#58 Kimi 3, Thinking Machines, Codex Micro, Gemini Pro 3.5 Delays

Hello there. Hello there. Oh, no. Hello there. How are you doing, beautiful people on the internet? We got to keep talking. So, we do a sound check right at the beginning before we start so that I'm not uber quiet. I think I've been playing around with the volume, but yeah, I think we're hope hopefully good. We should um So we are live on Instagram as well and cool cool cool cool but we don't how how many people do you remember how many people attended from Instagram last week? Well right now we have nine people watching and what are you drinking water from a bottle? Yeah. Yes, exactly. W I don't know. Cool. Okay, so we we it wasn't a crazy week, but we still have some like good amount of news with a big one obviously a big one for the AI community, a big one for the design community.

So, we do have some cool cool stuff to show. Um, I have an announcement next week. Okay. I will not be here. Oh. So, it's up to you whether you want to go live or buy your lonesome. Um, but I'm going to a wedding. I'm going to France for a wedding. Oh, wow. Crazy. Nice. Oh, wow. Okay. Uh, we'll see if I will join alone or not. Uh, I don't know yet. I'm going to do a little sound check. Do the sound check. I'm double checking the list of things we have. Um Um Oh no, I didn't make the cover. Right. You need to be quiet for a second while I listen. Okay, we got to be quiet.

Okay, it sounds all right. Maybe a little bit louder. Maybe a little bit louder, but it sounds good. Sounds good. My audio goes rubbish. Do let me know because last time was an embarrassment to everyone involved. I was way too quiet. Way too quiet. Um, what are you doing? Writing the intro. Yeah. Oh, I think we're just going to wing it this week. We got carried away with a new tool that we'll get into later on. So, that's exciting. New tools, new gadgets, new distractions. Yes, something to procrastinate over and wonder whether you should or shouldn't commit to this tool or not. I'm still having trouble with it, mind. I'm I still just can't edit files.

Oh, okay. Like I don't know what we'll check mine. Yeah, unless you have to. Hey, you you installed it through codeex, right? Both codeex and clot code. I I did actually two instances and you just gave it the code. Yes, because it only gives you it only gives you an example script for claude code. It doesn't give you an example script for codeex. Uh, it just worked for me. It just worked. Okay, I'm going to do one more try. Blah blah blah. Okay, we we we have the intro. We have the intro. the beautiful people of okay Instagram. Uh where are you joining us from and what are you excited to hear about because we have a really cool tool to show.

Um I'm mind blown. I mean I showed you we got carried away with it. It the results for me were like yeah this is this is the turning point. So uh let us know and once you are ready I will start with the intro. Uh I'm ready. I'm ready. Okay. So this week Grock got uh caught red-handed. The richest man alive was stealing your hardwritten code. Um, we have Kimmy K3 dropped that is apparently better uh in some tasks, better than Fable, better than Soul. Crazy for an openw weight model that you can never host yourself, but theoretically you could. Uh, we'll talk about this. And there is a new tool that could end Framer and Web Flow.

And for real, I I actually mean it. Uh I'm I myself I'm terrified and excited. We'll show you that. Uh but all of this and more in today's episode. This is Command AI. I'm Kabarza. And I am Sam and this is a week. This is a week in AI design and web. Nice. Nice. So, let's start with the the craziest news of this week uh that we don't even have listed. Uh but let me actually share my screen. Please share your screen. So I saw this news and I was like what the actual hell you know Grock build the you know the tool in terminal with which you can I don't know create a website create a tool whatever write code we demoed it last week.

Yeah. Uh it got caught red-handed. If you would like push something to GitHub with it, it I mean all of these tools do they do read your code, whatever you push to GitHub, but in this case, Grock would read your entire GitHub, actually clone it and upload it and send it to our Lord and Savior Elon Musk. So Claude got caught stealing your entire code for real. So yeah, obviously um some somebody found this and there was like a backlash and uh if I scroll down there is there is a response. So Elon Musk went further saying all user data uploaded before uh now would be completely and utterly deleted uh with nothing left behind.

So they are saying that this was a mistake and apparently maybe a bug and they're supposedly deleting the code that they were stealing. But that was kind of funny. And the results was uh XAI going completely open source with uh Grock build which is actually cool. It is really cool that it's now open source. So do do we think that they this was intentional and that they only retracted it because they got found out? Um or do you think this was a bug? At this point, honestly, I don't know. Um, it might have been a bug to be fair, but I mean, does it really matter? We know for a fact that Mark Zuckerberg was torrenting books.

Like, you you can't go worse than this. Like, yeah. So, all the information that the AI has been trained uh with has been stolen. And as we speak, as we will speak about it, Kimmy as well, we see that now they are stealing from each other. Everybody is stealing from everyone. And it feels like nothing that you put out there like this thing that we are putting out there, we know that it will be stolen. Like we know they will take it. It's it doesn't make it right. But it wouldn't have taken private like they could they couldn't have taken private repos whereas obviously if you authenticate with GitHub you're giving it access to private repos.

Yes. Or potentially private repos. It's like giving your Yeah. It's like giving your apartment's key to somebody to go and I don't know pick like note down something from a book you have, but they just take the entire library with themsel. It's just stealing everything. So bad. Really, really, really bad. Really, really, really bad. Um, shame on you. But yeah, Build is now open source and it's quite a nice little tool. quite a nice little I would say one of the nicest outside of open code. Yeah, I I I liked it. Uh I uh I started using it. It's really really fast and yeah, it's quite a nice tool. Very good. Very good. Um, next up, next up, uh, no, then we have the the AI questions.

That video that I haven't seen, uh, I guess. So, let me let me share this video and just let's just get a vibe check on this video. I'm hoping. Let me have a listen before we balls. There is music in it. I will I will because it's got subtitles. I will do the audio. Uh the Yeah. Check out this kind of I don't know. Let me know what you think about this video from uh Anthropic. Can you zoom in please? Uh a little bit like that. Do you want you want cinema cinema screen? Yeah. Muse. Openheimer Odyssey type of thing. Yeah. Yeah. I can't hear. I know cuz there's music. Can I be trusted? Who's going to hit the brakes if we need to?

Death. Do we really need to ensure that we're aiming to achieve really does benefit the majority of people? If it ends up taking life, almost all the jobs, then what does it mean to work? But wait a minute, why do we have to have this stuff? If a machine could pretend to care better than I can actually care, how do we draw the line there? If we had all all had a voice in it, then I feel like it would be better. Oh god, if I was a woman, then could I AI help people stop feeling misunderstood? Could AI help me build more connections in the community? Can AI help me be better teacher, a better mom?

Maybe it will cure some great things, you know, things that we've not even seen at the cusp of understanding yet. Will it create a group of people that ask more questions? What if we started to be more human again? We don't missed it. There's hope in Okay, there's hope in hard questions. Oh, there is claw.com/h hardquests. I've not been there. Not even been there. Um, the Samman thought it was a joke. Um, and the response has not been particularly great given the sort of message. Wait. Uh, house on fire, brain dead child, uh, hellscape ch Chinese system. Can AI be trusted? Graveyard. And I think this is I think this this was made by a company called I've got it in my notes here.

um, Mother, an an advertising agency. And I think this kind of it it accentuate it continues this trend of anthropic fear-mongering when it comes to AI. You know, it's almost like, you ever heard of the term neg? They're making you worried about AI and that they're going to be the saviors because they're always talking about safety and like trust and it's like they care. They're just trying to install this idea that everyone else is horrible and like Terminator 2 type of So, who are you going to trust? The people who are calling it out on what it is, right? That I think that's what they're going for here. Um accountful. Okay. So, let me show this.

I haven't checked it uh myself. So, so it leads to this website claude.com/hard questions. There is hope in hard questions. Uh anthropic exists to answer the hardest questions about AI. Okay. Safety, governance. So, we traveled across the United States to hear people's questions, concerns, and hopes about AI. This is what they shared. Uh, okay. Interesting. Interesting. So, basically, what you heard in that video or what I attempted to read out was people's testimonies on what they worry about when it comes to AI. Interesting. And I think in the edit this needs to come first before our assessment. Yeah, this is by the way this is a beautiful website. How did they make it? It's very simple.

Sanity and React. Okay. Interesting. Because they do use web flow but this is not web flow. So uh I have to listen. I cannot read it. So if I go to listen um I want to just scroll. Can I So you can just scroll to the next person's testimony. Is it? Yeah. So it goes like this. And by the way, it's a really nice website. You know, I'm a sucker for this. Uh so voice of people. So Scott, engineering manager. I want to potentially listen to a few of these, but I'm I'm not sure if these companies are addressing these questions uh in any meaningful ways. Like if AI is out there to destroy jobs like we we don't have any systems to accommodate people when when the like we talk we have talked about people hating AI and I would love to know from our audience if there is anyone watching and listening who hates AI because I assume that people who watch us or like stay u they don't hate AI but I see a lot of hate and I see a lot of hate especially with students.

I have friends who are who want to study and they there is this uncertainty and all of this is created by these companies basically saying you can study but there won't be a job for you at the end of your study. So there as you said there is this fear-mongering that is yeah I I don't I don't think it's real. We see what AI can do, but I don't think it's I don't think we can out of everyone trust the token dealers. Well, I think that's what they're that's what they're kind of they're playing the reverse uno and that they're trying to make you aware of that they're aware of how bad AI is. So the reverse psychology of it is well these are the trustworthy guys.

Now I don't think these people who they asked these questions are anti- AI. I think the people they asked are pro sorry that must have sounded weird are pro- AI but were open to sharing what they worried most about it. Because if you talk to anyone who's not who who's anti-AI, you know, I' I'd be interested to hear those conversations with Anthropic. I bet they would have got a kick in the shin and a spit in the mouth. You know, they would have been like, you know, you've you've stolen all of this stuff or like um uh you're taking away uh people like artistry and artist jobs and things like that. You've killed industry. You know what I mean?

Like it's we we've all heard it before. Um, is that can you see that flickering on your Yeah, I I see it. Yeah, I see it. I wanted to show like how cool it is. So, they have anthropic.com/path to hope. There we go. See, anthropic is the path to hope. Yeah. How does AI work? What is AI's impact on society? We don't know any of like we it's funny that we don't even know how AI works. Not really really. Wow. Who should govern AI? Who should govern AI? We know that it's governments, right? Like you've heard of Fable getting um banned, right? You've something happened. You've probably heard it. Yeah. Yeah. Something about that. Yeah.

Yeah. So we we know like the the government is controlling AI, not the people. No, of course we're not going to control. No. Yeah. And you were saying um No, I mean that's it. That's kind of it. It's just more fear-mongering that I don't think is what I was saying about, you know, people who genuinely hate AI have more interesting thing to say about what they fear about AI because they it's already happened. You know what I mean? The a lot of the the damage has already been done. Yes. Yeah. Yeah. Uh we have a comment saying, "My friend is a programmer and works on AR. any possibility of being friend and work with him.

I don't know. Okay. Uh yeah, just I don't know share work and we will see. Another comment saying that I love AI. Generative AI helps a lot during uh education to fill the gaps and go further without limits. I Okay. I I really don't don't know. Like with education, you should be really careful. I I hear all kind of like comments on that. I I've I've heard people being very annoyed with AI giving them wrong answers. I mean, it it sounds like this is an issue of two years ago, but still exists. And when it gives you a wrong answer, you you might lose all your hope in asking AI for anything. you might feel like, okay, I'm going to go back to my book and just read the bloody book to to make sure that I know what I'm reading, you know?

Yeah. Yeah. And you also don't want it to take away core fundamental skills like, you know, um uh what's it called? when you uh when you think about something uh critical to thinking, yes, you know, and knowing where AI's pitfalls are and you know, it's too it's too easy for it to to sound confident and tell you that it's done this and this is the right thing and whatever and we're too gullible to believe it. Whereas, we need to grow a bit thicker skin, I think, and trust that it's not. the latest video that I've just released, I did like a comparison between all of the Gemma models, and I just asked it, what are the the gym leaders of the Johto region on Pokemon, and what's their highest level Pokémon?

Two of the models couldn't answer it. One of them got about 50% of it wrong, and the other one, which is the most intelligent one, got most of it right, but it got one answer wrong. So, you know, even the best I know these are open source. These aren't multiple trillion parameters, but like AI gets stuff wrong, you know, even when it's doing a web search, you know, and I I basically went to Poker Mania or some website or something like that where the the information was on the page right in front of me, I could read it. I got Claude to summarize it because I didn't want all of the gump and everything in between.

I just wanted a clear, concise list of the gym leader and their top Pokémon and the town that they, you know, um, uh, reside in. I had to just double check even though the information was there on the page in front of me, uh, in front of the agent, sorry. I still had to double check that the summarization was correct and it was but point is we need to know where it's where it pitfalls are and and be aware of them particularly in education and not getting I fully agree a lot of this comes back to understanding at least partially how AI works I know I said we don't know how AI works I mean we don't know like inside of the blackbox what exactly is happening but If like once you understand uh a little bit how the token tokenization and you know it's just predicting the next token works, you can also learn about where AI gets things wrong more often.

Like AI does uh things like summarization surprisingly bad. Like this was a new thing for me to learn. Like what? That that shouldn't be a thing. that like it can misunderstand things very easily like misunderstanding X for Y like converting things it it can misunderstand things. So just knowing that AI is not really good at summarization is kind of like enough for you to be better at interacting with it. That's what I'm trying to say to to get better at interacting with it. Anyway, uh I think this is Yeah, we we started with the video and we ended at Path to Hope. Yeah. Uh I don't have much else to say. No, no, no me.

No me. Interesting video. Um interesting tactic that an Anthropic goes for and I don't think it's landing very well with many many people. Yeah. Um, next up, Kimmy K3. No, no, no. Actually, we're not going to go Kimmy 3 K3. We're going to go thinking machines. So Meera Morati who you might know or she was the CTO of OpenAI uh spread her wings and flew away and created a new startup which recently raised uh 12 billion I think called thinking machines released its first opensource model called Inkling which is a 975 billion parameter mixture of experts model um yeah open weights uh model. It challenges China's dominance in openweight AI and signals a major new player in the open-source space.

So, if I share my screen, you might have you might have remembered um her from when they got when she got asked about do you remember this? When she got asked like where did they get the training data from? She was like she kind of pulled this face. No, I I didn't see I wish I would have. Okay. Yeah. Yeah. Anyway, we did we just stole it like everyone else. Yeah. Yeah. So, I haven't had a look too much into this. I saw I mean it literally got released yesterday and you know 975 billion parameters under just under a trillion. You're not going to be running this necessarily locally. Um and in fact, let's see.

Maybe let's let's hop on hugging face. Is it even there? I don't know. Um think here we go. It is nice. So um I never know how to um just look at the size of the files or or size of the the open the weights. Um one model for free. We can download it. Excuse me for a second. So, can you run this locally? Like, is it That's what I'm looking for. So, small enough for that. Um, yeah. I never know how people just files info. Here we go. What's I think this is going to break. Zoom in a little bit. Hang on. Yeah. Total size. Oh, it gives me give that in bytes. How much is that in bytes?

Well, that is bytes, right? Actually, um yeah, it's it's really tiny. bear with me. Bearing with you. So it's nearly So this is nearly two terabytes. This uh this if if my calculations were done correctly is it's nearly two terabytes. So it's I think it's possible to run to run locally. Um but a quantized version for sure. This is going to cost you a lot of money. What I tend to do if I go to hugging face um thinking she I doubt anyone's done this. Oh, someone has quantized it. Amazing. So, someone's quantized it for 4 bit here. Okay. And this comes out at half a terabyte here. So, this this can run on two two Mac uh 512 Mac Studios, which is pretty cool.

So, you can run it Yeah, but you need something Oh, yeah. This is This is Yeah. Beefy. Many many thousands of pounds. That's the That's the thing with local AI. Like, it's it's not something you can do on a budget. Like, people get excited about these small local models. And and fair play. Like, I get excited about them. I love local models that can run on my, you know, petty petty 128 gigabyte VRAM M5 Max Studio. I love models that can run on that, but it's just the the less you the smaller you go, the worse that model is, the less you can actually do with it. Encoding models roughly start about about 30 30 um billion parameters.

So yeah, I mean pretty cool. I've not used it. Can we get it on open um open router? See, open route has had a rebrand. Let me share this. It looks like AI sloped still mind, but it's Yeah. Yeah. I was thinking about the same uh about hugging face. I wouldn't say hugging face is AI slop. I don't know the the design. Yeah, I wouldn't say that's AI slop. Well, it's not on open router, so I don't know whether we can actually use it yet. Um, maybe we'll say in this article here, let's see. On one benchmarks, the company says Inkling uses a third as many tokens as Nvidia's Neotron 3 Ultra. So it it's it isn't like it's not the a massive model but it is in the realm of like you know it's affordable.

It's not it's not as we'll get into Kimmy K3 level of um compute needed for it. Um that raises the question of who within the enterprise market it's targeting. Thinking machines is for now marketing inklings less as for now marketing inkling less as a finished product than as a starting point. Something for organizations to fine-tune themselves through tinker. Interesting. I want to get into finetuning. Yeah. So, I don't think I don't think anyone's serving the model. I think it's something you have to you have to run. But anyway, either way, some some good news on the um open weights front. I unfortunately think that was overshadowed by the next thing in our list, which is Kimmy K3.

Yeah, let's uh let's move to that. Uh, let's go. So, not long ago, we had GLM 5.2, which blew certainly from a front-end design perspective, Fable out of the water. Well, maybe that's a bit of an exaggeration, but either way, it was touted to be better than Fable. And then probably less than two weeks later, Kimmy, the other kind of Chinese open source competitor, lands Kimmy K3, which really does blow Fable out of the water. Like, this is actually pretty insane. Jumping right into it. I mean, to be butting up against Fable and GPT 5.6 on on Deep Su. This is by the way a closed source benchmark. So you can't you can't um benchmax against deep suite because you don't know you don't know what the test is, right?

Terminal bench is beating Opus 4.8 and just just nipping at the heels of uh 5.6 Soul. It's winning on program bench which is crazy. Uh same again here. It's nipping at the heels of of Fable 5. These are crazy benchmarks. It's a 2.8 trillion parameter model, which is probably one of the biggest open- source models I've ever seen. So, when we talk about open source and running locally, you need a lot of lot of hardware to to run this one locally. You're definitely going to be waiting for the quantizations of this model. Uh, which they haven't released yet. They haven't released the weights yet. I think it's coming January 27th or something like that. Uh July 27th.

So in 10 days they're hopefully going to be releasing the full thing. Um you have access to it right now through Kimmy uh and Kimmy code which is quite a nice CLI actually I have to say. And you which means that you're always going to be going through their servers. So again something to bear in mind. These are Chinese servers. they can see everything you're doing. And these are the sorts of prices we're looking at for those um coding plans. However, once they've released the weights, this is when other people can host it and probably be able to host it on US or European soil. Um and well, let's see what the prices look like when they're on there.

But it's definitely an amazing feat of engineering. Um, looking at these these uh benchmarks here, like there's GLM already the lowest setting there already beating GLM at roughly the same price. And yeah, I think this is an insane insane model. I think this is a Oh, and there's Inkling there. Look, just coming out. They were head to they were neck and neck there and then boom came straight out of the gate and just completely killed it. But it is 2.8 trillion parameters. So this is a third of the size. So you know the end of the day the more you put into something the more you're going to get out of it. So these are the open uh models getting better.

Yeah it is crazy. It really is. Now once you scroll a little bit more down I mean yeah here you it it's neck and neck to close to every like frontier model and in some tasks winning it's insane here apparently it made these games and in this case it made even the models that's insane. Mhm. like they have an understanding of the 3D world and you see there are like many examples of like games. It's it's just getting out of hand what these models can do and I I'm like how can Open AI and Claude uh Anthropic make money like I mean how can they justify Yeah. Like justify their entire existence. Yeah. I don't know.

Yeah, I don't want any music to come on. I mean, Cyberpunk web swinging game. Yeah, these are these are great. And again, for the price you're getting, it's it's pretty pretty darn good. It's interesting. Oh, no. Actually, sorry. Scrap that. Ignore me. Um, but yeah, I've I've got one more day of Kimmy. I've been using it a little bit today. Um getting it to solve some like light problems or something like that. I still don't have the confidence to give it like um anything outside of hobby projects just yet, like the confidence, but I think I'll need to use it just a little bit more to see what I can actually throw at it and what kind of yeah, what kind of work can we give something like this.

Uh again, this is on Chinese servers, so maybe something to not throw your uh big open source project code base at at all. Again, just little shitty little projects are fine for now. And then we'll see what happens after. I don't know if it will be will it be as bad as Grock. That's the thing, man. That's why that's so disappointing that that's they've already revealed themselves to be absolute scumbags by, you know, taking advantage of our trust. Yeah. Don't know. Don't know. Um, just as a last little thing, Kimmy K3 agents, download or update the latest Kimmy app from your mobile store. You can download all that there. Um, I mean, I guess you can use it in the in the in the whatever.

They've got their own chat interface as well. Um, pricing is 30 cents uh in uh for sorry for a cash input. Uh $3 for a cash miss. So quite a bit massive difference between that and that. So once you've cashed it's pretty cheap. But the first one is $3. This is this is set prices I think or old Sonnet prices I'm pretty sure which is quite crazy. Um but is it as efficient with tokens? Um I'm not that this is also something to look into. We we can look at um we can look at it is cheaper but it might need more tokens to do the same work. I don't know but just just something that to look for.

Uh I'm looking at their docs and it can also do like video editing particularly good. That's actually a really good point. That's a really good point. So because it's multimodal, right? Yes. So GLM was great, but it wasn't m multimodal and it is very uh helpful to be able to pass through visual cues on whether something's working or not. By the way, I I don't I'm not I'm not in the same camp as some people who are like, well, how does it how can it do front end if it doesn't have visual capabilities? Funny enough, you can interpret a lot through understanding code. You know, it's like when you look at something, oh, that's about 20 pixels or whatever.

Like, cuz you're looking at something. The reverse is true. If you can see 20, you can visualize things in your mind. So, I'm not 100% in the camp of like just because something doesn't have visual. I mean, like I say, it scor 5.2 scored better than Fable. Um, and it doesn't have visual capabilities. So, you know, proof is in the pudding there. However, um, yeah. So, it can do video as well. It can output video. Is that what you said? It can edit video really well because I mean it's it is smart and it can take screenshots and uh based on their own documentation, but this is what I would love to test and I I'm not sure like when we when we talk about something like video editing, what are we talking about?

Is it like using a tool like um what what was it called? the the VF well that would be the underneath technology but the motion the remotion and what was the other one I made a V hyperframe thank you it it's not using those right so it's just ffmpeg underneath and then figuring out yeah so uh Tar did a really cool breakdown on how you can edit videos with AI and it does a few things. It does a few passes. The first of all, it translates the audio. So, it does an audio like um what do you call it? Uh captioning. It not transl. So, it transcribes the video and it might use it might use another tool to do that.

You can you've got a local one called whisper which you can use. So Kimmy might choose to use whisper to translate something that gives it a JSON response of all of the text that's associated with a particular video. It's then able to infer the timestamps and whatever from the transcription to be able to make edits through FFmpeg. So it's all done programmatically. It's all done through CLIs and it can put it together like that and then um have a bunch of separate videos that have been edited independently and then merge them all together again with FFmpeg. So it's a it's an understanding I think it's an understanding of a few different things. Do you know what I mean?

It's understanding of context. this understanding of um nuance of editing specifically because my biggest thing about video editing and reason why I don't use AI for video editing it is mostly it doesn't doesn't have a sense of humor AI no doesn't have a sense of humor it doesn't pick up on timings it doesn't pick up on um breathing room it's very like literal so when you have a talking head shot and you're just talking something to the camera it's great for all of that but it's just it's So matter of fact to the script. I see removing arms, you know, and things like this. This is just an example. I don't know if this is probably too much of a derail.

Uh you can remove it later on, but just to show this is done with clawed code and um hyperframes. Oh, you got a call. Yeah. Yeah, I got a call. I mean, it's my dad. Let's I I have things on do not disturb is on but somehow ah because family they go through ah I was going to say what app is it telegram but it went through um Okay. So, yeah, my dad needs me and maybe I can do like a 10m minute pause. I'm not sure. Let's see. So, anyway, just to quickly show it, I gave Claude this video. This is just to test something. There are some things on my screen and then I have myself on the right uh on the left.

This is how I uh how I make videos. Uh it took some screenshots for itself to understand what's up. Uh, and it did a transcript of like literally every single word has its own start and end. Yeah. Um, it has the silences the that that's interesting. So, uh, map words. I don't know. Anyway, so it it it creates a few. Is this a J? What is MJS? Because this is a JSON, but I don't know what an MJS file. Um, it's a JavaScript. It's a JavaScript like thing. Okay. Anyway, so it did that and then it rendered a few versions and then let's see a sample. So this is a sample I asked it to do.

So it does like these animations etc as well. So it can also do like these transitions obviously. Obviously it can do all of this. So it's pretty cool to see it. But I was wondering how Kimmy is doing this if it's Well, it will be do it will do the same thing. It will do the exact same thing. So, do you know how it did that transition? Were you watching? Uh, what tools? No, I it I I know that it I was using hyperframe. So, it's technically just Gap. So, it's just putting the frame. Yeah. Yeah. That So, it's GP. Yeah. So, um, like it would need the raw video to have been edited. Like, hyperframes needs just one video.

Like, you're not going to edit a video. And so, it will probably do a pass of the video, matching the transcript, uh, render that out with ffmpeg and then putting that into a hyperframes website and then doing that animation and stuff like that. Like hyperframes can't cut out ums and rs and things like this. Hyperframe is just the kind of like the the tool and the composition tool. Yeah. Yeah. Yeah. Okay. So, is it everything on Kim K3? I think so. Well, um just quickly with the just go through the quick benchmarks like this is the intelligence. This is kind of worrying in a way how close it is to fable and intelligence. Like this is this was the biggest thing about Fable is it intelligence I think and it's like there do you know what I mean?

That's crazy. Speedwise not not so I want to I've downloaded uh GPOSS. I do want to try this because it's always up there on the on the speed like unbeatable. But again third place here got electrocuted. Uh, but you said something. You said C cost per token, didn't you? Yeah, I just said I just asked if it's not cost per token, but per task. If it's efficient with the token cost per intelligence index task. This is kind of it. So, it's like around 95 cents, which um Claude took 275 and 255.6 took $14. So just a little bit cheaper than soul. So it's not Yeah. So not cheap. The tokens are cheap but it probably takes more token to achieve the same result.

Yeah. Costing. However, more Yeah. Yeah. Exactly. Cool. That's that's all I would say. So yeah, Kim Kim K3 I would um I do want to keep trying out, playing around with it. Yeah. So, what should we do about this thing that we want to show because um I I probably got to go. Okay. I Yeah. Uh I had a heart cut at 30. So maybe I'll show this and go. Yeah. Yeah. Yeah. Yeah. Do that. Do that. Yeah. And maybe you can do the installation if you want to, but no. I I I we don't have to go I I'll just show it. I'll just throw it. No, no. Yeah, show it and I'll do the installation when you when you leave.

So, we can check it together and um or we can just add it on at the end and people won't be I'll just show I'll show what I show and then if you feel like you need to add more, you add more. Okay, cool. Yeah, there's a new tool that might put an end to Web Flow and Framer. And I don't say this just to say it. I'm genuinely excited and a bit terrified. So what you see here is a tool for no code that is AI native and it's all open source running locally on my machine connected to your codeex, your cloud, whatever you want. And yeah, it can it it can create designs.

It has a CMS. You can host things. You have users. There's just so much that I want to show you. So, somewhere to start, we we have a dashboard. By the way, this is this is the V, not even V1. This is version 0.0.1. So, it's very early. Uh, you have a site. the idea of like you add a site and then you can just get in and do stuff yourself or do it agentically which I will show you in a second but you have you know your pages you have your layers as you see and for my web flow friends I don't know if I can find an element here that has multiple classes here you see like I can remove the base class something like that you can't do in web flow crazy right 2026 so it has everything that you need here.

Literally, you have this view, you have the canvas view, you can like execute the code. Uh you can like if you look at here to show you like you you have like animations. So a lot of things like this, you have obviously your um different break points. And you know what's crazy? You can edit those break points. you can add your like new break points. So you have all of this. You have content and data which I haven't wrapped my head around completely yet. Uh so this is CMS but I I'm trying to understand what is the difference between content and data. It's super early in this but anyway content seems to be like a rich text and in this rich text I can do slash and do table crazy right 2026 we don't have this in web flow you have all of this here uh yeah I mentioned data you have a database you have the the media you have plugins I haven't installed any plugins and now you have also users uh So imagine you build a website for your client and you want to add your client as a new user.

You know what like we can just add a new role and as you see we can allow them to have access to the dashboard or not to the site to pages to content data media runtime storage plugins AI user roles audit like all of everything separately. This is literally $100,000 worth like this is like all not just web flow every enterprise website uh website builder that sells like this access levels they charge I don't know like a lot of money and now you have all of this here opensource MIT license uh so let's go to the site and see what we have here by the way like here we have also code which I haven't looked into yet.

But yeah, you have that uh you have uh a framework something that I guess is from the tool that um gives you but you can discard this. You don't have to use it. That's the beauty of it. You can manage framework and use none whatsoever to do whatever you want. It's yours. Uh so you have these but let's talk about how I made this. How did I make this? So, this was designed by a human. This is designed by uh my designer on uh my team in Figma. Beautiful work in Figma. It was designed in Figma. I did try to bring this to Webflow via the MCP. Let's just say it was not really good.

Uh and I just gave in this case chat GPT. I I'll I'll show you how. Access to Figma and said like do this and it it did it basically one go. No correction, nothing. I did a test for the first two sections to see if it does it. It did it and then I said, "Okay, do it for the rest." And it did it. So one go basically did everything. Uh so you have this AI assistance panel here. You can add an API, but that's not what you want to do because that's going to cost you a fortune. What you want to do is create an MCP. You create an MCP and you are basically connecting this tool to your AI whatever that you are using and then from there you take it.

So I'll show you how that is. So here we have my uh codeex and this is like my my chat creating this. Um, so the way I did it, I went to the AI assistant here, AI settings. By the way, depending on when you are watching this video, I can imagine a lot of the UI changing, but what you want to search for is it's also slash AI and here you go to MCP and you create a new connector. Now, what's really nice here is you can create a connector for local or remote. For example, the local one is for yourself. The remote one is when you want to give access to your client. So imagine you can create an MCP server for your client but give it only access to site editing like to the content editing but not publishing just for example.

So your client can edit but they cannot publish so they don't break whatever like whatever access level that you want. You can have different MCPs and give it give them to your clients. And once you create the MCP let's just create one. We give it a name. Actually let's say test two. I give it a name or I have to I need to put in my password and then you get a token or like this basically prompt. You can actually just ask your codeex or cloud code paste that in and say add that MCP and realistically you need to reload your session or like create a new session uh to to load that MCP. Once that's added, you can just talk to it.

So you talk to your AI and you say create a new page. So I can say create a contact page one section. So just just to be fast and now in this case codeex is talking to this tool and everything is still on my machine and it goes on and it it understands by the MCP what it what tools it has access to and creates it. Uh what would be interesting is to see if it still has access because we did turn it off actually. Oh yeah, but I haven't reloaded the session, so I don't know if it can do it or not, but you get the idea. There is nothing crazy about this. You you know how that works.

You talk to it and then it it um it it uses the MCP, the API layer to create all of that. So, why am I so excited about this is it's a tool that is doing the AI part really well. It is open source. It's on my machine. I can literally create not just pages here, not like these websites. I can also change the tool itself. It is open source. It's MIT. I can tell chat GPT to change the content of the you know the site or like the tool itself and it even though it's such an early version it looks very like complete to me or like not just complete it it sounds like to me it looks like they have a good sense of where they want to bring this to and yeah this can potentially bring your hosting fees down.

You can have your clients websites here and they have access to the CMS. They have you have multi-users. It's a really good sp uh starting point. So here you see the contact page just got added. Um yeah, it's it's not perfect. I didn't give it an a design or whatever, but it started it's it tried to use the same design that we have. And yeah, so just to show you what it can do. Pretty cool. Really cool. Really cool. Do you have to go? Yeah, I have people over and I have to go. Uh I don't know if there is anything people are asking. Oh yeah, because I won't be able to see if we leave your uh thing on.

I won't be able to read their comments. Yeah, my Instagram. Uh, I don't know if I missed anything. No, like that's really good. You you you used the side panel to do any edits. Did you you change the text or whatever? Show No, but I didn't do that. So just okay just to clarify why I'm so excited about like this being finally being something that feels like client ready is that I can have a client with edit ro where they can see the the thing and just edit it on canvas like I can edit this I can edit this like these things are fully editable that's why it's really nice and I can limit their access in addition to obviously having an MCP for them to edit through chat GPT or through the canvas.

So all of that together it does feel pretty complete. So I cannot wait to play more with it and potentially I'm not joking. Um my my site this is my new design by the way showing this for the first time. Um, shout out to Ryan for creating like this banger design. I might legit just continue building it here. Like what is the point of this in doing it in web flow? RIP Web Flow. I I really don't want to sound like that and and frame like I mean what Okay. Okay. Ju just last question before I go. I know framer community is much bigger nowadays. much more active, but we still have the Web Flow community and like genuinely asking like no fear-mongering, nothing from designer to designer, agency owner to agency owner, freelancer to freelancer.

Like why would you use Web Flow and Framer over this? Like why? I know there is a learning curve like a tiny but to sort it out but you you could pay someone just to get it up and running and then that's it. You know, you pay someone quickly to get this up and running on a URL and then you don't there's no learning curve. We'll go through it after you go. We'll go through how to get this all installed and stuff like that. But point is if you don't want to do that, it's a 20inut job and then you've got your own bespoke version of Web Flow for you and whoever to contribute to, right?

Yeah. And so there's no learning. Yes. But like even the even the learning curve if you know web flow and framer I promise you the learning curve here is 10 minutes 15 minutes. It's very familiar. Maybe it is ugly in some places like this is ugly, right? This is no web flow and framer. This is no web flow and but does it matter like the the whole point is that I want to do the to let the agent to do the bigger like part of the job and then I want to have controls to do refinement if I need to if I really need to. And this gives me all of that. And if I don't really like this panel, it's on my machine.

Like I can't tell Claude, hey, redesign this panel to do whatever that I want. Does the MCP allow you to manipulate the source code or does it just allow you to create the components, the pages and stuff like that? From what I understand, the MCP is only for the content essentially. Okay. But on the same layer, I have access to, you know, the whole tool. So, yeah. So, as long as as long as you're running it on your machine, you can edit the code. Yeah. Nice. Yeah. Cool. All right. I got to go. Um, but that question I really want to read the chat and see what people think. So, anyway, so I gotta go.

And what what is it called again? Instatic. That's the one in static. Oh, I I went the whole thing without naming the tool. That's the way to do it, by the way. That's the way to do it. Like a billion comment. What is it called in static? All right. Uh, okay. I'm I'm out. So, take care and thank you and uh see you around. See you not next week, but week after maybe. Yeah, we'll talk. We'll talk. Cool. All right. See you later. Byebye. So, I've put in the I've put in the chat. Hopefully, you can see it. Um, no, just in just in YouTube. Uh, the link to the GitHub. And this is it.

I mean, we we'll just install it. We'll take you through. And I know specifically that Cabaza's audience hopefully you can hear me and it's not. It's not too bad there. Um, and I know many people struggle with some of these concepts and there's a lot to learn in terms of how all of this sort of stuff comes together. So, let's just go through it step by step and and talk to you about it. So, you access this, you get the raw code. That's what we talk about when we talk about open source is you get access to the code that makes up this whole project. Okay, they're all here. Admin, core, all of this stuff is the code that makes up all of that beautiful UI that you saw.

So, we get access to all of that. So, we need to download this code. If we go here, we go to quick start, it tells you how we're going to do it. So, this is what we're going to run through. You need bun installed. So, you're going to have to go install. I'm doing this offscreen. Bun Mac. So, I've just I've literally just Google search install bun Mac and you're going to install that. And Macs and PCs, I think they all have git install. So, you don't have to worry about anything regarding that. And this is a this is the line of code that allows you to just download it or what they call clone onto your local machine.

Let's have a little quick look up here. You could probably also just go to releases here and just install or download the source code like like that. So that would be another way to do it. And I want to do a video on running like terminal cheap like how to use the terminal for beginners doing all of this stuff getting into all this. I've run through it in one of my latest videos on basic terminal usage. But point is um yeah it can be quite tricky. So, let me I'm going to share my screen as opposed to and I want to try and moni I can only see comments from my Twitter and my YouTube.

I cannot see comments from Kabaza's Instagram. So, apologies for that. If I remove this present, stop screen, share screen, entire screen, screen two, and we're here. Cool. So, I'm inside a terminal. This is warp, but you can download if you just um command space on a Mac or or download PowerShell or Terminal. These are both applications that basically give you this awful black and white UI. And you'll want to cd into a folder, change directory into the folder that you want to work in, that you want your website to live in. This is just all locally for now. You you don't have to worry about anything else. And oh, sorry, I'm not sharing my screen.

This is the problem of doing it on my by myself. So there I just did CD. This is warp. And the easiest thing I would suggest you do is grab your finder. I like all of my websites to be in the sites folder. So, I would potentially just drag sites into that. Just got rid of the CD and that will navigate to the sites folder. This could also be uh documents. You can type it out or you can drag it in. Uh Samuel Gregory documents. Whoops. Documents. Drag that in. CD into that. Right now I'm in that folder. I actually want it to be in the sites. So I'm going to navigate to sites. Now I'm in my sites folder.

Then we run the git clone. this I wouldn't press enter just yet because this is going to download it's going to create a folder called nst static and download it into there whereas you need you need a unique installation for every single website you don't just have the one installation of this app and then create loads and loads of websites no no it's one installation of instatic for every website that you want to create so let's as an example let's do the command AI podcast website so a space after the git and then the name of the project you want to do. So, we enter that and it's going to clone and download all of those files into the sites folder on my computer.

If I go into my finder here and I go into sites, there's that folder that I've just programmatically created. Can you see that with my terminal? And if I go in here, there's all the files that it's downloaded. And this is exactly what is on the the the GitHub. These are all the same files just here. And we want to cd into that folder that we created. And remember you installed bun earlier. you do bun install which what this does is it installs all the libraries and everything that this um this needs to function and and to no time at all these are all the libraries that it needs and then we just do bun rundev and all of these instructions are hopefully this is all making sense I'm being very very simplistic here these are all the instructions here the only thing we don't do is we created a folder ourselves and then cded into the folder that we created instead of just in static.

If you just run this, if you just blindly copy and paste this, then this will work totally fine. It's just if you want to have multiple websites using in static, then you don't want them all to be called in static. You'd have to rename them or whatever. So now that's running and looks like everything's good to go. So I can visit that URL and here we go. command ai podcast site. Uh I'm going to use uh test atemail.com, but you'll obviously want to use a proper email. I'm going to put in a password. And this is now creating our site. And if you look at the URL, I'm not sure whether you can see this, but it's the it's a URL/admin dashboard.

So, if we just do this and just do that, it's going to error because we haven't created a website yet. So, let's create a new page. And I'm not an expert in this, by the way. I got I got shown this literally just before the the show today. If I click publish, and I'll put my password in to publish the website. This is all local. This is all on my local machine. Here's our white website that we just created. If I added something here, I don't even know how to add stuff. I really don't know how to add stuff. Paste HTML. Oh, that's cool. You can paste HTML. But realistically, I want to add what's code insert module.

Let's add a container. There we go. And then add uh some text as well. There we go. We just added some bunch of text. Really sorry that this looks like utter balls. But if we refresh that there, there's our text. So we're creating our website. We're done. And obviously you'd use this um and you create your website. You can either do it through the panel here or like Cabaza showed. If we click on this AI button down here, open AI settings. You can create an MCP. Uh let's call this Claude code. Name it whatever you want. really doesn't matter. And I would say no expire there. Select all. So it can do everything. These are all it permissions that it can do.

And now what you can do it creates you a token and a URL. If you're using clawed code, just copy that into your terminal. Oh no, it's gone. That's annoying. If we create create a new one because you can create multiple two. Select all. So yeah, we can add this into clawed code now. So if I split this with command D um and then do that and then run claude. Um can you access my instatic website? Hopefully, this will just create it because it knows. Here we go. Check connecting in static MCP server. Here we go. We're good to go. Now, what I might do is get something from Figma because this will be a workflow, right?

You might have a you might have a um let's these are some of my um thumbnails. So, let's just let's just copy the link to that frame. There we go. Right. Take this Figma frame and make a new page in in static. So, this should use the Figma MCP to go and get the Figma design, translate that into Instatic, and create a new page and do everything. So, let's just see how this goes. Hope this is this is very off pieced from the normal uh normal Command AI show. We we normally just go through the the actual um the news, but why not just go through stick put it on auto mode and hopefully you can hear me because I'm all on my own.

I don't know what's what the sound is like. Here we go. I'll pull in the Figma frame design then build the instatic website first. Wonder if we can see it whilst it's doing its thing. And by the way, this again, this is all completely local. Completely local. There it is. Oh, just created a page right in front of our eyes there. So, it's not done anything yet. Shout out to 104 people on Twitter. On my Twitter. Um, let me know how you Why is there 104 people? Like, I know I'm chatting all the time, but that's uh that's a lot of people. Let me know where you found me from or why you landed here.

Was it the title? It's probably the title, isn't it? Probably something to do with Kimmy K3, I bet. structure is in. It's using the command AI design skill, which is what I don't want, which I which are really what I don't want because that's not something I wanted to use. So, it's probably going to change the design a little bit. Um, there we go. Boom. And I know it's a thumbnail design that I've created in Figma, but could this just could just as well be a website. And it's done it. And now I can edit that. I can change. Well, let's have a look here. What is I don't again I don't know this interface.

Spacing, typography, font family, inherit. So let's just do uh what's uh times. There you go. Times New Roman. There we go. So you can change all that kind of crap. So there we go. Um then the last thing to do and then I promise we'll move on is and I won't do this because it's going to take too long. Now that you've got all of these files, if I go here and then why can't I see my files? Um let's just go like this. Um ls. These are all files. Then you literally just upload these to a server. You literally just upload them to a server. And then now you've got this online. This all this is all locally running on my machine, but you'll want this online.

So if we go to here, I think they've got some instructions to deploy it. Here we go. So you can deploy on railway, on render, on docker, you know. So you you'll get this online. And then then you'll add users so they can log in and do all that sort of stuff. So that's that's in static and installing it. I hope that was useful and I hope it help is helping people um you know understand the whole terminal stuff. Let's move on. So, Apple Intelligence has been approved to launch in China, but instead of being powered by Gemma or Gemini, it's going to be powered by Quen, which is pretty mental. Uh, here we go.

So obviously there's a bit of um let's say tension between uh Apple and China and America and China and things like that. I'm not 100% sure if this is something to do with the tension between America and China. However, obviously Apple were running into issues getting Apple Intelligence approved in China. And they've come to a conclusion to say China's Internet Content Regulator, the Cyerspace Administration of China, approved Apple's AI services in the country on the back of the deal to integrate Alibaba's Quen AI model into Apple's operating systems. Now Quen is the biggest competitor to Gemma which is Gemini or Google's open-source models. Um so it's really interesting how there's there's this obviously this um sub rivalry that's going on or sub tension or something like this.

Uh and uh there we go. Apple was facing issues adopting its models for Chinese customers. So um it it was uh said to be exploring integrations with deep seat and bike dance which is Tik Tok. It didn't provide a time frame. It also said the integrations would involve AI capabilities like text and image understanding and generation. So not a lot more to say about it there. Just a really interesting collaboration between Apple uh getting their devices in China. So, you saw me flash something up there. Let me let me get things sorted. So, you might have heard that Open I've been working on a device, a physical device with Johnny IV. Well, this is not that, but it's something very, very interesting to say the least.

Open is teasing new hardware for codeex. This is the new device launched by OpenI in combination with work louder and it's this basically this fancy little device here which has knobs, it has screws, it has joysticks, it has lights which look really really quite nice actually. um the fact that you've got these different lights there and I wonder whether you can change the colors or whether they're set. Now, if you're a longtime viewer of the show, you'll know that I already have this device. This is Framer's collaboration with Work Louder. And all Framer did, I've got the All Framer did was give you a blue knob and I can It's up here. It's somewhere. Hang on.

Stay here. Stay here. There we go. And they gave you a publish button to go with it as well. So, that's that's all that's different there. Um, however, this AI this this this one, sorry, I should have uh made it bigger. It's It's exactly the same, but I just get this. And this is a multi Actually, to be fair, this is multicolored. Everything's Everything's multiolored. Maybe I can even turn it on for you. There it is. So, yeah, that's blue now. But if I go I've got it set up so that if I go to terminal, it's set up for claw code. Oh, I haven't got the uh helper thing running. This is what I'm waiting for.

Looks like there's some updates to be taking place. Don't think we care really to be fair. Either way, it's the same device. If I go through to this. Um, but they've obviously given you some key caps because again, if I show you this, I got a bunch of key caps that are more tuned to framers requirements from a device like this, which obviously I don't use. I've kind of made up some there. Um, yeah, it's quite a nice It is a really nice to the secret source. I thought I'd be using this all over the place, but really the secret source to this is uh I use it for editing, but you need because you can change basically depending on the app, you can change all of the configurations and layouts and whatever, not layouts, but the configurations.

The issue is because there's no screen on the buttons, you kind of forget what the buttons do. So, uh, that's why those stream decks are really cool because they actually have screens on them and you can see what the buttons actually do, unless you use it for hours every single day, you know. Um, I can I use this for editing a lot, so I remember all the buttons, but as soon as I switch it to clawed code, which I have it set up for, I forget what all the buttons do. So really, they they tempt you with the idea that you can have different configurations for all the apps. But again, unless you're using those apps all day, every day, you're going to forget what the buttons do, unfortunately.

But it is still really nice. And I love the white. I really do like the white little um codeex logo. And here's the thing. Chat GBT, OpenAI, they they're trying to get rid of the codeex thing because it's so confusing. And now they come and bring this back. So, it's a bit of a I don't know. Very much a um Yeah, I don't know. I kind of want it. I kind of want it. How much is it? I think they're quite expensive. I think they're about 150 bucks. 230 bucks. I think mine was 170. That's a lot. They're sold out anyway. They're sold out anyway. But yeah, so not the device. So not the device that we were hoping for.

However, a pretty cool device nonetheless. Moving on. Last but not least, we're at the end of it now. So, we've been hearing rumors that Gemini 3.5 Pro is about to launch and then it got delayed and then it got delayed again. This is the third time that Gemini has been delayed and it's a bit of a shambles and I can only assume it's because they're just not hitting the benchmarks. like Gemini is a good model like for just chat intelligence doing all that kind of stuff. It's not a bad model. It does the job for for a lot of people. I've always thought that I like it as a an image model as well. Um but for coding for real world usage how to sorry I'm just addressing a comment here.

How to make an LLM model? I am CS students. I want to own models to make my own AI. This is this is a very very expensive task to make your own AI model. Very expensive. Anything AI is very very expensive. Prohibitory for a lot of people. What I suggest you want to look into is training your own model based off of something like an open-source model. But these require still require a lot of hardware and you're going to be doing it for days and days and days. Check out PewDiePie's video. PewDiePie trained his own model or I think he trained his own model or fine-tuned his own model and it took a lot. So yeah, uh I want to look into my own personal channel which is uh 0x5M5.

You can check that out on YouTube. That is I'm going to be looking into fine-tuning my own models on that channel. So yeah, cheers for the comment. Um, but yeah, Gemini 3.5 Pro has now missed its launch target three consecutive times. June, then July, and now. Reporting Wednesday that the rebuilt model is still played by hallucinations and inconsistent outputs that prevent it from clearing basic reliability sounds. So, it's not actually an intelligence thing. It's not a benchmarking thing. It's actually that it's just fallen to Um, Google IO set the expectation. Oh, yes, I remember because I think we've got 3.5 Flash. I think we have knowingly lesser models for 3.5. It's just the Pro model that we're after that we're waiting for right now.

Um, so yeah, I don't know. I mean, at the end of the day, oh, some benchmarks here. Neither gap is cosmetic. Recursive tool call instability is the defining capability for injected coding model and inject coding is still is the use case Google positions the entire 3.5 generation to own. 3.5 flash already demonstrated the category was achievable out Gemini 3.1 Pro and Terminal Bench. MCP Atlas and Google confirmed both figures in their 3.5 launch post. Let's actually rem ourselves with this. So yeah, this is so they released Gemini 3.5 Flash which outperformed 3.1 Pro. So it's still even 3.5 Flash is is good in comparison apparently better than Opus 4.7 and not as good as Super 5.5.

So who knows? By the way, I mean, Google are just a bit of an embarrassment really, aren't they? To be honest, not not a great look for them. So, yeah, that'll do it for this week, I think. Uh, Kabaza had to shoot off a little bit early. Not a lot going on this week. Um, give us a like over on YouTube if you are still over on the Twitter. Shout out to 108 people over on my Twitter. I love you guys. Thank you so much. Um, let us know if you want us to cover anything specific next week. Give us a like over on YouTube. I've been Sam. Catch us next week for all the news and AI design and web.

Nothing left to say, but keep on vibing and enjoy your weekend. Hopefully, it's super nice where you are and there's lots of food and sweets and ice cream for you to have.