#48 GPT-5.5, Image 2, Deepseek 4, SpaceX x Cursor

Are we a minute early? Are we Are we Are we on time? [snorts] We're not. This might be a first. This is break. This is breaking news. Yep, we're live. Gosh. Nice. God damn it. Who'd have thought? Well, I'm going to give us a like on YouTube. Get the get the ball rolling, if you know what I mean. Haven't even I haven't take a sip of my coffee yet. This is the first time I have had a post 100 p.m. coffee in many many months. I normally try I normally try and do not drink like caffeinated coffee after like midday. Actually, to be fair, I had a I had one yesterday. I had a really nice day yesterday. I went out. Oh. Oh, actually [laughter] stay there. Actually, what's happening? Okay, so Oh, I see something lime green. Sort this out. Hang on. You're still there. I am here. Still there. Right. I'm back. Sorry. Check it out. A Neo. Nice. Nice. And you got the green one. So cute. Citrus, actually. Citrus. Green. Matching matching wallpaper. Matching matching keyboard. Just slightly off white. It's not the one with uh Touch ID, right? No. know cuz I've got a look how cute it is. Very Do you know I I kind of I do kind of like the idea of a you know a beast desktop computer and then access editing machine and then cuz Yeah. I really like lounging around with a laptop. Yeah. um a small laptop but then sitting at my desk it is nice that's a very you know that's a a massive luxury to be able to do that and I don't necessarily I'm not necessarily able to do that but it is nice to be able to just have a little laptop for little whatever you know I was coding a new website I was moving our website onto Versel um and payload and stuff and XJS and I was just chilling out, sat in my bed. I just there's something about coding on a laptop that I really really like. I feel like I'm like I've I don't know. It's just so hard to explain. I just really like coding on a laptop. Anyway, I went out and because I returned the MacBook Pro because I had a MacBook Pro that I tried, the M5. I don't know if you've been seeing my videos go up. And then I took that back and then bought the MacBook Neo and just sat in a coffee shop and I took my I took my big laptop and I just sat in a coffee shop and worked for the afternoon. Bought myself a little sandwich and had a and had a caffeinated coffee. That was the That was the last time I had one. But I need this one. I was just I was treating myself yesterday. I didn't need it. That That's nice. Yeah. Um what do you what do you want to use the MacBook Neo for? So, I'm going to set up a a a um rig where I've got local AI running on my big laptop connected to the Wi-Fi and then access that local URL on the MacBook Neo. I also think the MacBook Neo is ideal for open claw because it just doesn't need anything if you're I mean again you know if you're not running local AI of course it's a perfect machine I think because you've got the screen you know if you're buying a new laptop you've got a little screen you can do is open a thing still I think yeah I think it's like a I think it's like a staple now yeah oh uh I'm using dispatch. I've I haven't been at my desk as much this week as much as I wanted kind of a lot uh to deal with. So been working quite a lot with dispatch. Yeah. Uh just from my phone. And I noticed T3 code has added connection. So I can run T3 code directly from my phone, but I have to be on the same network. So, I guess in order to to access it from uh outside, I need to set up a VPN at home or something like that. Well, no, cuz dispatch will dispatch to claw. No, no, not T3 code. I I know. Um, dispatch has computer use, so you should be able to dispatch computer use to tell it to go use core code. Uh, T3 code. Sorry. I I did I did um funnily enough I did and somewhere in the middle I was like what's T3 code doing and it was like oh you know what actually I just spun up an agent to do it with clot code directly and didn't give it to the agents in T3 code and I was like well it kind of makes sense but I wanted to have it in T3 code so I have the chat and the nice UI later when I come home otherwise it doesn't make it. I hear they're working on an iOS app. So, that's um that's coming. Um also warp, you can use you can use claw code in warp now. So, you should see a little remote button down below your claw. So, that just sends it up to the cloud and then you can I don't know how you interface with it though. I don't know how you can use it. But, um have you been building anything cool? Yeah. Yeah. I so I know one of the biggest news that we have this week is the image gen and I had an idea with a friend uh who mentioned something and I had this thing in my head on my head for quite some time. It's not new and I I just prompted to existence mostly from being away from my desk mostly you know doing it while I'm on the way and I can share that it's actually pretty cool. So before I share that um I want to explain a little bit or just just a tiny bit. Why do you think restaurants don't have menus with pictures? Like it's a genuine question. Yeah, some people some people actually think it's very tacky. I actually too tacky. Yeah, some people think it's tacky to show photos, but then I don't know what that's, you know, derived from. I don't know what that idea because at the end of the day, you know, you just need to know what's in it. But we all know that you do eat with your eyes sometimes. You do look at something. Is that Yeah. But like if you're a not native person especially if you are not native it becomes even like harder for you to rock your head na n n n n n n n n n n n n n n n n n n n n n n n n n n n n n n n n n n n n n n n n native to I'm I don't mean even the language but native to that food or place like I'm not native German and when I go to even after 10 years of being here if I go somewhere sometimes I'm like what is what is that like okay how does it look and I end up googling or I I noticed my I asked my friends they are like um they told me that they look at the neighbors the tables at the the next to them and they see if they have something nice and they would ask the they would ask for that they'd be like what is that maybe I can have that because as you said we we do want to see it right so why don't they have it okay maybe tacky. That's true. Also, probably very expensive, right? Uh what as in to take photos of the meals? Yeah. Yeah. A lot of people just use stock photos and you know it's not going to look like that when it gets to you. It's like blatantly a stock photo. But yeah. Yeah. Yeah. And it's not on brand and it's not what you get and it's not their uh plates. It's not their, you know, their restaurant. Okay, so here what I built. I was like, what if I use AI to help these restaurants create the images, you know, of their own food. So the idea is the idea is that they upload a picture of a dish that they took with their phone. Imagine this being taken with a phone and then we recognize the name with AI or they can type it in and then they can choose a style. for example, what should the the table roughly look like or exactly and then maybe the direction and then the direction of like the photo and then generate. That was one of my ideas. As you see, I have 14 more instances of this flow because I was playing with the UX. I was telling Claude and GPT 5.5 to create different alternative ways that a user could do this. So this was kind of like a middle one, but we have more advanced ways of doing it. We had also I I made more simpler version. You take the picture, you press generate. Everything is a prompt that I write behind this UI that goes to the API and basically retrieve the image um after we get the response. But then I was like uh what what else can I do? And I will show you a really cool one is this one. So, I came up with [laughter] this like full on studio mode where you can have a light pointed. Wait, let me first upload the dish we uploaded. I project it to the table [laughter] to the plate. [gasps] Surely this only works if you upload an overhead photo, though. Yes and no. So, I'm going to before we generate the main image, I'm going to actually send this image [clears throat] to maybe a cheaper model to create an overhead image of it and remove the background. Okay. So, so it's just like for the studio purpose. And then you move the camera, you move the light, you can add another light, you can change the intensity of the lights, the color of the lights individually. And everything together creates a prompt. And I'm thinking about like even including this a snapshot of this as a as a reference imi reference image for the AI to to also have this as a reference. So, I'm not that far yet, but yeah. Mhm. [clears throat] [laughter] It's it's quite uh more fun because now you can you can actually change the table. These will be like proper structure um proper materials for now. This is just for me to play with it. Uh the the plate, you can change the plate. Uh maybe I allow them to upload their plates, right? That that's easy. Then they can add fork and knife and spoon. [laughter] And the AI should pick this up and create something realistic because in the prompt I say make it realistic, right? Oh, of course. Of course. That's all you need. It's just a prompt. And make no mistakes. And make no mistakes. And you a loaf of bread and a drink. Yeah, that's actually Is it working? Not completely. It does work. It does work. But right now I I broke it, but it does work. It's at the uh in the admin I have like a bunch of prompts and I'm pulling all of this information together and it's creating a prompt that goes to whatever for now I have this UI for myself but obviously when I push it live it I I will build like in environments and everything I will uh take care of that with AI obviously. [laughter] But yeah, uh and and you know what I was like, why am I going to stop at this? If I'm making the image, I'll put the image in in a gallery, right? This is the gallery of your images. What do you do with these images? You put them in a menu. So this is a fullon CMS that they can you know like edit and you can you can add look you can even multi-edit stuff even web doesn't have this and I built all of this from my phone. Um then you can go to individual dishes change them take a picture make a picture for them edit it whatever. Um, so that's their full-on menu. They can import it with AI, so they don't have to type in all this information. Uh, and and then I was like, what do I do with this? Well, let's let's add some design. So, you can design your menu in a Yeah. with different fonts and different color schemes and different designs. and you know you have design capabilities and you can preview it and they can publish it and just link it on their website. So there's a full-on menu working menu that is mobile responsive that has like I will add search to it. I will add like nice information to it. This is a full-on menu. And then I was like, these these images that they have, do they want to only use it on their menu? No, they can share it on social media. So I added social media. So like here you can make a new post like a carousel. You can add the images and then after you select the images, you can let AI generate a caption for you [laughter] and just publish. Why lift a finger? Why lift a finger? Why lift a finger? So [snorts] like the fullon life cycle of an image for a restaurant. And then I was like, if I can do this for a restaurant, I'm gonna duplicate this and do this for different industries because definitely restaurants are not the one with the most money. No offense, but they probably don't have the the biggest budget to work with, but yeah. So, what other industries are you going to do? Uh, next one is real estate. Uh, the pictures that the real estate people take of the the house. uh uplift them with AI and maybe do some cleanup. For example, a house the tenant hasn't moved out yet. We can just remove it. [laughter] Yeah. Dirty washing. [gasps] Yeah. Um or like you know when when they move. Yeah. There are so much. So all of these are prompts in a way nowadays, right? the way it's just they could go to chat GPT and do the same thing with prompting but the idea is you can appeal to people who don't want to prompt who don't want to tingle with styles that that I build these guard rails right uh these presets and I I think the presets are the most important part so let's see let's see how you go Next year you might be a millionaire. Yeah. With with 50 apps. 50. [laughter] Nice. I'm I'm going down more of the route of Yes. I've got like projects and work with with people, but I'm building tools and things for my own work, you know, ways to speed up my own and and also save my own money as well based on subscriptions and things like that. So, you know, I'm I'm thing is because when [clears throat] you build an app for someone, you've got to really care about their industry and you've got to really get nerdy about their industry. Like it's all one thing to like it's one thing to like come up with an idea, but you've really got to you've really got to project yourself in the future just working on this one app like all of the time because it's if it if it blows up, you know, it's a business that you've got you've now attached yourself to. That's true. So, um and it all feels really fun. It's it's so I've been there. It feels it's like it's so hard to imagine at the beginning. You're like really excited. this is really cool and then it start people are starting to get demanding about this tool and it's like oh I've lost all the fun in it immediately. So you got to really be sure it's an industry and a tool that you should uh see yourself uh you know is it uh levers level levels IO levels io so levels io he or levels yeah levels he's he's making apps left and right and he's trying to see what sticks and I think parts of that is he's trying to see what also gets him excited. For me, this thing with restaurants, it was just an a problem for me. I wanted to see that in the world. I see this it's missing and I'm trying to build it and maybe I will get bored next week. Maybe not. I'm talking to a few restaurants. I'm talking to a few people working at restaurants or with restaurants and trying to see what's their pain points if they would be interested in using that. That's why I'm not so keen on polishing the UI yet. I'm building like different flows. I think people really get caught up in how the UI and stuff needs to look and it's just it can be a it can be a Figma prototype. It doesn't actually have to work. You need to just go as far as showing the concept that you can tell people what it is, right? And and that's that's the right way to approach it. Yeah. Um I think the the one thing that uh I think will be missing I mean you'll find out from from the people but like is that payment aspect of them yeah they can publish their menu but they want to be able to like buy like buy their and pay for their order. Yeah. Yeah. That side of it that would be different like I'm not going to attack the e-commerce side of things because then you have delivery to deal with. This is mostly for and this this is mostly to to have a nice menu. And I also talked to somebody who works at print. So I could create these designs and also create the printable version of them. So they can actually order the printed version. And it's it's easy to set up things like this. But yeah, you this is as you said like once you get deep in it, it's a it's a universe of its own and it's no longer your ideas that you want to do anymore. It's other people's ideas that you have to do for them and and their problems, their specific pro Yeah. Yeah. Cool. Well, let's let's should we should we crack on? Yeah, let's crack on. Uh let us know in the comments what you're building. I want to know what other people are building. I want to do more. I want to do more. I think I'll I'll get through my little phase of like Mac and local AI because that's the beauty of YouTube stuff. It's like you can go through es and flows of like what you like to do. Definitely ride the waves of stuff that's obviously doing well, but I look forward to a a point in my journey where I'm showing what I'm building, you know, and building content around that because I think it's good. I think people need to hear what each other are building. to give them ideas, you know what I mean? They'll take it and they'll make it better or they'll make their slightly adjusted version of for them sort of thing because they they didn't know something was available. So, um yeah, definitely share if if you want to share what what you're doing because it's quite cool. Yeah, I like hearing it. [snorts] Should [clears throat] we um smack that intro? No intro. Okay. No intro. [laughter] Okay. So, this week, actually, let me check a bunch of other stuff. Yeah. Just Yeah, this week we have a state-ofthe-art image generation model from Chad GPT. It's just so freaking good. Uh, I cannot wait to show you. SpaceX is buying cursor. What like what world are we are we living in? We have flip book. It's it's a crazy idea that I want to share with you guys. Uh, we have Kimmy K2. Uh, Kimmy K 2.6. Sam, you will be talking about that. And we have Deepseek 4 and something about anthropic testing. Uh, Antropic trying to get Claude code. Uh, what what is he? What are they? Oh, they are trying to remove cloud code from the pro plan. So, I'm stuck with the hundred euro plan. Wow. Yeah, you're wrong. I'm on the 101. Yeah. Um, okay, cool. So, we have all of this and more in today's episode. Uh, this is Command AI. I'm Kavarza. And I'm Sam. And this is a week in AI, artificial intelligence, design, and dev. I still think we need a a theme tune there. Something like that. Um, well, you missed out the Versel supply chain. Yeah, I Yeah, I don't know what that was. So, well, we I mean, let's just talk about it. So basically uh Versel uh announced on all their channels that they've had an attack basically through uh and people's environment variables got leaked like the secret keys. I think there's a bit of I have my doubts I have my doubts about this and we'll get into that but yeah they said if you have un like you can there's this idea of sensitive and unsensitive environment variables in in versel obviously sensitive ones you can't see like in the in the UI you can't see sensitive ones however you can see and edit um uncensitive ones someone managed to get in by overtaking someone's uh Google account who worked at Versel using a tool called context AI which let me show my screen. I took a look at what context AI is. Um [snorts] it's an agent orchestration platform. So a supply. So basically everyone at I think context AI is slightly less known than versel. So context AI are also [ __ ] If you use that you got to be careful about this. Um, but yeah, someone gained access to someone at Versel using context AI and got into their Google account and then got into the Versel account, attached to their Google account and then basically leaked a bunch of uh, you know, environment variables. If you if you've got if you're using it correctly, then you'll have a bunch of secrets that are protected under the secret thing. But anything that's unprotected is you've got to rotate your keys basically. The reason why I think this is a bit bit [snorts] suspicious you know a bit but is for develop so you've got your production branch or you got your production environment on Versel if you um you've also got preview and development URLs as well for seeing pull requests and they build and you can review them and whatever before they go into production. It's a pro feature to enable secrets to be sensitive on those production deployments. So it feels it just feels annoying because it's like well okay so everyone who's not on a pro plan sort of it's sort of it's sort of like leading you on to a pro I'm not I'm not fully putting on the tin for hat. I I genuinely think this happened, but a small selection of people who were affected by this have been notified who they know, but they do recommend rotating any keys that weren't marked as sensitive inside of their cell. But it just so it just it's just a um it's kind of uh what is it? What do they call it? It's kind of coincidental that a pro feature is marking environment variables as sensitive on development and preview branches. So, I don't know. I saw it as an excuse to get in contact with some clients and say to them there's been an attack you're we built your website on Versel. Do you want us to take a look into it? So you might want to reach out if you if you've built any products on Versel uh for your clients. Definitely reach out and say hey this has happened. Can I sort should I sort it out for you? And I'm doing it in a way where I'll do the review for free. So I'll look I'll look to see if they're affected. If everything's set up, then I'll leave them to to, you know, carry on as normal. But if they've got any keys marked as unsensitive that should be sensitive, then I'll I'll go in and rotate those keys and that's what I'll charge for. So yeah, opportunity for you there to earn a bit of money, but um yeah, it just seems a bit suspicious to me, but there you go. You make money because somebody used white coding. Yeah. Yeah. Well, who knows? Who knows what this has actually caused this but someone um definitely was uh well final thing I'll say on it is that this person apparently knew uh Versel's architecture and APIs uh very intimately. though suggests it might have not been an inside job but maybe an ex employee or someone like that like it was very it was very um uh what do you call it sophisticated highly sophisticated attack with detailed knowledge of the sales internal system. So [snorts] yeah, rotate your keys, get in contact with your clients. Stay safe out there, boys and girls. Cool. I I I missed that news, by the way. I Yeah, it's it's it's been such a weird week, guys, because like I'm I'm normally on it. I I felt like I was totally out of it and then but loads of stuff kept coming up during the week and um I've been playing around with our AI claw bot producer and I it was a point where um basically if my laptop's turned off it doesn't work or closed it doesn't work. So I wasn't getting a bunch of news. So I'm like I have to remember that. I have to remember that. So there was so much stuff that yeah may or may not make it into this week. If you hear about anything that we don't speak about, let us know and we'll talk about it because last week design claw design dropped midstream and we're in a frantic rush to get it out and cover it. We covered it twice to make sure the video was good enough. So, uh yeah, let us know if you hear about anything while we're while we're live and we'll cover it. We'll go through it together. But [snorts] yeah, this was very early on. This was like, you know, Monday, I think. Do you want to talk about GPT images? Yeah. Did you do a test of GPT 5.5? With images? No. Or with a design? Design? Uh, no. I don't have a design yet. Maybe you can run that. I can't. That's why. Yeah, I can't. Ah, you don't have it. No. Uh, yeah. because they haven't released it on the API. So no one no other coding tool has got it yet. That that's why they're keeping me on my toe. I still have my 20 bucks subscription. I cannot cancel it. Just like every week there's a new reason to to keep it. That that's that's how they are making money. I guess I did have fun, you know, jumping between the two to be honest. Claude and GBT. I do this all the time. I I noticed that like many people don't do, but I I mix and match everything. But I don't do it I don't do it to get better quality. I do it literally because it's just a nice change. And um if I run out of credits on one, I'll just use the other. It's not like, oh, Claude would be good at this or Jackbt would be good at that. I'll just be No, I I don't believe in that. By the way, I don't think one of them is particularly like miles ahead of the other one. Guys, I've been I've been using Kilo code a lot recently and especially today even I'm using their own model, Kilo's Kilo's um Frontier model, you know, fraction of a cost. And we'll we'll go through um GLM 5.1, Deep Seek 4, and whatever the other one is, Kimmy K2 uh 2.6 six. Like we look at these benchmarks, we're splitting hairs. Do you know what I mean? Like these models are all incredible. Spend a little bit you might spend a little bit longer on some of the cheaper ones, but you know, it's it becomes arguable whether we are just, you know, we're so spoiled. Jovin's paradox, isn't it? Yeah. Okay. So, let's All right. [clears throat] So, we have a new image model and it's so freaking good. I did a bunch of tests, compared it to what Google has, Nano Banana 2 and NanoBanana Pro. I will show you all in a second. But first, what is this? We have ChatgPT images 2.0. And it's incredible. It's incredible in doing anything with UI, with graphic design, and especially with text. All of this text is made, all of these UI, everything here is made with AI. It is incredible, and it's like unbelievable how good it is. Look at this. You have images within images with multiple languages. Um, knowing Farsy, by the way, this doesn't make complete sense, but it's it's really really Maybe they didn't say what it should say, so it just said something. Anyway, uh, it is really really really good. I did even um a test of something in Farsy. So, I will show that. It's incredibly realistic. It's great at UI design. It's incredible at graphic design. The text rendering is sharp. It's It's so good. So, where do we begin? I did a bunch of tests through the chat GPT app and through Hicksfields and here are the results. So as I was showing earlier in the show, I was building this I was designing and building this app and then chat GPT image drop the image V2 [snorts] and I was like hm let me use it for UI design and I did I just gave it a screenshot of like my very wire fra frame looking like very not nicely designed. It was a wireframe right? I gave it uh to CHP and it gave me this UI. This is pretty good. I mean this might not be perfect in terms of like if you look at it like a UI designer and if you look at it like as a front-end designer but it's amazing for one thing and that is getting some ideas for how the design can look like. So I I did this again and again and again. I did a bunch of designs and I was getting some interesting ideas by doing these like different designs and I even started putting some stuff literally on top of each other and be like I want a piece of that and I want a piece of that and it's really good at combining these things that I give it to create new UI. I mean look at this. After I got these UIs, I was like, maybe give me something for the light temperature and light intensity controls. And it gave me a bunch. And all of these are really nice looking. And I took two that were simpler. I didn't want to go with something too, you know, out there. Um, like some of these are a bit too much. Uh, and I gave it to Claude and it made it for me. Again here [snorts] I was like I was creating this sphere design and in one of these example it came up with this idea where we have kind of like the base the table and I was like why didn't I think of this? This is much better. Who you don't want the camera to go under the table doesn't make sense. So what this is good at it's generating these ideas and then following these UI design ideas I came to I I I I made this one. Uh yeah I was really impressed with this and I just continued. So that's one side of things creating variations of UI design. Uh and I'm really really happy with that. Yeah, because by the way as well this is a this [snorts] is maybe one of the first thinking image models, right? So this is able to help you iterate by thinking. Yes. Uh by the way for all of these my prompts have been bad and minimal and just my prompt I I haven't done any like prompt engineering. All I've done has been giving it some context for what it is and offloading the prompting to the chat GPT app. So you turn on the uh the thinking in the chat GPT app and it does thinking first and then does the image generation after. So it seems like it tries to understand what you actually want to do and then create a prompt and then hand it over to the image generation model which you know creates these UI designs and I'm really really happy with these designs. So, uh, there were a bunch of stuff I saw people were doing on Twitter, for example, creating some sort of I don't know what what do you call like recipe. Yeah. designs. This is uh a Persian um dish and it's it's my favorite and this is really really good. it. Just to be clear, there are mistakes, but for a quick prompt, no iteration, just like one prompt and very broad, this is really good. Uh, it understands the language, it understands the concepts, it, by the way, it does searching like you can tell it, you can tell Chachi to go and search how to make a certain dish, for example. It searches it. It finds images and then use all of that to create the final image. So yeah, this is this is quite impressive. Um amazing. It is like just amazing. Maybe it is taking it from literally a web page. I don't know. We we know we we might never know how it does it where it's pulling the information if it's like from Wikipedia or from a a cooking website or it generates it. I don't know. All I know is I gave it a generic prompt and it made this and it's really good. Uh another one I did was web flow redesign. Um the do do a rebrand in red. This is not the cleanest. I see all the mistakes. But to all UI designers that want to go into the comments and be like, "Well, this doesn't make sense and this is not perfect." I see all of that. That's not the point. The point is you can create directions just the way I showed it with the UI design. You can create ideas uh mood boards if you wish and then the the human with taste will get into that can can choose the can choose the good ones and do the cleanup. Uh I did some food photography for this thing that I was doing. Well cleanup um and it's incredible at that. And then here comes the comparison. So any anything you uh you want to add or question before we go into com comparison? Anything I want to add? Uh no no all good. All good. [clears throat] Okay. So here let's compare uh the image v2 versus nano banana 2 and nano banana pro. I used the same exact prompt and I prompted actually chat GPT through two different places. First from chat GPT app just like the normal app that you have and also chat GPT images API through Higsfield because I want to see if that makes a difference and spoiler alert that was one of that was one of the questions. So Higfield for for Nano Banana, but in the UI for GBT5 because it's not available for all. Yeah, all three of these are with Hicksfields by the way. And then chatgbt is through the UI chat GBT interface. Yes. So yeah, sorry. Made in Hicksfield. All three of these made in chat GPT. Yeah, these. Yeah, same prompt. I just made a few more with CHP, but same thing. So, the prompts were to create some street wear, some merchandise for our brand. And I included I included some of our brand assets and then a list of words made with chat GPT 5.5. Well, that that's for a separate video, but yeah, with that model. So, here's the results. The results are [snorts] incredible. I'm just I want these. I want to wear these. And like genuinely, these are just so freaking good. Look at this. Look at this. These are just so high quality, so freaking realistic. Don't get attached to models. This is just so good. This week's best model. The cursor is thinking probably my favorite. Outdated by tomorrow, you know, all AI jokes. Um, taste is the last interface context window. These are really good. And look at these few last ones. Like these I I did with this holographic. I had this web flow sticker and I was like, let's do it like this. So, I did it holographic. uh with holographic design and ah human taste, machine speed. Come on, we got to make this. This is the second week in a row we're mentioning merch, guys. Like, let's make it happen. If you want to buy it, we'll do it. I mean, human taste machine speed. Come on. This this is really cool to wear. I I love these. Um so, yeah, these are the images with Chad GPT. [clears throat] Yeah, keep on wiping what you say. Uh and then I made the a few from these are also done with CHPTV2 but through Higsfields and honestly I don't see a difference necessarily but I did have a sense I had a feeling that it's better with chat GPT directly especially cohesive especially that I could chat at it's such a nice experience that you can like chat with the model a bit more and it's a conversational thing in one and done. Yes. But what's nice with hex fields, you can just pop a a 100 uh with a few clicks. You can't do this with chat GPT. It's quite slow. You can do one at a time or you have to to have multiple chats. But yeah, uh these are pretty good. Now, let's compare it to Nano Banana 2. It's not the same. There are creative stuff that I like about Nano Banana models creative wise, like this shirt or like the idea of maybe this style, but the the details are just not there. designers with terminal access. Um the realist it's just so much more realistic the ones uh from chat pt. These are well I would say to to me the they look like more like photos. They look more like photos. However, it's the graphics on the clothes that have this very 2D look and they're not they're not they don't look fluid with the clothing. It looks like a bad Photoshop job. Whereas, as I say, I think the photo itself, the aesthetic looks more like a photo. That to me looks more like a photo. Uh, if we compare these two, for example, these two are very like close in style. The bottom one is from Chat GPT. The top one is from Nano Banana 2. [snorts] This has more details like the hair, the environment detail. Overall, it's quite dark though and like it has an aesthetic and you can definitely you can do that like that looks like it's taken on a uh disposable camera as the one above looks more like it again maybe a disposable camera but the flash is actually on. [laughter] Yeah, the prompts were the same by the way, not for every image, but I use this the same prompt in every model, but yeah. Um to me these are a bit more and this is Nano Banana Pro and it's just a level of details and um the the ones from Chat GPT do look more realistic the physics more ma make more sense [snorts] the math is mathing uh in Google it not so much but it's like this is still really good. Just to be clear, this is still really really good. I love like there are some details here, but the the chat GPT1 are just higher quality, higher detail. It feels like uh the the setup setup of people in a room makes more sense. uh from the ones from chat GPT, it feels like that they actually exist in that space where as for Google, I feel like it it it feels like more photoshopped in. It feels like things were just put in there a little bit less realistically a little bit and it's not a huge one, but yeah. U this one is really nice. I have to say I maybe disagree with you a little bit there, but yeah, the merch for me looks so much better from GPT uh Imagen 2 and the way that that the the graphics are placed on the clothes is a lot better in GPT uh Imagine 2. Uh but and maybe it's a taste thing. Does the the images look more well only marginally more cohesive, but the the the to me these look photoshopped next to each other. They don't they all don't look like they're standing next to each other, whereas to me the photos from Nana Banana 2 look like they're all next to each other. That to me looks a bit flat, you know? Yeah. Now, with that being said, I I I I think you're right. uh you and it it's marginal at least it's very marginal the difference uh there is an important tell though with chat GPT images there they have a specific look that nobody else is talking about and once you see it you cannot unsee it. Have you used the stamp uh the stamp tool in Photosh? Yeah. It has way too much of that and once you begin to see it, it's everywhere. Especially in places where it's out of focus. Here um you see there is a noise a big noise pattern. It's a not it's not a tiny noise. It's like the Photoshop clone stamp uh pattern that is everywhere. You have it there. Everywhere there is detail. You have it in the hair. I know the hair has has the style, but it still has this oversharpening in that specific style, in that cloning style. And and here, by the way, you see it very minimally. Like these are like the good images. I've had other images, not like from these, where you see that just a bit too much. Like here on the food, I see a repetition of the pattern that is very photoshoplike. And once you see this, maybe it is difficult to come through the uh the stream because this is a very detailedly thing, but once you see it, uh it's there and it's Yeah, I guess that's just their distinctive look for now. Um, I mean, you see it here. Here it it repeats itself uh quite a lot and it's it's it's not a it's not breaking the model or anything, but I do certainly notice it. Yeah, it's great though nonetheless. The the model is amazing. So, yeah. and what it means for graphic designers. Uh actually let before talking about this let me show a few uh one more thing actually another thing that people uh started to share are 360 images. So you can because the resolution is so high you can create 360 images and then build a 360 viewer and it's just like really interesting creating these experiences and now with that you could imagine how you could create procedural environments where you could create worlds where you can actually walk in. Yeah. With 360 images. So that's another thing happening. So realistic images, mockup images, and especially graphics with, as we saw here, graphics with text on it. So RIP graphic designers. What do you think about that? [clears throat] Yeah. No, it's uh it's it's an exciting ideation tool, I think. And um you know it and it's like when you say RIP graphic designers like of course there's going to be need for very very bespoke graphic design but for a lot of people like small business owners they're going to just go straight to this and get be very very happy with images and not having to pay anything for them which by the way do you have an idea of price of what it takes to generate these images? [snorts] No idea actually. I I should know because I'm building a tool around this but I haven't checked yet. Let me actually check so I can because I guess it's I guess it's covered under the JGBT because they haven't released it on the API. So we don't have No, they do have the API because I use it through I used it through Hicksfield. So that means through the I thought you said they used through Chibi BT. That's why I was trying to clarify right before you went to the nana banana and the thing I I mentioned it. Look, so everything here is labeled. These three these three rows are made through Higsfield. Yeah, these are chat GPT images 2.0 made in Hicksfield. And these are the same prompts made through the app. So yeah. Okay. So what's the price? I don't know. I'm checking that, but uh I don't know the answer yet. Uh I don't know if they have it. By the bottom, it said pricing and availability. You scroll past it. There we go. Yeah. [snorts] Chat GPT images 2.0 is available starting today to all GPT and CEX user. Advanced outputs with thinking are available to chat GPT plus pro and business user. The image two model is available and the API pricing varies depending on the selected quality and resolution of the images. So let's check that API pricing. So these are for the coding models uh for the thank you ah images too. So, $8 per 1 million token for inputs, $21 million tokens for cached inputs and $30 one million token for output. But how does it does that? I don't know how tokens work with images. So, yeah. Yeah, roughly how many images does that get you? Don't know. Let's see if they have it somewhere in the FAQ. So what is this? How is pricing calculated for images are converted into tokens and charge per token? text model price image tokens as standard text tokens rates while GPT image and GPT realtime uses a separate image token rate models like 4.1 mini nano convert images into token so pricing calculator uh the width who wants that I want like at least,500 by,500 pixels Okay, that's is that the price? So, what if I go with 4K? I mean, I know 4K is not exactly that, but let's say something like this. So, this is the pricing. It's not too bad if I'm not mistaken. Oh, [gasps] but which one? I don't know because we have a bunch of five versions. You see that's a bit confusing. So, yeah. 5 point. Yeah. But, uh, imagen two is the model. Yeah, but it's not here because it converts it to text. So, if I choose the the most expensive one or like the the newest one, 4,000 pixel by 2,700 pixels is this much. It's So, it doesn't doesn't tell you how much image in image two is or real time. I don't see that now. Weird. Okay. Never mind. I will use it and we'll see because I have like [clears throat] $6 left on in my tokens. Yeah. Yeah. Anyway, yeah. Okay. Yeah. Nice. Cool. Nice. We can we can move on. Um cool. Well, let us talk about SpaceX buying. That's crazy. Which Yeah, SpaceX is set uh secured an option to either acquire Cursor for $60 billion later this year or pay $10 billion for their new partnership. So either way, Cursor's quid in it. [laughter] [gasps] And by the way, did you know the owner I think is 25? The tall ginger guy. So 25. No. Um, but reading this article has revealed quite a lot of uh tom foolery going on as Elon Musk is known to do. Um, and a crazy [snorts] amount of compute that I really didn't realize XAI/SPACEX had. Um it also provides cursor with more computing capacity to develop AI tools. So they're they're actually providing the compute capacity for yeah um composer I think they're called the the AI model and everything really. The combination of cursor's leading product and distribution to expert software engineers with Space X's a million a million H100 equivalent Colossus training supercomputer will allow us to build the world's most useful models. Colossus is XAI supercomputer cluster in Memphis which has touted as the largest in the world. I did not have I did not know this was all going on behind the scenes. which is kind of crazy. That's crazy. And as I say, the deal is that they got to buy it for 60 billion or pay that 10 billion, you know, whether they go through it or not, you know, which is just mental. Um, got some monos here. So, I don't really know what like why would you why would you go through with this with this deal? Like why would you I don't get what Elon Musk is getting out of this. Do you understand? If I can guess, well, maybe because they they see that um Claude was right like Anthropic was right with what can make money. People are not paying Chat GPT just to have a chat app to talk to because that part is going to be free. We know that uh what people will probably pay for is um all of these tools around these LLMs to make uh to generate code with essentially to to create products with. Probably that's going to probably that's going to make uh the most money or that's what they see. So that's my guess. They they they see that the future of these like LLM businesses are bound to these tools that they can create and not just a chat app experience. And cursor is a in a in a really good position with a lot of a lot of users having a good model and all the data from all of other models are going through them. uh and they actually they have a better position than many others because they see what code is actually being pushed to production. So that's that's also I think is a big u important factor there. Um, so it's probably that. Yeah. Well, like I say, they've got the right to buy Cursor at the end of the year. It doesn't mean to say they will. And again, like what if Cursor pull out that means SpaceX have got to pay him 10 billion like like it just seems win-win for Curser. Um, but also yeah, read enough about this this whole thing. It's this whole manipulation of like a XAI are now owned by SpaceX. So they bought and then XAI own Twitter. Is that right? Yeah. And Grock. So now SpaceX Yeah. Grock basically is is the model XAI. Yeah. Is Yeah. Um so now SpaceX just basically own Twitter. That's what I get from that. [laughter] Yeah, that's crazy. Yeah, that's the the space bird. What did you say is like from one pocket to the other sort of thing. It's seems to be doing Yeah, it makes sense if you squint a little bit because of course SpaceX want to put compute into space. Yeah, I I I don't reckon they will. I don't reckon they'll put it into space. I don't know. I think they will try it. I don't know. I reckon that was the that was the story to stakeholders to, you know, allow them to buy Twitter and make it justifiable that Space X buys XAI and Twitter sort of thing. So, how will they be profitable? Because well, we know that Starlink is growing like crazy and there are talks about and they have done testings for bringing like actual satellite connectivity to your cell phone. That's insane. um that there will be a lot of money to make there. I don't know if they will send like actual data centers to space because it's that's what I mean. We talked about it. It's heavy stuff. Yeah. You I I hope they will because it's cool. [laughter] It's not my money. So, I guess I want to see it happen. Uh [snorts] so, I hope I hope that they can do this and we'll see. We'll see. It's crazy times. Well, yeah, more compute. We know that the thing all that to be said, you know, like uh I'm not really using cursor. I don't think I think cursor is just becoming bloatware. So yeah, yes, they'll be able to buy it, but like I don't I just Yeah, I I sort of think that people are using cursor less and less because these AI agentic code tools are pretty like codeex uh claw code desktop uh T3 code. This is the way that people are going now, you know. But I don't know, maybe we're in a bit of a bubble, but yeah, let's call them agent managers. these agent managers are the way that it's going and IDE actually writing the code is is kind of in the past now you know yeah I had two cursor subscriptions and I cancelled both that's the length I went to um but they came with the new cursor design and I really like that design it's the same as you know every other agentic tool but it it feels like every week I'm developing my taste a little bit because we we are still figuring out what is the better way of doing something. Like for example, for web UI, there was a time where we knew that Web Flow figured it out. Web Flow was the first company that actually figured out how to connect CSS to a UI properly and it was it was just so good. And everyone else just, you know, did what they did with maybe a little bit of twist. We are at this point with AI where we don't know how agentic task management should be. Uh and this week I feel like dispatch is not a bad idea. Mhm. [clears throat] So I don't have to because completely removed the the not just that uh okay I've been using dispatch even from my computer which makes you why from my computer like you can go to the dispatch and just tell it to do stuff. You don't have to use it from your phone. You can also continue the conversation from the app from your computer. what what I like about it. So, okay, here's the thing with with codeex app, with the new cursor, with T3, with uh claude code desktop app, with all of these, you have a list of conversations to maintain and you have like separate conversations and separate contextes and separate tasks, right? And you manage these different ones. With dispatch, you remove yourself even from that. You tell an agent to do all of that for you. So there is a agent sta staying on the top of all of that that decides when to create a new cloth code instance and when not to and it has a memory of its own. And I've been enjoying that quite a lot. The only thing is it's missing. I wish I still could kind of like switch between like having a master conversation where has all of that but also being able to go into individual not instance of cloud code but topics where multiple instances are stored potentially. So it's like having a team like having a proper team. I either talk to the manager who runs them or I could go walk to individual developers who has like a lot of context uh of multiple tasks that so an open claw in telegram in a way was that because in telegram you can have open claw and telegram topics in a group which are separated. I don't know if you know Telegram topics. It's just like tags within a group and it separates conversations so you can have a general one including everything and also specific ones. Um, so that's my vision for this week. [laughter] Great. Thanks. Maybe we'll I don't know. I I'll we'll see. I think we are going that direction because why should you manage your different conversations when an agent can do that as well and then so all that to be said what is Elon Musk buying you know like what is he actually buying do you know yeah cursor has a good model do they yeah yeah I I didn't like Oh you mean you mean the AI model I mean business model well that's that that as well the business model of creating software like enabling more people to create software that's what Claude has found to be uh lucrative at least more compared to you know compared to GBT trying to [snorts] create Sora or whatever. Yeah. Well, we'll get into GBT 5.5. I think that's there's definitely a reframing that's happening with uh with the new model and and G chat open AI's outlook on you know yeah helping people get work done. I think that's that's the play. Uh let's move on to do you want to do flip book? Yes. Uh do you think we can where's GPT 5.5? Where's that gone from our notes? I'm looking our internal notes. So, it seems to have disappeared. Interesting. Yeah, our AI is like taking control. And um Wow. Yeah. If you want to wrap, let's do How long have you got? Uh 30 minutes. 30 minutes. Let's do GT. Do you want to do 5.5? Should we just go over the release of 5.5? Uh if you can do it would be better. Uh, and I will do the Okay, let me find that because I don't have it prepared. I cannot get over our merch. I'm I'm switching the tabs and I'm like I don't care about you guys. I I want to print one at least one of these and human taste machine speed. That that that one I I I want it. So GBT 5.5 literally just landed yesterday and I think it's starting to solidify a new I don't know we'll get into it. A new way to work I guess is the way to think about it. So, first impressions, it we're releasing GPT 5.5, our smartest and most intuitive to use model yet, and the next step forward, a new way of getting work done on a computer, and what I said about the uh the the focus on the workforce, GPT 5.5 understands what you're trying to do faster and carry more of the work itself. It excels at writing and debugging code, researching online, analyzing data, creating documents and spreadsheets, operating software, and moving across tools until the task is finished. Instead of carefully managing every step, you can give GBT 5.5 a messy multiart task and trust it to plan tools, check its work, and navigate through ambiguity and keep going. To me, that sounds like a huge shift in the way that they're talking about this model that it is something that you should use to get your work done to actually do things with. It's not about being smart. It's not about being friendly. It's not about being a chat model. I don't know. Do you agree? This feels like it it the tonality and the way they're presenting this model is way more about real work. is agent stuff because no one's using like well I say no one that's an exaggeration but people aren't using AI to like chat with anymore you know what I mean like the casual user the one who who's zero value to a company dip in to chat GBT and chat with it every so often they're not you're not chatting you want it to do stuff for you and so that's what you want to hear about and that's what the where the focus is something Claude has obviously been doing for a while. JGBT seem to be getting it. The gains are Yeah, Google. Okay. Uh the gains are especially strong in agentic coding and computer use, knowledge work, and early scientific research. So, just to cut in as well, GBG 5.5 also has a pro tinier, and they're just talking about GBT 5.5 at the moment. um delivers seup intelligence without compromising on speed. Larger, more capable models are often slower to serve, but GPT5 matches GP 5.4 per token latency and real world serving while performing at much higher level of intelligence. Um stronger safeguards um and here we go. G5.5 is rolling out to plus pro and business and enterprise users in chat GBT and codeex. You've got this right. and GBT 5.5 Pro is rolling out to pro business and enterprise users in chat GBT, so not in codeex just yet. Um, a API deployments require different safeguards and we're working closely with partners and customers on the safety and security requirements for it serving at scale. We'll bring GBG 5.5 and 5.5 Pro to the API very soon. So, this just shows you the importance of a harness, how much they're able to control the raw power of a model because API usage, you just get the you just get the model, like do what you want with it. Um, you get the raw you get the raw output based on what you tell it. So, yeah. Yeah, they're obviously able to control it a little bit there. Uh, smashing all the benchmarks. We won't go through um hugely on this, but Terminal Bench absolutely killing Opus there. I did I said when we went over Opus 4.7 that it's not great on terminal bench at all. 82.7. Um what we what do we care about? What do we care about? Browser comp uh well pro is 90% but it's still 84.4. Um Opus 4 uh 4.7 79. Just yeah very very very good model. The thing I wanted to show about uh show you is uh so it's this week's best model. It's this this it definitely is this week's best model. Definitely this week's best model. Um I can summarize it if I don't see it. So bear with me just a sec. Well, these are interesting as well. Look at these apps that it's building. Like these cool I can show mine. Yeah, the 3D stuff. Yeah. Uh, earthquake tracker. Bit of UI there. Really nice. That's really cool. Dungeon game. That's cool. 3D game there. [snorts] Um, yeah. I won't I won't I I won't be able to find it, but it's somewhere in this and it was maybe it was in the live stream, but GBT 5.5 uh use very similar terminology to um Opus 4.7, which is it does more of what you ask it. It does exactly what you tell it to do. And with this comes with this idea that when like for better or worse an AI will go off and do a bunch of stuff. Sometimes it will fill in the blanks. Sometimes it will go in excess and do things that you don't want it to do. and and a lot and and they're warning you or they're almost advising you to change your behavior and the way that you prompt these things and and um get to know the model and how it works and the sort of stuff it likes to leave out because it's more direct. It will do what you tell it to do. So, it's interesting now that um that's why I said about solidifying Opus 4.7 literally gave you like a little bullet point to change how you work. GBT 5.5 is also saying the same sort of thing that it's it's better at doing what you tell it to do. So definitely worth playing around with and definitely worth being mindful that uh maybe it's becoming more and more important to be you know prescriptive about what it is rather than lazy. You know maybe those maybe those lazy days of um telling the model to like you know just do everything. Oh I I think it's my camera. Go on. Okay. Um, yeah, days of just like figuring out Well, he's coming in and out now. Those days of figuring out like or letting the AI figure out a whole new feature is, you know, or at least I think going into planning mode, um, asking it questions and having having it sort of fill out everything for you in the planning stages is kind of where you want to sit with it. [snorts] So, definitely uh definitely exciting new model. Don't know where he's Is he back? Are you back? No, he's not back. I wonder if he can hear us. Um I'm just scrolling through here whilst we wait for Kabaza. I can hear you. Can you hear me? Oh yeah. No, I couldn't hear you at all. I can hear you now. Yeah. Yeah. I I don't know why it's it's bugging. Uh maybe I close a few apps. I [clears throat] need a new MacBook, I guess, or I need to restart it. Yeah. Uh, this is quite a nice thing, the GB GDP Val, which is one I'm becoming more and more interested in. I really I said last week like I want to set up a trading bot. Uh, and it's kicking ass, better than Opus 4.7 on GDP val, but it's definitely welcome. Um, yeah. I think that's that's kind of all in all the the the thing. Did you have something you wanted to show about it? Yes. But I struggle with getting anything open right now. Yeah. Yeah. You look like you're frozen to me. Yeah, I I do see myself as frozen as well. Um, let's see. Should I refresh? I will refresh this tab, maybe. Oh, yep. You're coming back. You're in and out. Oh, also something to note as well. They're working with Nvidia uh to basically get get you better compute. So, they're I think they've deployed it on a bunch of new hardware, probably reserving some of the older hardware to um some of the older models and things like that. But I did f I did I didn't mind GBT 5.4 to be fair. So it's a shame because I I would like to I would like to use this um Google Chrome renderer issue. Um it's Google Chrome if I close. Okay. So whilst you do that then also it's worth saying that the pricing has kind of gone up. Um GT5.5 will soon be available response and chat completions at $5 in $30 out now. And then pro $30 in $180 out. We are getting Opus prices there. Let us let let us look on um uh open router for prices 25015. So it's double the price. It's double the price of GBT 5.4. for. So it ain't cheap guys. In shitification continues for sure. Well, we know we've known for some time that these models are very expensive and they are basically giving away tokens, but they can't do this forever. Yeah. No, there we go. I think this is it now. I think this is it. The money's running out. Okay. One last try. If I can't show it, I won't. Uh, it was a quick comparison. Maybe I can now. Okay. So, you see Mhm. So, I gave chat GP. So, Okay. So, I gave GPT 5.5 and 4.7 from um Claude the same exact prompt to create this 3D scene with these lights with lights that I can add. Uh, and one of these are made with Chetch GPT. This and author, you're gonna make us guess. Yes. And this one. Which one do you think is with CHP? Which one do you think is with Claude? I think the Oh, maybe this one actually cuz I'm looking for the cards. I'm just looking for the cards. Uh, flip back to the other one. I don't think you can guess with the cards because I gave it a design. I think this one's Claude. So I gave it a design from um from a Figma screenshot which was an image generated was uh generated with image gen 2 to to to develop it. Uh no 13 is chat GPT 12 is claud. Okay [snorts] to be fair I did tweak them so this is not the first uh thing that they came up with the first instance. Yeah, I one of them I Claude got a few things regarding like the physics better but GPT got a few things regarding the looks better but I cannot draw a conclusion from that and tell you well you know one of them is better at this and the other not I I I think with AI this is too much of you know uncertainty to say it's like that I I think and I said this in a community meetup we had anything beyond GPT codeex 5.3 is good like claw 4.5 even anything beyond that they are all good you you can do like actual software with them as long as you know you are limited with the amount that you know and there will be always better models for sure, but they will you know what you can how you can scale through the roof when you learn and this is something that I need to work on. Um that's a different topic but I don't feel like I'm necessarily learning as much as I'm you know building. Uh that does sting a bit. I would like to learn a bit more. Yeah, I've been thinking about that a lot recently like you know it becomes more important to know how all of this fits together. It's all one thing just to rely on the AI just to do everything. It's another thing to actually like I I I often find myself just like holy [ __ ] if I didn't know this like oh the Docker container isn't running or or um you know looking where to look and and on logs and like you know something's not going wrong I know that oh this seems like a Versel thing so I'm going to go and Versel logs or this seems like a clerk issue so I'm going to go on clerk logs or browser logs or whatever it is just knowing how all of this stuff fits together. Yes, you can you can learn just by banging your head against the wall and just going through it and then having that aha moment of like oh god I I'll like you know you open up the console as an example and then you find that um oh I won't make that mistake in future but like it takes hours and hours and days even of banging your head up against the wall to get to that point whereas when you know well to be fair I bang my head up against the wall you know with with other things as well however when you know, you just know, you just know. Um, but saying all that, you know, you mentioned about the models and this and that. Kim K 2.6 has just launched yesterday even. Look at these benchmarks, dude. Like, and I know they haven't got 4.7 and I know they haven't got 5.5, but look how it competes against what you've just said are brilliant models. And this is a cheap uh very very cheap Chinese model on par with pretty much all of them, you know, and these are the good models. And I've been playing around, as I say, I've been playing around with um uh Kilo Code and just flipping between models and stuff like that. Uh DeepSeek version 4 has just landed, so let's take a look at that. [snorts] It's crazy. This is like the first time since I got into cloth code that I'm actually thinking about cancelling cloth subscription and go with one of these models. Kind of me too. I'll be honest. You know, you know me. I'm I'm I'm like a a Claude fanboy, but I'm now just like, man, like I'm I've just done so much work and it's taken off a couple of dollars off my kilo code thing. Yeah. So, um, and I'm starting to understand all this stuff now. The the the fact that this is Oh, this is V4 Pro, but there is V4 Pro and V4 Flash and V Well, yeah, V4 Pro and V4 uh V4 flash 1.6 uh sorry, 49 No, no, no. Yeah, sorry. 1.6 6 trillion parameters and it's a mixture of expert model and each um it's a 49 billion active parameters. Now just to bring this into context I've been playing around a lot with local AIS and I'm pissing around with like oh should I use the 4 billion parameter model or the 9 billion parameter model because it can't fit on my MacBook Pro 9 billions. This is and and these are sorry and these are nine billion parameter models. This is 1.6 trillion parameters. It is a hefty model. We are talking where are we where are we talking here? If we go to um files you know wouldn't be cool it would be cool to see these are all the these are all the tensor files. These are all make up the model basically you know. Um, I want to see a total. I want to see a total. I don't know. But saying that as well, you can download it. Yeah, that that's crazy. That's actually crazy because MLX community have already quantized it to MLX at 8 bit and that is 302 gig at 8 bit, you know. Oh, this and this is the flash one as well. This is the smaller one. Yeah. So, we you you can play around with it. And the interesting thing about this is that they released it, open source it the same day that they released it. This actually appeared on hugging face before they actually announced it. And it's often a bit of a two-step program. Um, and saying that, GLM 5.1 got released as well. [laughter and gasps] And these all have very reasonable coding plans. Like uh if I go to um great share this. This is the Kimmy coding plan. Starts at $15 per month. Uh well actually monthly starts at 19, but if you were to go annually um you know these are great models. Then uh GLM coding plan here. I think it's I actually haven't people have been telling me this has gone up because I made a I made a video when it was $10. Okay. So yeah, it has monthly 18. So it's gone up to $18. But you can get a yearly Oh, Jesus Christ. That's a that's a lot. But you know these sorts of prices for these models are crazy and they are all very good. They are all very very good. I want to see benchmarks on GM 5.126. Um yeah again GPT 5.4 4 is like a ahead of GBG 5.4 ahead of Opus 4.6. This is software engineering benchmarks. Gentic coding not quite as good but it's just these are all cheap models that are all really really good and these got released this week. So take your pick. [laughter] Yeah, take your pick. As you see, I'm like moving into like developing some software that is not, you know, um familiar to our community, you know, the web and design community and I'm just trying to see if I what I can do because there's just so much software to de to be developed. There's so much tools there are so much tools and ideas that you can create. So all of these are uh axes that you can choose and use. It's so it's yeah I felt like a kill in the candy shop today just being like holy [ __ ] I can just build anything and the tools that I'm building for myself is so fun. And then all that to be said, you know, we don't dunk on anthropic too much on this channel, but all that to be said, there was a bit of a slip up and they removed Claude code from the pro plan, which is what I have. And whether this was an accident or whether this was some AB testing gone wrong, people no AB testing, yeah, people noticed and there was uproar about it. I never saw any indication because I've been using clawed code on pro plan you know that whilst it was all happening and I was like hang on I've been using pro plan for and nothing's been affected. So I think it's literally just a copy change on the website but again is that in shitification is that a sign of things to come where they're going to remove the pro plan from from the from crawl code? Yeah, who knows? Uh I I think I saw a tweet and I'm I I should have digged in but I think I saw a tweet tweet saying that this was AB testing for 7% of the users and they are just testing. What are they AB testing? Is it like actually that they're going to remove it or is it AB what is it AB testing? That's what I understood. AB testing to to remove it and see if if people actually care. But that that's dumb. I don't know because they know who's using cloth code. Why would they AB test it? What are they AB testing? They are just AB testing if people will get fe furious about this or not because they know who's using it. To be clear like um they it wasn't always part of that plan. They they did actually add claw code well after the pro plan was even a thing. like you couldn't like that's why I paid the $100 because you couldn't get it on the $19. Then they added it. So now it sounds like they're potentially thinking about taking it away. And I saw loads of comments on the tweets and things like that saying people don't use the $19 for code. And it's like I hate this attitude that what really grinds my gears. This attitude that people literally just think that the way that they use things or the way they are is just the like they can't fathom that people are coding on a $19 a month plan. And I'm I'm here to tell you I do. I'm quite happy to wait on my limits. I have other means by the way of coding, but I'm happy to wait for my limits. I'll go out like make my lunch or do whatever, you know, touch grass. Yeah, touch grass, you know. Um, not when I'm on a project, but the point is people can code and do it quite well on a $19 a month plan. But yeah, um, I'm uh a bit in a rush. Sorry. That's why I I need to say this and go. I can also addit it in the chat. Um, that fine with you? Yeah. Yeah. Yeah. Yeah. Okay. All right. So, this is a crazy prototype idea. This is a window to the future. So, imagine every pixel on your screen streamed live directly from a model. No HTML, no layout engine, no code. just exactly what you want to see. What? So these guys uh built a prototype to see how this could actually work. And look at this. So this is imagine as a website. No no HTML. Everything here is u an image. All of these are images that you can click and a model creates the next frame. based on what where you clicked. It just they gave it awareness to where you click and it expands on that. So you explore a completely generative world. It looks very interesting and it got tons and tons of views and they played with responsive and uh with other concepts as well as you see here for the text and RIP accessibility [laughter] by the way. I know. Uh [snorts] but this is again a prototype and Elon Musk said this and I was like what is he talking about? A while ago on Joe Rogan's show, he said on your phone everything will be stream u a streamed video like the UI will be just a continuous you know stream of frames essentially. [snorts] I don't know if I like this future or if this will be the future. This is extremely um expensive to do like because of the amount of compute that you need to do but fair these people have also mentioned in the tweets that yes they know this is really slow. These videos are actually cut and they've been clear about this and speed sped up because models are not this fast to generate videos. But it's a concept and it's interesting. It's an interesting concept. You can go and try it, but I haven't been able to because too many people are using the service. Uh but uh seemingly you could go to a go to an environment and start exploring um or upload your own and start exploring. And here you can read a uh a little bit more about it. uh what comes next. And as you see, this demonstration uses pre-generated video and has been cut for speed. But yeah, crazy idea. That's is wild. Is that flip book? Yeah. Yeah, it's video as UI. Video [snorts] is UI. Yeah. Well, I mean, it's a concept and it's a concept. Yeah. It's not It's not real. It can't be real. But there you go. Well, we that that's the thing. Will this become actually a thing maybe for some specific use cases? So, I would be interested to see what our audience thinks about this because this is a look into a possible future and I really like this idea with alternative futures, right? we as the way we can have alternative realities. This could be a direction that we might want to pick and you know spend a lot of compute for it or not. So yeah, I personally want to stay with HTML for now. Fair enough. Fair enough. Yeah. Cool. All right. Well, that's all we've got time for. Uh unfortunately, you want to continue or stop? No, I'll stop. I'll stop. I'll uh yeah, I haven't got the energy to to to follow that. I appreciate I appreciate it. Cool. Um yeah, tune us in in next week for all the um news and AI design and dev. Um if you're on YouTube, who's on YouTube? Give us a like on YouTube, subscribe, and all the rest of it. Um that's it. That's all we have to say, is it? Yeah. Do I say anything else? No. Uh buy our merch. Well, we Yeah. Yeah. Let us know if you want us to do merch. Some of that stuff looks sick. I thought you were going to correct me and say, you know, keep on vibing. That's all I've got left to say. All that's left to say is keep on vibing, baby. Peace out.