#15 GPT 5 fallout, Prompt API, China Chip Sales, Claude Supports 1m Context, Miyagi!

Welcome to Commander AI, where we press command F on the internet's AI news and somehow still can't find our sanity. My name is Samuel Gregory and I'm Karza and we are here to decode the weekly chaos in AI web design and development. This week we have Chat GPT Fallout essentially chat GPT 5 being really bad. We have Google uh Google API. No, Google Chrome's API integrated in Wait, I I messed that up. What? Google is embedding the uh Nano AI API inside of the Google Chrome. Yeah. Right. Yeah. Okay. We have the US government taking away 15% of the sales that AMD and Nvidia makes to China and Anthropic uh just increased their context limit to 1 million tokens.

We'll talk about what that means. We have a new app for Web Flow called Miyagi. Miyag Miyagi Miyagi. We have a new app for Webflow called Miyagi, which is an AI app that helps you do anything. We'll talk about that. And we have Perplexity trying to buy Google Chrome from Google for a whopping $34.5 billion. It's It's bonkers. They are worth half of that. We'll talk about uh all of that in this week's episode. So, not so smooth. So, first of all, the the craziness of GPT5, the diversity across the board is wild. Sam Alman even kind of issuing an apology here where it couldn't have gone, it could have gone a little bit smoother, let's be honest.

Um, people love it, people hate it, people miss it. It's just all of the emotions across the board here. Um, first of all, yeah, let's go through the tweet here. It's Sam acknowledging basically that 40 is meaning a lot, which is interesting, you know, that people grow are growing attached to these AIs. um even if GBT5 performs better in most ways and there's that objectivity there of just like just because it's better doesn't mean it's more favorable you know um who defines what is what is better really so so sorry to cut you off but like to backtrack a little bit let's just explain what happened uh OpenAI released chat GPT5 last week and there were tons of videos like my social was full of people saying how amazing Chetch GPT5 is.

I tried it and I was not blown away. I tried it and it felt like the same thing. It was the same thing to me essentially. And now we are living the fallout like people uh being really at angry at open AAI and also at creators like at YouTubers at you know people making videos why they they said GPT5 was so good while we are experiencing it being not so good. So, as you were showing that a lot of people prefer the old version uh over the the newer better version as Sam puts it. So, yeah. Um not not uh not a great start for them. Um people just comparing GPT40 to 25 which is so interesting because you get obviously would can't generalize like too much but you know some people just don't care.

they just want the model. They just want it to do its thing and whatever. But when you actually show that there is a difference between models and they can see it, there is actually a desire to actually maybe go back. Are you all right? You could Could you hear that? Maybe I Okay, fine. I was like, um, but then we now we now let's get into the kind of meat of what Sam's trying to say. Longterm, this has reinforced that we really need good ways for different users to customize things. We understand there isn't a model that works for everyone and we've been investing in steerability research and launch a research preview of different personalities.

Um, some users really like emojis, some never want to see them and probably never wanting to see them at all. Cold logic where some people want warmth and whatever. So it is complex you know you're trying to trying to have one one one sizefits all when it comes to just you know providing something for everyone why not add this to memory though this is my question like sorry why not add this to memory though I what do you mean so the the way I see it their naming sucks like I I shouldn't be the one who chooses between models even though I'm quite ticky and I enjoy having a lot of customizability but at the end of the day if the model does what I want I don't want to go and choose the model for each use case I think in that sense the idea of the router that they had to to have one name that uh in the background essentially picks the best model that idea I find it good what is really bad is that you know pe people have different preferences and what is bad when the AI model basically is not customizable to any person and the way I would at at least like think of approaching it would be similar to DIA and memory uh and just when I tell chat GPT never use emojis just just remember it remember it for my new and future chats as well just don't use emojis I know it's not just don't it's quite difficult to make these models do what you want to do.

But that's about the approach that I would like them to uh to take. C can you confirm can you confirm or deny that it hasn't like have you have you asked it to not use emojis or have you asked it to not do or do something and is that now committed to memory and you're saying and now you're saying chat GPT5 when it rolled out it ignored a bunch of stuff. So, what I'm saying is I tell chat GPT all the time, don't use m dashes. And I still repeat myself. I should Yeah. Yeah. I still see them. They they appear less, but I still definitely see them. Um Yeah, m dashes are annoying because Yeah, I think that's one of the Any other, you know, preferences that you have that you've thought you commit to memory, but chatbt ignore chbt 5.

I'm I'm talking about the transition now. I'm talking about the transition between chatbt 4 to Chat GBT. No, nothing specific. Anything that you can say that didn't Okay. So, it would be it would be interesting if someone if someone out there has set some sort of preference in chatbt40 as they came as a chatb5 that has ignored those preferences because what you're saying is basically that um apparently it ignores them. And it's not a new memory. I'm just saying that even they Yeah, it's not me. It's the same old memory. But doesn't matter which model it is, it still refuses to do what I asked it to if not outright um you know make a mistake which has been the case like I've been using it all day today and yesterday like final finalizing one of my components.

Yeah. it it has so many mistake and I'm like why don't you understand my simple requests I'm sorry I'm just I'm focused on the transition not the performance of a model sorry yeah got you well let's uh let's keep going because um sure yeah it's the he's saying as well here that they uh they're 100% the rolled out to 100% of pro users and they're getting close to 100% all users and then we're going focus on some changes to chat GBT5 to make it warmer. Really good per user customization will take longer. So that's what they're sort of saying as well. There's some I don't know the specifics, but there are some partnerships um in cursor as well that they're they're going to be working on.

Um mention he doesn't mention that specifically in in in this thread, but yeah. So I think I think this is kind of ownership of you know not not going as smooth and they've done this a couple of times I think but um the other aspect is um Theo I felt really bad for him in this in this video because he's obviously taken a lot of beatins for it and what what he's essentially saying in this video is that he experienced a different chat GBT or GBT 5 than what he experienced on launch day, which is what I want to kind of focus on here. Um, which is really interesting. You know, you'd think that they would get a reaction and then stick with that version and launch it, but they they're obviously tweaking it probably until the last minute.

And this is something that got me thinking today and maybe we'll discuss it, but like I swear um AI goes through mood swings. And I don't mean mood swings as in grumpy and happy. I mean just some like yesterday for example I was doing a bit of coding and any change that I would try and make it would just try and clear all of these errors. Even if I'm saying do something in this file or do something with this function I ended up having to put do not do anything else in this prompt. Yes. And it was just it was just in a mood. Whereas today I've not had that issue. It's just done pretty much exactly what I want.

But it's anyway I'm going slightly off topic. The same thing is happening. Yeah, to me and I believe this is like to everybody literally like earlier today I told um I was using the app told it to not delete anything but just change five lines. It literally deleted everything and just replaced everything with five lines. And I'm like why don't you understand this? But going back to that video, there is one thing I don't agree with Theo. I saw this video um where he says he he doesn't believe that Chad GPT that open AAI gave him and other creators a better version and then they released a worse version. Uh he doesn't believe in that and I don't understand why.

Like I would not say why shouldn't like why shouldn't the this be the case? it would actually make sense for um for why this happened and also in terms of performance not not just ill like pure ill intent of open AI but what if they gave these creators an earlier version on I don't know how much this matters but like uh on infrastructure that is not so much limited or like they gave them more free like more time for the models to think or whatever I don't know how they are managing the performance but that could couldn't that be the case that they had a better version and then when they released it and everybody started using the models the models degra deg degraded because of the performance issues um I don't know whether AI's degrade in performance because of capacity I think they might degrade in terms of or in terms about performance in terms of speed because the GPUs are like working overtime, but it just takes longer.

Again, I don't know, but it seems like it would just take longer for the answer to come out, not necessarily the quality of that answer. So, that's my assumption. The other the other thing to consider is that he could have just been hyped. like he keeps stressing that ch and this is a really common maybe it's my personal belief but this is this is something that I don't believe people quite understand that when you're paid to advertise something um that mean that doesn't necessarily mean you're giving cold hard cash like if you're receiving something for free in order to make content or do something on it that is a form of payment and it's a and more specifically It's a contributor towards bias.

So, when I do see some creators say, "Look, they haven't paid me a scent. They just sent me this camera lens. Um, they're not don't get any say on the video, blah blah blah." By them sending you that camera lens is a slight bit of bias. So, I don't know whether he was just super hyped, you know, he was excited to be there. Um, he was genuinely really impressed with the model. I have no doubt about that. But ultimately it came down to uh was it like buyers remorse or something along those lines where it's like actually it wasn't as good as maybe maybe it was just a day thing. I don't know. But like regardless, you know, it's he he shared his opinion and this is what and this this is uh I don't know whether we were live before or after I mentioned this, but I actually wrote to him and I said like it's crazy that we live in a world where you just make videos and share your thoughts and opinions and then people get upset about it and you have to produce a sorry I should be sharing my screen really but you have to produce a video like this to apologize to these grown adults because they can't understand or they can't put down someone's opin opinion uh or disregard someone's opinion politely and whatever they just get upset and angry.

But um anyway, um point is he Theo experienced apparently a different model. I don't know whether it was whether he was hyped, whether it was bias or whatever, but you know, um but we go the other way. You know, you got McKay Wriggley here. Um he says it's a really good chat model. He really likes it. And again, maybe it's a personality thing. Um, API pricing is incredible, which it is, although again, I've seen I should probably have that tweet up, but someone mentioned that actually it's wasting a lot of tokens as well. But, you know, uh there was the efficiency of the token usage was um apparently a lot better. Um, I trusted in that, which is why I moved all my things, my a AI stuff over, but I found it to be too slow.

But that's neither here nor there. He's saying he's going to stick with clawed code. I'm going to stick with clawed code. I really like clawed code, but yeah, coding wise, it was just too faking slow. It was taking ages to do stuff and um, yeah, overall just wasn't I wasn't that impressed with code. So, all in all, just a mixed bag, I think. Sam Alolman realizes the roll out could have been a little bit better, maybe a bit like I I don't know. It's just kind of all over the place. You got different people saying different things. Um don't know. It's And it all just comes down to like, you know, just use what you think is good for you.

You know, all of these all of these AIs respond differently on different days. Maybe you're a bit more, you know, annoyed. So your descriptions are shorter or something like this. Like they're so undeterministic that I think it's just a case of getting your hands in the code, sorry, in the in the model and and using it and and seeing if you like it and if you don't like it. So um that's my spiel. Um, we will get into Kimmy K2, which but a little bit later on, which I've I've in in the roundabout way kind of been enjoying the last hour or so. Kim K2, but yeah, GBT5 could have gone a bit better. Um, anything anything more to add?

No, it's it's it's mid. Yeah, I I mean I I'm you know, I take it I would use Why is Notion so bad? Um I only thing to add really is um I've got it here is I know we don't we we hate benchmarks and blah blah blah but the the the fact of the matter is it is a ve it's scoring very very high in all of these benchmarks which again just goes to show um that it's the usage of the of the app of the why do I keep saying all this code stuff my head's going in app mode. using cursor way too much curs. Yeah. Yeah. Yeah. So there's no doubt about these numbers and the score and I think I saw something that actually won a coding challenge an AI coding challenge I I'll add.

So it is no doubt you know scoring very very high points in all of these things beating out some of the best models out there. long reasoning, agentic capabilities, all of this. Vibe checks, which is an interesting one. But that doesn't necessarily equate to a good model. So that's that's it. Yeah. Okay. We have got something super exciting which unfortunately I wish I had more demo to talk to you about, but Oh, actually I'll go to should we go back. You want to sh uh uh share my screen is it? No, you aren't. But I can show it. What's that? What are you What are you saying, boy? No, now you are. Cool. Um is that chat uh Chrome have actually released a language model.

You can see uh Add's saying here you've got a language model inside built into Chrome and it's called Gemini Nano. and it's pretty cool. So, he's just writing a prompt here and he's uh he's console logging the prompt and it's pretty pretty cool. Um it does uh image understanding, it does speech transcription as well like um voice. It's got text summarization, the usual stuff you'd uh come to expect from from using a a model. Works across the board obviously because it's built into Chrome. Um, Akut actually said about this like on one of our previous episodes that they're they're building it directly into the uh the Chrome, but I think this is a statement of that it's actually usable.

Again, I'm sorry. I tried. The documentation is pretty rubbish, but I'm guessing you have to be in some you have to be in some sort of group or something or other. But um yeah, uses 4 gig of RAM as well, which is really cool. Very small. Obviously, we talk about 16 gig and all the rest of it. Yeah, Chrome just taking 4 GB more. It's nothing what they're already taking. It's just a tiny bit more. Yeah. Um the only thing is I'm uncertain about and they've actually flagged it here. Let's go here. is it takes 22 gig of free space and I don't know whether that's the model but 22 gig of and then they have this little built-in model should be significantly smaller so I'm not quite sure what this is tried playing with it I tried using it downloading it or doing something I I was you know wrestling around with it but I couldn't couldn't figure it out you should be able to just run this and it will download the model and then you set up.

Oh, not that one. Sorry. So, wait a second. So, this is a model that you download on your machine and it is a part it just runs inside of Chrome. So, this is an API that your website could use. Yeah, this is like super cool. Uh, and is this the end of alt text for images? Well, why why would you say that? Like if your browser already can analyze images, why do you need to write all text? Like the all text can be generated if it's empty by the model that is running in your Chrome. Mhm. So I would say no because two for two reasons. One for permissions like it's not just going to you know uh in it's not just going to like run AI on your machine annoyingly.

I think we're far far away from AI running like behind the scenes 247. Yeah. Yeah. Yeah. Because people are too sensitive even though it's already been happening on pretty much everyone's Apple device. The other aspect is is that the website itself might not have like implemented the call to that because it's an external API. It needs to be the website that takes charge of doing that. So definitely not yet. But like I yeah I think that the probably who people who need that feature probably have some sort of plug-in don't they? I don't know. So here's a question like what are the benefits of having such a small model? It is I understand it's quite good for the size but it's not comparable to you know bigger models.

And if you are inside of Google Chrome, you are connected to the internet. So why not, you know, call an external API and just use a much bigger, more performant model than a smaller one that you know takes uh the speed of your machine. Um I think privacy again it's privacy it's uh it's speed is going to be a lot faster because it's a smaller model and it uses less like resource. I mean they must have they must have done comparisons and the trade-off of having a local model you know with not not the world's knowledge necessarily but what is deemed pretty uh universally you know so maybe yeah maybe yeah maybe because like a lot of these cool projects are just cool projects people do it without like comparing it or have a real world you know example or use case they just do it because they are passionate about and and we see if how well they get implemented.

Google has done hundreds of thousands of these projects. But it's it's really cool. I am actually looking forward to it. But I just don't know like what is the actual use case over another model that but you mentioned privacy. So sure. Yeah. Yeah. And like I say it's an API that websites can use you know just like you can use you know scroll add event listener and blah blah blah. So, uh, we we have, you know, because obviously you got all of these Chromium forks, DIA, Comet, pretty much all of the AI browsers are Chromium based and they're built on top of it. This could potentially provide a foundation for all of those brand all of those forks to to leverage off.

So, it's pretty Yeah, it'll be interesting to see what the community does with it, how clever ways they can utilize it. So, I'm I'm down for it. You know, I'm really getting into the local stuff right now. I'm even I've had two conversations recently about setting up a completely localized uh AI system and I might even buy a nice PC or something and have just local model because privacy is a thing, you know. Yeah. At the price of that, uh you probably won't be able to travel as much. No, exactly. Travel and access your own PC at home. That's Well, that's it, right? you I'd be I'd be tunneling in I'm trying to figure all that out.

First of all, my local network, then the world. Exactly. So, for the next one, uh we should stick to the topic Google Chrome uh and perplexity trying to buy it. I find it I mean, should we should we run the ads and then we'll get back to Sure. We'll jump right into that. So, we'll have an ad and continue talking about Google Chrome and maybe Perplexity. No, complexity the other way around. Yeah, not it's not mine. Support for this episode comes from Flowbase. If you are building in Web Flow, Framer, or Figma, Flowbase can help you build much faster with over 4,000 components to choose from. You have a huge variety of wireframes and super clean, nicely designed sections that you can put together by just simply copy pasting them to your project.

over 15,000 icons as well, over a,000 illustrations. They also have a super helpful Web Flow app called Boosters. With Boosters, you can add things like sliders and mares and countups to your project with a few clicks and without any coding. So yeah, big thanks to Flowbase for supporting the show. If you want to try them out, go head to flowbase.co. And just for you guys, I want to show discount command flowbase all capitals for 20% off any plan. This is a limited offer, so get in there early. All right, that's it. Let's continue with the show. Here we go. Yeah. So, Google Chrome. Uh, do you want to buy it? Should we get hs? Well, this is not a huge news, but this is more of a funny one because Perplexity is worth probably around 14 billion somewhere in the article says and they are trying to buy a company or like a yeah a company essentially that is worth twice as much as their own worth.

Like how does that work? So apparently the like Perplexities investors would pay for that. So it's basically yeah I don't know how all of that would work exactly in um in action but yeah they are not such a big company and trying to big uh to buy a much bigger one. And this is coming out after they they released Comet their own AI browser. Yeah. Yeah. Very strange. Very very strange. I mean, I think they would just ditch. They would just I don't know. What are they actually buying? Yeah. Chrome. Well, the users like I'm sure Google own the users. They just they just log in. It's like Yeah. So, what are they buying? because Chromium is open- source and Google Chrome is, you know, not one of the best browsers.

Chromium is amazing, but like Google Chrome is just meh. Yeah. So, I I don't know what they exactly they are buying. This is similar to the situation we had with OpenAI trying to buy uh what was it again? Windsurf and and we were also confused why they were buying that. Um, it says here as well, the company is not has not offered Chrome for sale and plans to appeal the US court ruling that found it held unlawful monopol an unlaw lawful monopoly in online search. So, I heard about this, but it doesn't necessarily mean it's for sales. So, what would like are they just throwing money out there? probably because the US government try is trying to split Google Chrome from Google because they own a monopoly and they are paying what 20 billion I believe a year to Apple 12 or 20 billion something like a I think it's 20 it's like big number just so that Apple uses Google uh as their default search engine and yeah things like this and they want to, you know, uh, split it in pieces.

Google Chrome being one of them. Uh, yeah. So, Perplexi does not disclose how it's going to fund the offer. Seems very strange. That's why strange you could offer. You could you could make an counter offer like you don't have to disclose how you are buying it. How very interesting. How very interesting. again like you say already has comet and it's pretty good. I'm I'm What browser you using? You still on Dia? Yes, I am not 100% transition to Dia, but like I use it like pretty much like 95% of the time and the other 5% is split between other browsers where I need them. DIA is not 100% what is it like reliable for me yet and the Google meet screen sharing stops like it just doesn't work.

So whenever I have a meeting with my students or with my clients I actually use Safari because it just works. Yeah. Um I've had no issues on Comet. No issues. Maybe I should try Comet. Um, yeah. Um, I'm not I'm not really I'm using it. Yeah, I'm using it to chat to pages. I'm using it to chat to pages, but I'm not really actioning things like I thought I would. I would thought I would be doing say take something from this page, put it in that page, this and that. I think we got burnt very early on trying to do something in Web Flow and just wouldn't work. And it was just like I might as well just do it myself by this point because it was just trying to do stuff and like Yeah.

You know, I'm doing the same. Yeah. Uh I'm not I'm not using the AI as much as I thought I would and I'm just going back to chat GPT app because it's a separate window for if I want to focus on one work which I like to work in that way. Enjoying multitasking all that much. I focus on one thing um and just using chat GPT and then do the the thing manually myself. M um you know we get into the discussion on productivity and working smart and not hard blah blah blah all this kind of stuff and it's like if I wanted to like burn myself burn myself out I would probably set up a task on perplexity to do this that and the other leave it do its thing then go on to another task and then do you know if I was an octopus in my brain I would do something like that but to be honest like I don't I don't want to be that I don't want to have various different things flying off all the time.

I actually don't mind just setting an AI off and then just chilling while it does its thing. Like, yeah, you know, you're I'm already hyper productive just doing that. I'm already doing hours worth of work in minutes. So, why would I, you know, we spoke to um Max from Warmwind who said this, their system is more for background task. And I still think the AI inside of comet is for back background task. If you are using comet, you are a heavy AI browser user. I would love to know what you use it for and and how how effective do you find it really? All this AI. So what do you do while your agent in uh cursor is trying to write the code?

Just chilling, baby. Just chilling. Maybe I'm thinking, maybe I'm just like, okay, while it's doing my thing, like what what do I want to do? And I'll make like a little list. maybe and the prompt because you can cue prompts sometimes. Um well certainly in replet you can cue prompts just trying to like think through user journeys. What do I need to do and stuff? I I use that as thinking space because oftent times we're building an app like I'm not Yeah, I'm building I'm building and creating apps. I'm not like I'm not I haven't just got like a list of things to do. I'm I'm thinking like three or four steps ahead, you know?

So yeah. Uh what are you doing? grabbing a drink, touching grass, grabbing a drink, go touching grass. You love touching grass. I I also mostly think about the next step. So, if it's either doing something manual or um just thinking about the prompt for the next one, assuming that the current prompt is going to work and you know do the thing correctly, then I think of how I should prompt the next one. Yeah. M this another one of your one you want to take us away with this one. Yeah, this one is insane to me and I don't know why the coverage around this they don't find it too crazy. So you tell me uh Nvidia and AMD are pay are going to pay 15% of the China chip sales to the US government.

So the US government in a way is is it like blackmailing them? I it it is crazy to me that the government can do this because in my head like if there is a so the governments are based on rules and these rules are based on principles. Um, in this case, in this specific case, we had Nvidia and AMD and other, you know, US companies being banned from selling their chips, advanced chips to China. Why? Because they don't want China to get access to these like highly advanced technologies. They don't want them to get ahead. And that the reason that they say why they are doing this is because of security reasons. So that is a principle that is like okay based on that we we have our own worldview we want our um we we want to keep everything safe.

We don't want China to have access to these advanced chips and now suddenly it's fine if you pay us 15% of your sales. That is just rubbery. That's that's not working on principle. This is just essentially blackmailing them. This is insane. The government I don't know whether I' I don't know whether I'd class it as blackmail. I would Yeah. There's a different word for it. I think it's just it's just I'm I'm thinking the same thing. It's like, oh, privacy, priv privacy, and then, oh, just pay us some money. We're under a bridge. No problem. No problems. We'll be fine. It's never about privacy. It's just lying. It's about money. It's all about money. And that's Donald Trump in a nutshell.

And we we've known we we should know this. We've known it all along. He's just a businessman all out to get money. That's all he wants. And this is just really weird behavior. Really weird. And and China, I saw China's comment. China was like, what did they say? Um where is it? something along the and it was only Nvidia who said it was something along the lines of we we just we'll do whatever we need to do to take part in the world market. Yeah. But something along those lines. It's like we follow the rules the US government sets for our participation in the worldwide markets. It's like Yeah. But do you not see that that's like these are like these are not rules.

These are made up. I these are completely made up because it's not based on anything. Imagine you either have a national security problem or you don't like it. Yeah. Like imagine you not being able to do business with a specific country and you'd be like okay sure because it's based on whatever you know and then the government says well actually if you pay us money on top of taxes or whatever then it then it's fine. It's kind of like, you know, I don't know the ins and outs of it and maybe someone can fill us in here, but like it's like a US citizen going abroad and getting citizenship elsewhere. They still have to pay the US taxes.

They still have to pay US taxes even if they've moved to a different country and you know, whatever. Whereas I if I get citizenship somewhere else, I'm done with the UK. Like, oh, it's the same in Germany. We have to pay. We have to pay like the US. we have. So if I if I go live somewhere else because I have a kid here in Germany, they will actually make me still pay taxes because they can say, "Well, you still have a life here. You have family or whatever." You can um skip it. if you can actually prove that you have nothing to do with a with this country and you don't live for like uh you don't come even to visit more than I I don't know how many months but some months there's something similar for the UK there's something similar I think it comes down to what are taxes for right they're for they're for public services they're for but if you're not in the country utilizing those services and and you know apologies for the lack of word but you're not like a drain on societ society.

I know it's the wrong word to use, but like with the US, it doesn't matter. You could be you could be fully like, you know, in a complete different country, different different passport or whatever, new passport, but you always always pay US taxes. It's just ridiculous. It's just a I don't want to say they're like a lovely country, but yeah, it's just the workarounds are like the shell companies and the offshore accounts that you know these billionaires and millionaires do. So even though they are in the US, their money is not in the US or whatever. So wild. Absolutely wild. This is wild. Um, let's move on to Claude Sonnet very graciously giving us a million context window.

Million token context window, which is a 5x increase. Process 75,000 lines of code or hundreds of documents in a single request, which is pretty cool. I'll take it. My my comment to this is that context size is a little bit overrated, but I'll take it. um you know being aware of what they call uh what's it man in the middle or whatever that you you kind of lose things in the middle of when it's a large context and all this that and the other but again if you know how to use it take it um tell me what do you think of um context window like how important it is to have whatever like 1 million context window and does it actually like benefit you in coding if you go from 100K to a million.

Well, I I well, as I say, my comment here, context size is overrated, but I'll take it. Like, oh, no, I'm not even sharing my screen. Boom. There's the man of the hour. And I'm very much someone who does not I'm I'm I'm much more meticulous with my coding. I'm like here's the file that I want to here or here the files that I want to for you to look into. I don't I don't necessarily give it broad you know whatever. And that's my way that's my style of coding. So I'm not all about context size. I've never run out of context size. I'm always like when I'm done with my task, new chat, compress compact.

Um, so I'm not that fussed on having a million. I, you know, my experience with Gemini Pro isn't actually that great. It's got a massive context window. Maybe that's something to do. I don't know. But you're you're more encouraged you're you're more encouraged to be less meticulous with what you're trying to do and you're greedy. You you you'll use that million context. But saying that, I know there are tons of people who work on much bigger code bases than I do, working on much bigger, more important products than I do. So, for me personally, take or leave it, I'm not that fussed on million contacts because I was never hitting the 64,000 or whatever it was before or 200,000 it was before.

Because saying that, it's not even available for us mere mortals yet. It's only available for uh API users with tier 4 and custom rate limits. So, it's actually for the privileged few that are allowed to uh to use it right now. broader availability will be rolling out over the couple of weeks. I had a little look on open router still says 200,000 tokens context window. So I haven't had a chance to enjoy it yet but it is you know it's a welcome change. Yeah large because saying that as well one thing I do do is I do this a lot. whack on Claude Opus 4.1 pay all the money that I need to pay when I know I'm making a very big very important change and I won't get it to edit any code I will say give me a plan before changing any of the code and then I'll implement it with Sonnet or again depend or depending on the um extent of the of the of the change I might implement ment it with opus but it's I just you know that's just my style baby that's just my style I'm not right I'm not wrong I'm just I'm just Sam so yeah uh the but but what I mean by that is that is a change where I'm not being meticulous with specific files I'm like look at this code base because there might be things that I'm not thinking of or planning for or something like that so please go ahead and look at it look at it So, but again my code bases using cord opus have not hit that context limit.

So, um there is a lie or there is a change. Sorry. Under 200,000 we're still sticking with the same price plan. If you're over 200,000 it's double. So, you are definitely paying for that increased context window. So, it's just something to something to take into account. So, yeah. very like again I'm not going to complain. I I mean you know it's there. Oh and it's just Sonet as well which is important to say it's just Sonet not not Opus. So um yeah I dig it. I like it. Sure. You aren't actually I keep forgetting what model do you use? You just you just use whatever mix. I know I make I I Yeah, I No, actually I I do choose different models to see if I get better results, but like my experience hasn't been so amazing with Claude.

Not much more different from any other model. Again, you're using it through cursor though, right? Yes, I go to I I use all of the models through cursor. And as I mentioned, I think it was like one or two weeks ago, I made a very simple React app and I showed it to my uh developer and he had a stroke. How bad the code quality was even though it was a very very simple React app. Apparently the models that I was and I did it twice by the way and it was just trash apparently like the the code quality the using like uh depreciated syntax or like Yeah. Yeah. Uh yeah, it was apparently not and and it didn't work the way I wanted to.

And this is Claude. I use Claude, Gemini, and OpenAI. Yeah. What as a mixture of, you know, jumping around creating the working on the same code bases. Yeah. I don't know which one I started with, but I did actually struggle quite a lot like hours and hours and didn't get it even close to what I wanted to do. Uh, and it was things with shaders. Um, and a bit more. I think I reckon you do better with something like replet or lovable where there's just, you know, you're you're riding a bike with the with training wheels off, you know, with Yeah. I I don't want to I couldn't because I I needed like I can't disclose what I was working on because of NDAs, but it was like specifically like uh components that I couldn't like have within another uh platform.

It had to be local and um yeah try try them out. So I did try and didn't work at all. But even from like a project setup standpoint, I'm thinking like just setting up a project that has the kind of like the foundations and the structure of like just a an app and uh and then when you're in cursor sticking to your files, which again you you you're riding the bike with the training wheels off because you need to know like I only want you to work in that file and I know you only need to look into that file. So you know Yeah. Um yeah, it's a it's a long and long and treacherous road you're on.

Um rocking cursor with the training wheels off. Yeah, but it's fun. I'm I'm you know, learning um the coding by doing these like things that are way too complex for my level, but I'm still tackling them with AI. Jake has asked, what do we think is the best AI for writing? um especially for job searches. What kind of job for a start? Um when it comes to writing, what do you use? So I always go to Claude, but nothing more than because a marketer, my friend Kieran, says that he he uses it for writing and I've heard it's good for writing, design jobs. Uh that's what that's what Jake said. Um, and what kind of writing are you doing?

Like what what do you mean but when it comes to writing are you are you talking about literally like job applications because my mate has had very good results with chat GPT and applying for jobs through that. Um, but yeah, it is more so Claud Claude and Open AI. Um, I'm all over to be honest, Jake. I'm all over like if it if it needs a lot of research and it needs a lot of up-to-date data, I'll use perplexity. But the good thing about perplexity is because perplexity is a layer on top of gro 2.5 claude. So, you know, you can actually choose your base model and then perplexity is doing its magic over the top of it.

But that's when I need research and input and all the rest of it. So, that's like blogs and social media. I'll use perplexity. I'd be tempted to try Grock, you know, for writing. I I should try that as well. Uh I use just chat GPT and the app because of canvas, because of how easy it is to go into some lines and change it. Even though it gets frustrating when uh I ask it to to tell me something without changing the content inside of the canvas and it just doesn't listen and rewrite the entire thing and then I have to undo things and you know redo things again and again. I just use CH GPT because of the canvas.

Uh and I'm not really happy with the writing style. Uh, honestly, I don't get it to sound the way I want. I'm very particular when it comes to to writing. I have I I want to write like a human. I don't want to sound overly complex or, you know, the this the simple M dashes that they use. I don't want to use any of those. I want to write the way that I want. So, I'm actually defaulting more and more to just using grammar check. So I just in DIA as well I have a grammar uh skill I think it's it's called that I just do I I write something and I just say slash grammar and it just checks the grammar and rewrites the grammar without without changing the tone and everything else.

Yeah. And that's another thing as well like he's Jake just said thanks but um the tool itself plays a part in it as well because perplexity has spaces you know where you can create a new thing in the context of the space. So you can set a lot of you could potentially have a cover letter writer which it says write in this tonality and does it have web search and what tools can it call and stuff like this. So you create your own writing person. I mean chat GBT has custom GBTs, right? And things like that. So I think the tool but I don't think I don't think Claude has anything. Does it? I don't know.

Does it have something true? But you I mean you got projects you got projects describe what you're trying to achieve and stuff like that. But yeah, I think the tool will matter as well more so than necessarily the model and what what it can how it can, you know, create a taking in a set of parameters whether it's got file input and all the rest of it. So yeah, not much of a writer though to be honest. I I do social uh perplexity and that's kind of it. Um let's look into Miyagi. Yeah, which Bradford Huba um it's been in sort of beta for like a little while and I think Miyagi and now it's live.

I think he collaborated with the guy who did XAM as well. Yes. um and on this which is really cool and essentially it is a vibe code plug-in for web flow which is really really cool and you can choose different models and this was my first experience with Kimmy K2 and got to say I was pretty I was pretty impressed. Let me see if I can I'll stop sharing my screen a sec. And so what does it do when it uh write code? So you open the app inside of web flow and then you ask it to make a component. Right. Right. Does it also create the HTML inside of Webflow? Yeah, it creates the HTML.

It creates it does it does whatever. Basically, it's from my understanding, it's basically an AI agent that interacts with the Web Flow MCP. Ah, okay. So, can you tell it to create uh a slick slider using CMS collection? That would be like the next level. Let's play because I'm just logging into Facebook now. If all of my Facebook now, do you know what? I've not used Facebook in ages. I had no idea why I said Facebook. It's funny that I'm um you know, shaming you for using Facebook. So you can see here I'll create this beautiful website with amazing. There you go. There you go. Look at that. Um and obviously it looked a bit better.

It wasn't much better to be honest. I was just playing around of what it can actually do. And one thing it couldn't do here, if I just open up um give me design mode. Design here. Okay. If I bring up Miyagi launch. So you can do two things. H come on. Let me um do this so I don't have to I have access as well guys but I haven't gotten the chance to use it but I plan on doing it. Um actually my co-founder had also this idea um a while back to build to build this but I didn't it's very well done. Yeah, I didn't want to tackle this because I don't have I want to be like very clear about I don't have any uh internal information from Web Flow saying that they do this, but I would easily imagine that they would do something very similar inside of Web Flow.

Mhm. Wouldn't wouldn't it be wouldn't it make sense for every tool every single tool to have an app like this that interacts with the tool that can do at least like the basic stuff. Um and then you take it from there. H um potentially potentially. Well, as with most things when you try and do a live demo, it's not friaking working. But I have a feeling it's more my end than Miyagi's end to be honest. Nothing to do with Miyagi. But there's two modes. Maybe I can we can potentially work through the screen share um the video here. So there's there's a design mode which you can create elements you can you know change styles things like this let me find generate code deploy code I want to here we go this is what I was looking for so you can apply you know this is this is you um choosing the context and you can say just the element just the element element with styles or include the element and its children and then whatever you type is exactly what it could do.

So I really want this to work because I really want to do like a live thing of it. Launch. Uh see it's just keeps asking me to log in. So really really really annoying. Frustrates me. Um where are we here? And yeah you can ask it to do a bunch of stuff. What I found really interesting is that you can uh ask it to create Gap animations and it will create like a Gap animation which by the way you do need to copy and paste into the code. So it's seemingly not interacting with the web flow interactions panel. It's literally just giving you a coded animation to do which is you know it's all right.

You can you can choose between Kim K2 which was really really fast. Really fast actually. Okay. I didn't try 2.5 Pro but and 2.5 flash. I didn't I didn't try those but those are the models that you can choose. You can see it's got communicate 2 just just in the screenshot here. You can do anything that you it basically leverages the web flow API. So whatever you can do or whatever you need to do or sorry whatever you can do in the web flow API you can do through Miyagi and I get that's why I mean it's just a a wrapper around the M MCP to then interact with the web flow API. So really cool change of style.

I managed to, you know, on my my website here, even though it looks like complete, you know, chaos, this was me adding uh variables. I told it to choose a p choose a color palette and add them as variables and this and that. That's why I was just, you know, just me really really messing around with it. I told to make an animation for this this intro thing, but I haven't got installed, so that's why it looks like it does. Uh, but you don't have to install Gap. Yeah, I do. I do. Uh, I don't have interactions yet. So, uh, presumably if it if you got interactions installed, it would just work, but I just don't.

Uh, can you go to your page settings? I can't be asked to do that. Do do go through all that now. Um, and what else we got? Um, and it's got this quite nice thing as well, which I found they've got a like nice prompt library. So, this might even give you a bit of an insight about what it can do. So, they've got here add alt text to all images. So, this can't work with Kim K2 because it doesn't have image recognition, but with with Gemini, you can do image recognition, which is really cool. Organize all of your media, fix insecure links, which is really cool. And they've got this weird I mean, I do this sometimes, actually.

I wrap my things in HTML elements like this. But yeah, export all the classes, find and replace HTML attributes. Really cool. So, it's got direct in access to the HTML code. So, when you say can it do like a slick slider, I have no doubt it can do a slick slider and it will work through, you know, what the HTML that it needs and then gives you the code. And in fact, I tried to do an unlimited thing for this. it didn't work. But it also but it gave me the CSS that I needed to copy into the the the page. The other thing is though is that I think it's got a little secret weapon here which is basically I mean I can't really show it.

Let's try let's try logging in again. Um or at least opening up my login panel. It's basically got Slater built in. Oh, and this is what they call code mode. So, it's got two modes. You got design mode and it's got code mode. See, it just does this and whatever. And in code mode, you can vibe code or you can just create files, JavaScript files, CSS files, and then through the options, you can choose which pages those JavaScript files and CSS files actually load on, whether they load on all pages or specific files or specific pages. So, it's really cool how you get basically uh Slater built into Miyagi. Really, really cool actually. Do you have to log in via Slater or do you have to like No, no, no, no.

What I mean is it's it's Slater functionality. Ah, so it's not like actually I see. Yeah. So, uh it starts at 49 bucks, so it's not cheap, but like I say again, you get Slater really, really fast, really effective. I was really impressed with actually how well it was um well it how well it worked and I've spoken to Bradford and he's kind enough to just share this code for our listeners. If you basically sign up for the app just like normal you'll get a little bad banner notification to say we're giving away tokens and if you put launch party um all caps hyphenated you get that you get 300 credits or 500 credits here uh for free.

So you get to try it out. So it's quite a nice little nice little gift he's given us. By the way, the pricing the pricing starts at zero. Like you can just have it you can use it for free. Yeah. Yeah. 30 credits um for Yeah. 30 credits I believe per month or um $19 for 150 credits. Yeah. So there we go. And then it goes up from there. Yeah. So this is the plan you get for free if you use the code. Exactly. So, pretty cool. Um, I haven't tried some serious like design prompts yet. I haven't had I would have done it on on the call, but like I say, it's just it's struggling to connect.

Uh, it might just be because I'm using a Firefox browser. That's probably it to be honest. Uh maybe I can quickly do something. It's super early as well. What I don't um understand is like design versus coding agent. So if I want to do a slider, I need to to click coding agent. Right. Let's share this screen. Um this screen invite share screen window. Choose window. Boom. Boom. Boom. Right. So, this is a Chrome. This is this is comet now. So, let's take back control here. Oh, I need to accept the request. Fact. Let's just close this window down. Let's Let's play Let's play with it. Oh, come on. Let me just log. Uh, I got it open on my on in a Web Flow project.

Okay, if you get it up. Yeah, the the code version. I'll share my screen if I get anything interesting. I said, "Create a slick slider with five slides, each taking about 30% of the screen." Uh, I know this is like a not a really good prompt, far from it, but something quick. Uh, and I assume a lot of people would prompt it like this. Um, so if you just look at my screen right now. Yeah. So, you've got the design agent. This is and you got CMS agent which is coming soon. Yeah. So, that's going to be interesting to be able to interact directly with the CMS. Uh design agent is as you'd expect, you're interacting with the the actual canvas itself, right?

You select this and then you can add the context. You can either, you know, apply it to the current page. You choose a page and vibe out a section from this, you know, from from a different page or whatever. Selected elements. Um, and then yeah, so if say you've got an empty div, or you could just say, what is it? Current page. Uh, add, uh, slick slider to the page and see what it does. It's pretty quick. Look. Uh, let me add a sexy slick. I don't know why, but I was like, create a sexy color scheme, and now all of a sudden it tries to make everything sexy. um failed to create elements from canvas.

It seems like something going on there, but it's still chugging through. Look, you know, and it might give me some JavaScript or it might give me some Yeah. script to add to the actual page. But current action take is too long. Please wait or cancel. It's probably we're we're probably sending it all the users through the roof now with that free launch party code. How giving them loads of traffic. That's why it's so slow. Here we go. Gap animation. And this I mean I'm not expecting to oneshot this, by the way. Don't I'm not one of those mugs who just think, "Oh, it's just going to work." Like, I'm just We'll no doubt have to tweak this.

But let's just see how it go. Let's just see what we end up with. But it's quite nice. You know, who who do you think this is for? Who do you think who do you think would work would benefit from this? There you go. Look. Um, it's hard to say like beginners because if you don't sometimes it is really nice just to like apply a bunch of styles with a prompt like doesn't necessarily mean beginners like just say just put the put the border radius and background and blah blah blah and it'll just do it. You know that that's actually nice. So for like big design systems, I'm not sure if it can, for example, add a component to all pages because this is something that Web Flow should really have by default, but they don't.

But with an app, you theoretically could do it. So that is imagine you have a website with 100 or 150 pages and then the client asks you to add a new CTA and then you have to go to every single page manually and paste that component. Yeah, imagine that. That would be that would be mental. Or you haven't created a component, you could say find all instances of this button, replace it with this button and make that into a component or something like that, you know, something like that. But yeah, like it is it, you know, prompting is a nice interface. The other um thing I wanted to show you real quick then is just how the coding agent works.

Now, this again is the replacement for Slater if it loads. But does does it like uh publish the code to like their environment and then load it from there? Um I believe so. I believe so. Yeah, because Web Flow can't store files, right? Well, techn Yeah. Not not currently. And you could store them as text, by the way. That's how some people do it. Mhm. And then you load that. Yeah. Yeah. Um, so yeah, here we got some these are my chat sessions. Um, oh, they've changed this. This looks different. Browse files. So, so here we go. Here's some CSS, some JavaScript. You have to copy and paste like a line of I think this is how you, you know, publish it and whatever.

Is that what it is? Script connection. Can you So, it's this here. So, if we look here, okay, files get mommy ugly. Yeah, they're they're storing the files on their server and they're alternating between staging. They're doing the automatic alternation between staging and um production or you can have that control yourself. So, yeah, I see. What I don't love is uh that I have to do this script connection manually. Why can't the app do this for me? Yeah, because it can't add code in yet. I don't know whether that's a limitation of web flow like you can't add code to a website through the web flow API. Pretty sure you can it when you first load up Miyagi it will take you through a thing and that thing says just Yeah.

Where is the What did I just see? I saw something. Here we go. It might even say something here about that limitation. I see. And it might it might, you know, it might be to do with it adding it to the head and body, you know, the site settings. It might be able to add it on a page or it might be able to add it as an embed element on a page or something like that, but it can't do it normally in the head. But maybe that will change. Maybe it won't change. But yeah, overall pretty impressed. Pretty cool. pretty cool to um really clean like they've they've done very well in the launch like everything's clean, very professional.

They've they've done very well. I think it's a fair play to them. Yeah, it does look really good. And I really hope this um like gets better because I I want to actually use this. I want to do like broad design changes and um the CMS one also is nice. like let's say you tell it to go to a bunch of CMS items and update something. That would be really nice. Uh because sometimes clients ask for things like this that are Yeah. Uh I I want this app to do it for me. Well, that's it. Like I think you know AI in general at the very least should take away long boring tasks. Like I know we want it to create I know we want it to do like more or enhance us but at the base level it should do the do the rubbish crap that we we don't want to do you know that we can do but we don't want to do that's I think that's where everyone should exist with at the very least with AI least do your dishes.

Yeah. Yeah. That's the hope that we can, you know, do the fun creative stuff that while the AI does the dishes and the laundry, but so far it has been the other way around. AI has been creating like movies and short videos and images while we do the laundry. I've been thinking about this and of course you're going to get all that because everyone's having fun and playing with it. Like a lot of people just are hating on it and it's like this. I've got a video coming out about my thoughts on all this, but um let people have fun. Let people create let you know. Do you watch AI slop videos on your YouTube feed?

No. Yeah, I don't either. Like I I mean I I've always said I watch the voiceover stuff, but I don't actively sit there and like watch any I've never watched an AI create. But um doesn't matter. Again, you know, I think I um I sent another tweet out earlier on in the week. It's like let everyone do what they want to do. Like I'm really I I'm flipping the script on my perspective on the world right now, which is just let everyone do their own thing. Like let me just stay in my lane, have fun, enjoy my life. Yes, there's going to be destruction and crap around me, but who am I to get upset with all that?

Just just let them do what they want to do. Yes, there's just ridiculous stupid crap on the internet and people just doing stuff that I just do not understand. But you know what? Just let it go. Well, I understand your perspective, but that's fine as long as it doesn't affect you. So, spamming does affect you. Like I get emails from um people who are not real people. These are like just AI bots sending me these emails that look like they are from people and I see the m dashes and I see the the way that the writing style and I see that they they haven't seen my stuff and they act as they have and I just discard them.

I just go next. I do not interact with them, delete them. And if somebody if I'm in a bad mood, I might even set them as spam. But you know what? Somebody reached out to me with a Loom video without his face, but he was showing my YouTube channel. And you got that? And he had like specifics about it. And I I was like, "Okay, I'm going to actually book a call with you." Oh, you actually did it, did you? No, I haven't yet. But because I couldn't book the call too far in uh in the future, I had to wait. But uh I'm going to It's a young guy like um showing that they can reach millions of people.

Of course. Of course they can. Of course they can. I have my own thoughts on those types of people anyway. But um yeah, I I get that. But that's not the point. The service is not the point. The outreach is the point to me is that that human touch. Yeah. That I know they're not sending this to two billion people. They are send this guy has seen my stuff. So he he knows what he's talking about. At least when it comes to to this contact. It's not an AI tool just, you know, randomly sending a generic thing. That's my issue. What I find with those, maybe we're going off script now, but what I find with those is that they and I but I I went through the process with someone and they they do all the technical stuff.

They tell you all the the baseline technical stuff that you've known and heard about and yes, they'll they'll whip your ass into gear to get you to do that. Create a video on your profile page. have uh you know um ways to contact you in your on your you know again your profile page and you know do this type of content or whatever you know but what well not even do this type of content and that's kind of where I'm going with it. What they kind of lack is a real understanding of your niche to actually help you drive through with your niche and what you want to do. They again it just goes back to technical SEO.

It's like, yes, you could apply, you know, this is just a oneanddone thing. Maybe a bit of a plan, you know, to to make videos once a week or whatever, but they don't really work with you on a strategic level to identify who your niche is, what sort of stuff you're talking about. It's very surface level. And I guess you can get that with them, but that's that's where their advice. Yeah. But that that's about the service. And I'm yeah I I was more commenting on the outreach the same question because yeah you know I know how to reach a lot of people with like short at least not millions though not millions. Well I have like I have two videos uh with millions of views and one with now currently at 800,000 the last one.

Mhm. But no, my YouTube videos, they are not at millions yet. But you but you know what you can do if these emails come through and you see those end dashes and you know it gets you riled up. No, I did I just discard them. I just Well, practice stoicism and the thing you can control is your own emotions. Like just Yeah. Just bat them away. Yep. And When I when I say if I'm in a bad mood, I send them on spam. That is because I don't want to get more of them. So angry. No, I I don't actually. I, you know, I sip on my drink and then I just continue coding. Yeah, I'm I'm I'm I'm on a very vibing out kind of like mentality recently.

Like again, I felt really sorry for uh Theo in his latest video. He's clearly really cut up. He's clearly really struggling right now with a lot of hatred from people and it's like these are just nobodyies with nothing better to do. The loud minority, you know, that comes with masses. So the moment you um you let like more people see you or like did that happen? The circle grows and the haters come and like uh on my last video that I'm just showcasing a website, there are haters. like I'm just showing you a a cool website. If you don't find it cool, sure it's not cool to you, then just scroll. Just just scroll away.

Just don't. And it's simple math. Well, again, yeah, just let them do it, you know? Let them do it. Like the the it's just maths, right? You look at the number of views you got. You look at the number of likes you got. You compare that to the number of dislikes you got and compare that to the number of hate comments. It's so easy to hate. It's so easy to hate. It's much harder to spread positivity than it is to hate. So, as you just do the maths and you say, "Well, there is 10 times more, probably a hundred times more people who have not done that than there are who have left the negative comment."

Because most people most people just silently enjoy your content and silently just give you a silently give you a thumbs up on your content because they I I barely ever interact with a video. I barely but it doesn't mean I don't really enjoy videos. It's just you know but anyway some people make videos in a way that are they uh bait you for comment by the way. So there are videos that are made don't mispronounce things. Yeah. Well, these are funny like the there is this guy he makes uh intern like website videos and in every video he has fans at this point. His fans look at the video in their view. They don't care about the main content of the video.

They care about the things that this guy puts as what's it called? Like Easter eggs. So the Easter eggs are he shows a website but that there is a tab open another tab like how to hide a body. Not not always like this but sometimes it's just something outrageously funny or he mispronounces somebody or he mislabels somebody you know showing Albert Einstein and saying neotone or something like this. So yeah, he does these kind of things and um well that can be a that can be fun. That's my point. It can be fun. It can be uh without causing harm. It can be just cheeky, whatever. I I don't mind those actually. Um but there are videos people who make videos that are designed to make you angry.

Is that why you mispronounced Miyagi at the beginning of the video? Just Yeah. So ri people up. Maya Maya. Yeah. Anyway, let's wrap it up there. Um, cheers for sticking around to those who did. Uh, next week, who knows what will be what what the world of AI, web design, and development will hold. Yeah, this week was not too busy, but next week we'll probably have I don't know. We should do predictions like Grock who's who will be on top of the leaderboard. I don't know. I don't know. Just Gemini. Who's next? There will be one of these companies. Yeah. Make your bets making an incre increment. Um yeah. Yeah. See you later. Peace out.

Peace out.