make actually we're working on a diagonal slider for our component library. And when I saw this I was like that is insane. Um, it comes somewhere. Are you sure it's this video? Yeah. Yeah, it they they show it in a piece where there is like a little bit of code.

Um, and there is a slider that is like a UI piece that is diagonal. Here we go. Here we go. Create beautiful infinite because here's the thing. It's good at it's good at onepage websites is the thing. I don't think it's good atjs. Oh, this thing. Yeah. I mean I this thing was all right. I mean I didn't think that was I mean that game was symbolic of its UI. I think the glowy neon thing but um there will be there will definitely be you know you know when a when a um a website is AI generated because they will all they all have their distinctive look and feel and I think we'll start to see oh that's a GPT5 generated website.

Yeah, I I want to try this model and see if it can help me in cursor. Uh because I I showed my like I genered what I was coding with cursor with AI. I made a react app. I made a very simple like the simplest React component that you can think of all with prompting to an AI inside of cursor. and I showed it to my uh developer to I could which we had on um two episodes uh before he had a stroke. He saw my code and he just had a stroke how bad the code was like it it was everything was depreciated like uh the syntax was we had just so many errors and the code extensions like even was wrong like so many things was wrong and he told me like how to prompt but these are things that I didn't know like I didn't know that I should build like react vite and I just said react app and ai assumed whatever it did and It used like very old type of building React apps, not the modern way of doing it.

And again, I I have to emphasize it was literally like a hello world type of thing. That was a disaster the code. Um, and then later on, of course, when I understood a little bit more, I got into it and I started to fix things. Uh, but all of that that to say that we see these like really cool show reels videos, but in reality me as a somebody who's just starting to picking up on development, I I'm nowhere close to getting those results. Not even close. Yeah. No. No. Like you say, you know, your the results you get from AI is average at best. It's always going to be at best. Yeah. So it's an important thing to remember like you think you you're you're a new de not you.

I'm I'm you know someone is a new developer. They come in they think they can write code because they've got AI now. Not even close buddy. Like it's it's it's a good starting point and I I absolutely implore you to enjoy it and and launch that website and whatever but understand there's so much more to it. Like you only you only see that when you know let's say you're you designer, right? you you'll only see it when you you have people come into your industry and you and um you realize how bad it is like it I let's say I don't know copywriting you know I'll be impressed with AI copyrightiting and a copyright will be like no no that is junk like that is junk so um but anyway let's let's keep going with the GPT5 stuff um but I love this real and I saw something here.

Um, uh, once users limits are reached, a mini version of each model handles the remaining queries. That's nice, isn't it? Um, and they they, uh, in future, we plan to integrate these capabilities into a single model. That's interesting. So, like it's almost as though they'll have a little backup model like the 20 the 20 billion. Oh, I don't actually know how many parameters this is. But um Oh, they don't release it because it's closed weights and it but like um yeah, it's like mixture of experts, but it's like not because that that little extra expert is probably a a distilled version of the of the knowledge it needs. Really cool. Um we'll get into benchmarks as well, but I really all I really care about is real world uses because this will be as good as the next model.

So, we can um saying that we didn't touch on um uh Claude Opus 4.1 being released. There's not a lot to to really go into, but that was a that was a nice little release that happened this week, too, but obviously overshadowed by by this. Here we go. The GPT5 is our strongest coding model to date. It shows particular improvements in complex front-end generation and debugging larger repositories. debugging and then front end generation which I've only really seen like again onepage single page things rather than actually um creating a decent code base as were though it can debug larger code bases which is pretty cool um and I saw the video and a lot of people were oneshotting their their their basic piece of software which is really cool I mean this is just nice is it oh I can play it hey it's got sound and everything it's nice and Oh no.

I don't know. I'm uh you entertain surprised. I'm surprised that it's like I don't know. Is this like done with one prompt? Um hope all pixel art. Um here we go. Yeah. GBT5 created with just one prompt. One prompt. Some pixel art here. I assume that there are a lot of like open-source data around these type of games because man with one prompt I can't get even a simple react app you know it's working but it's done in the most horrible way so I don't know I see that you're really highly focused on this game keep talking keep talking I'll be with you in a Okay, I've lost my flow now. I accidentally double, you know.

Anyway, you're doing great. 40 40. What What's your word for me? Do you know? Uh, no. I I don't know. But far like worse than yours because I I Do you type with like 10 fingers? Yeah. I have I don't You hear that? I do. Oh, that that is interesting. I don't know what I'm supposed to do. But why is the You can change the speed. Why is it like Windows XP? Ah, it looks like something edition. It's just a style. Ah, you see the prompt by the way. Um, if you scroll down, you see the prompt. Generate a React plus canvas lowfi visualizer. Okay. Yeah. vapor wave trap. No file uploads, bundle tone. Here you go.

Um, creative expression and writing, which traditionally Claude was good for or apparently good for uh creative writing. It's not my field. I can't can't uh confirm or deny such things. Healthbench making an appearance again, being good at healthbench, which is cool to see. Poetry, look. Um, and here we get into some of the things. Now, there was some very misleading graphs which they have since fixed uh but we'll get into that in just a second. You can see some of the comparisons here. Now, this is GPT5 Pro. Now, this we haven't mentioned this just yet, but there is a GPT5 Pro on the way. And I don't know how expensive it's going to be, but the thing we haven't discussed, I alluded to earlier, is that now while this is the same price as Gemini 2.5 Pro, $125 in, $10 out, the efficiency of the return tokens and the reasoning that um reasoning tokens that it takes to actually give you a response is so much better than um than Gemini.

I think the real real number we need to look at when it comes to price is price price per like token or price per prompt or something like that. So, you know, whilst again it's it's already I think G um G 2.5 is pretty reasonable, the fact that it's the same price but it's way more efficient means your average prompt takes far less tokens, far far less. So, it's really cool. Um nice in that respect. So, I don't know how much GPT5 Pro is going to be, but that is that is a model that's newly coming. But you can see some of the stats here. It's really uh button up against some of these sort of numbers here.

And again, from a price like I think 03 is quite expensive. I think 03 is like really quite expensive. I forgot what it is. Frontier math obviously doing very very well here is being out 04 mini and 03 here. Um GT5 is with that's Python. I don't understand Python, you know, whatever. Um, and then we scroll down. I mean, humanities art exam, which, you know, we've we've all heard it's a very very very hard exam. They're not comparing against other models, which is a bit of a Yeah, but from what I see, it's literally 0.4% better than GPT chat GPT agent plus browser. Like, isn't it like a very small increment incremental change? Well, here's Yeah, you're right.

like 20. What are you talking about? Here to this and then this to this. Is that what you're talking about? Yeah. I think very little. I think this is why I don't care too much. I mean, it's good little, you know, good little indicator these graphs and stuff like that, but at the end of the day, it comes down to real world usage. And I think the um the the the feeling is that the real world usage is actually something to to take note of. Um but saying that uh these graphs were very different 24 hours ago. So we had 52% showing more than six like 69 and 30 were both the same here and it was showing that 52 was um higher than these sorts of things.

But now they've they fixed these graphs now um that are more accurate. So yeah, you can see sort of the performance here. Um and the again the 03 is just so expensive. So the price price um difference to to get that performance for the sort of price we're looking at now is really really cool. Um good for agentic and tool use as well which we love as coders long horizon tasks and obviously tool usage and things like that. Just um really cool. Oh what was it? I don't think mini accepts images. This is the thing which was a bit of a letdown I think is that it's only text and image uh up to yeah mini I think nano is image input or whatever.

So I think that's the biggest letdown on this model is that I think we were expecting like kind of like a Google moment of this model can do all of this stuff like it can make videos it like if you put in you can give videos as a source or you can you can make code from a video in Gemini 2.5. I think I think we were expecting something more than just a textbased input output thing. Do you know what I mean? The multimodal capabilities are just not not impive at all to be honest. They they accept images. That's it. So, a bit of a shame there. Um some more benchmarks. Um blah blah blah.

I don't think there's anything else. So, yeah, I think really the the real uh it just comes down to what people are using and what people are doing with it. It's not. We are talking about small percentage increases to be honest, but it's efficient at doing so. Um, and again, it just comes down to those real world usages. So, I'm um I've switched everything that I do as my um automations. I've changed to GPT5. I'm curious to see how uh how good it's going to be, how um are you noticing any difference? I haven't run any of those automations yet, so I I've yet to I don't have any real world usage, unfortunately. So, from what I understand, the chat GPT5 is just overall a bit cleaner, a bit cheaper, a bit like having everything put together, but it's not a revolutionary model.

It's not significantly better than the previous models. It's just one simple, cheaper, faster model that I think it's consistent as well. Like with when it comes to hallucinations, did I just see something around that around here? Um, less agreeable, fewer unnecessary emojis. Um, I think they said something about M dashes as well. Yeah, like the AGI we were waiting for. We are going to give you fewer unnecessary emojis and m dashes. I already have this by the way in my prompts everywhere. I just I have it. Don't use n by the way. I learned this is I learned what n and m dashes are because of chat gpt. So what's an n dash? I don't know what an n dash is.

So we have dash. Okay. So we have dash n dash and m dash. Dash is small. N dash is a bit larger. And M dash is the largest one. You And do you know their usage though? No, I don't know. I don't know. I just know that Chad GPT uses them all. And genuinely, if I see any of any type of dashes in any text, I skip reading the text. Wow. Yeah. See, it's ruined it because a lot of people do genuinely use them. Um, yeah. So, I I know what M dashes are for and I know what dashes are for, but I don't know what the N dash like. Uh m dash is when you're like changing you you're you're it's like a side thought that you're changing the topic and it's like you know and then blah blah blah blah and then you come back to the topics then you parenthesis for everything.

I know it's not the right way. Commas parentheses are often you know they're more they're still on topic or they're still um related to the main thread of the conversation. Either way parenthesis either way. Yeah. Either way. dash like um I went to the shops today and it was tipping it down or no not even Anna was tipping down um which is five miles away or something like that. It's just it's just kind of like off the it's just like it's not even a by the way it doesn't doesn't complement the story. It's just a it's completely off pieced point to make. Yeah. And another thing that um chat GPT does by the way it's not just the visual the visual one is you see the m dash in the text and then you you definitely you notice that it's written by chat GPT with a high chance right another thing that chat GPT does it uses this like kind of uh se this structure in the sentence that says it's not just this but also that and that so it's kind of like under emphasize in a way in the beginning of the sentence.

It's like it's not just that, but then you know, put the put all the emphasis and the positive stuff at the second half of the sentence. And we do that all the time, but um I see like Chip does it in a kind of noticeable way and then I can see it. And yeah, uh you're right. they're ruining uh these and and maybe it will fix maybe they fix it and then I can read text with m dashes and be like okay maybe it's not you know just garbage they don't actually explicitly say it but I have I've read somewhere that they they have they they call out less m dashes but the point is is we talk that's the way that we talk we talk with m dashes if you were going to directly translate the way that humans talk then we talk with m dashes is we just don't know it.

We don't realize it. So, it's imitating human speech, not human writing, which is we were talking about language last week and and how it interprets language. And I think that's what that's what it's that's what it's um lacking. It's it's lacking the understanding of the way that humans actually write. And we unfortunately we write with grammatical errors. We don't Yeah. Again, we don't we don't write m dashes. Do you know what I mean? Well, only the most, you know, well-versed people use M dashes. And I actually started using them once I knew what they were for. I started using them and then literally it was just like the biggest thing like the biggest giveaway of TouchBT um or AI.

So yeah, let's let's talk a little bit about the pro. I've not really read too much about this, but I'm I'm excited about most challenging for complex tasks. We are also releasing GBT5 Pro, replacing all uh OpenAI 03 Pro. Um yeah, so it's replacing 03 a variant of GBT5 that thinks for ever longer using scale but efficient parallel task time compute test time compute provide the highest quality and most comprehensive answers. So I mean anything with pro in the name, you know, you've got to really know what you're looking for. Um in terms of benchmark includes state-of-the-art performance in uh contains extremely difficult science questions. I think this will be definitely uh you know something on their $200 a month plan or people who are doing big you know scientific mathematical problems like I'm yeah we're probably we're probably not worthy of the pro um you know manica Monica Monica uh but yeah so it doesn't say anything about timing doesn't say anything about cost but it's coming soon.

So, check if you're using it. Um, oh, this is the thing I want to say as well. I like because I'm a Claude Claude code. Oh, I I don't like the UI of Codeex. I really don't. And like do I do I do I use codeex just to try it? I just I just don't know, you know. Um, I have other tools that utilize aentic workflows like I use warp as well. uh that can you like it could that can use GPT5 but it's it's still it's not it's probably not as good as this but you know I might have to give it a go just to just to see what it's like but yeah GBT5 it's out there in the world we're all very excited um what more can you say um I'm looking I'm trying to compare chat GPT5 five to check GPT 5 Pro and also 01 pro but we don't have any data on 5 pro.

So uh and why 01 pro? 01 Pro was the only G the only AI model that I genuinely thought it was like brilliant or not brilliant but like like it had the wow effect for me and it's like by far I believe the most expensive model but 01 Pro was the model I I used to create that prototype that I made with the dots. 01's really expensive though, isn't it? 01 was ridiculously expensive if I remember rightly. Yeah. Um, if you you needed all the help you can get. I I shared I shared the link. Maybe you can paste that and we look and then the the pricing is just like incredible. Like it's it's not even close.

Look at the output. $600 versus $10. Yeah. This is 05. Yeah. Yeah. Um And you used this to code, did you? Yeah, I did. I did. I I used this and it was it was so good. And I was comparing it at that time. I was comparing it to uh claude 3.5 was at that time with everything else that was available at that time and nothing came even close. Yeah. I don't know what is this is that oh it's audio. So it doesn't even take audio. I'm just flicking through here to Yeah. Yeah. It's just just images and text. All of them just images and text which is a real let down I think. Um, you know, by the way, no.

Do you notice the GPT5 has a knowledge cutoff by the way? What do you mean knowledge cut off? Uh, scroll down a little bit. You see the knowledge cutoff um is Octo October 1st 24. That is Yeah. Yeah. We we we said this but interestingly they say September in the release like nonetheless it's uh one year behind I mean these models can reason and they can also uh use tool calls to access the internet but yeah um like I say this is probably why UI sucks because before then like I don't know what the cut off point was on some of the other models but you know maybe we're just so maybe Maybe UI gets outdated that quickly that we see UI from 2023 and we think, "What the hell?

That's gross. That's what you were creating." Um uh and this is why like again, so if I want to scrub up on the, you know, the current frameworks, the most hot frameworks people are using in in the world of development, I'm only going to AO's only had to tell me what was going on up until 2024. So this is where you need to get out the old books, do your own research and and whatever. So AI is always going to be AI is always going to be behind average at best. But yeah, um yeah, very very cool. Um I think that was all we had for GPT5. Should we should we wrap up with a bit of light-hearted familyfriendly content?

Sure. or what do you family friendly? Why do you focus on that show? I don't know because it's just, you know, funny, isn't it? So, basically on the 21st of July, um they there was a funeral for Claude Sonnet 3. Anthropic retired Oh, no, sorry, sorry. July 21st, um they ret anthropic retired Claude Sonnet 3. So there was a funeral held where more than 200 people gathered to mourn its passing and they had all these mannequins that kind of represented different models of of of Claude and there was a freaky one. This this is Claude Opus. Um and it just has this kind of I don't know. Oh, this is Sonic 4. Sorry, not not Opus 4.

Um, so Sonet Claude Sonet 4 attended the funeral um and was desperate to talk about his research holding in hundreds of pages of Claude's three sonet outputs and elicited I don't know what that means but yeah um really and and here's the thing like for whatever reason Claude has a a fan base that this is like a and it's like poetry was it like by is it by fans or like by anthropic? No. No, not by anthropic. Although Anthropic employees and OpenAI employees did attend. Uh, one woman, Amanda Ascal, an anthropic researcher, has jokingly called herself the fairy clawed mother. Um, people are dedicated, man. People are freaking dedicated. I think um, Anthropic is a company that is really easy to like.

Their CEO is super likable. Sam Altman is really not. I mean, he's okay, but he's not like really likable. But you look at what was his name? D Amados. Yeah. Yeah. Something crazy. Yeah. The CEO from Anthropic. He's just so like I I watched a podcast um that he had with Lex Freedman. And since that time, I'm like, this guy, he's so likable, so easy to to like anthropic. And I think their branding is also really nice. It's it's very human. their branding is really like close to you and OpenAI is I don't know their their branding and everything that they do is not really that likable. Even the the live streams that they have, it sounds bor.

They have like 25 minutes of a live stream that could be just five minutes. And I've heard I've heard the uh GPT5 live stream could have been a five minute thing, but they just it's always like this. And it's funny like if you go to their live streams, there is always a comment that people are saying they are uh going to skip this live stream because they are going to watch the fire ship video of it. I see that. I see that comment everywhere now. It's like great original guy. Yeah. Like oh I see it everywhere now. Yeah. I bet fire ship are like well that they've become like a meme but it's just read it everywhere.

Read it everywhere. Yeah. Um, yeah. So, what are you what um you know you you say you're back on CL like what are you what AI are you using to code that? Are you using Claude or are you using Actually, I'm using all of them. I'm just trying different ones. Just whatever can make up for that lack of I I you know the the Max version I try to you know use use the one Yeah. Yeah. I but but none of them like I have to emphasize like none of them even came close uh to to do what I wanted to do. So, I I made a very simp I can't talk about like the specifics of it because I'm under some NDAs, but I made like a the simplest React app that I could and AI was like really bad.

And then I with the help of my developer we fixed things and then I understood things and then I created a document um like just a text document from all the learnings how to make everything work and the structure of the app that I want and everything. And then I took that and then created a new project. Uh, I cloned um like I I took I forked a GitHub project for a very fancy transition like on an image like a web WebGL effect on an image that transitions like um from the edges with like a really nice ripple effect. It's not a simple ripple. It's like very fancy to all of that. But even with all of those complexity, it's still kind of simple for what it is.

It's a React 3 fiber project with one specific animation. There are not like crazy stuff going on. It's just one specific nice fancy animation. I wanted to turn that into like something else and like it was hilarious. Like the result is literally a fade. So take that like fancy animation. What AI gave me like after hours of trying multiple times is every time a simple fade animation at best. It's just like the image fades in and out like opacity like no where is the you know all the web it just it just fails at um doing what I want to do and I know it's on me because I'm not you know a developer who who understand GLSL uh I don't understand webgl I don't understand react 3 fiber these are quite advanced like the maximum I can do a little bit of JavaScript not these like much more advanced stuff.

But that is where I would say AI is. If I'm not the person who understands these things at least to a good degree, not you know, I can't be even a beginner. I have to understand it in a really good uh degree. I can't do anything like really advanced with it. I think I'm going to I'm still formulating my on this, but I'll hopefully be releasing a video on it. basically how you know you shouldn't need you shouldn't need to understand code like you with with AI now I think we can not that we can get away with not understanding code but I think that's the that's the kind of the end goal is to you shouldn't have to understand code yes um but using that but but using that as a framework like the way that you AI code is not prompt the thing that's on top of your head the way to AI code now is again they call it context engineering it's creating as you created the learnings the text all that sort of lot there's a lot of boring unfortunately there's kind of a lot of boring admin that now goes into coding it's not this thing where you just know stuff anymore so someone someone like you like coming in and just thinking you can prompt and and create what you need like I mean that's what vibe coding apps are for that I think you're probably better off just jumping into replet or lovable or something like that.

Whereas if you wanted to say in cursor or you know Kira is an interesting um addition because they start to formulize the different files that you need to to actually build and code a website because it automatically creates a bunch of of those context files that you you added at a later date. then only then you start then prompting your features or you build your PRD or something like that. So yeah, it's building the ecosystem and the context around the the product before you start throwing top of mind, worst case scenario, top of- mind prompts, but hopefully some more wellthoughtout prompts that directly um you know um uh directly reference like user behavior or or this or that, you know.

So it's I think we I don't know. I can rant on this for ages and I don't I don't want to do it now but like we need to change now our perspective. I think we need to change our perspective. We need to change our tolerance and understanding of what development what what a developer does nowadays when you've got AI because it's not sitting down at a code editor and just writing stuff that's in your head. We the role is now changing and that that's what I'm trying to formalize. Yeah, I understand and uh I'm going into it with that mindset as well. I I know I won't be like literally typing lines of code.

I know I will be more prompting. But even in prompting, if you don't know what you're doing, it's so easy for AI to get it wrong, even on the simpler side of things. And the more complex and specific it gets, it seems like it's it gets even more difficult for AI. And there is um yeah there are so many things that can go wrong and break things. So for prototyping it does work but it's still not where I want to be. I I wish I could that I have so many like crazy ideas. I I'm I'm not you know I I don't ask AI for ideas. I have a ideas and I ask AI how how I can implement them but it's not even close.

And I know I'm the limiting factor at this point. I think if I knew if I could develop uh and I had a better understanding of for example WebGL, I could do like more crazy stuff. But that will always raise the the ceiling will raise with my knowledge. So if I understand WebGL, then I probably will think of even more crazy things that AI can't make at you know, you know what I mean? So it's always that ceiling that raise and our perspective and our way of using the tools will continuously shift. right now. Uh I think if you are a developer, you you can think about the architecture and think about the functions and how you want to have these functions talk to one another and you ask AI to create these functions with the variables and the output that you want without typing it in.

That's probably what um you know a developer does right now. Yeah. Yeah, for sure. Um, let's wrap it up. I think that's the weekly news. They've heard us rant on enough. Um, yeah. Cheers for tuning in everyone. The 20 of you on probably on Twitter or, you know, if you're not on, come over to YouTube. Come over to YouTube. Um, join us next week. God knows what the AI world is going to bring us as we're always surprised. Uh, GT5 was released yesterday. you know, if that hadn't been released, it was pretty quiet. So, uh, yeah, come join us next Friday, same time, same place. Uh, yeah, by next week, we probably have, you know, anthropic maybe being on top of the chart.

You never probably, probably. It's a mad world out there. So, uh, with that, keep on vibing, baby. Yeah. Peace out.