#46 Claude Mythos, OpenAI $100 Plan, Netlfix VOID, Managed Agents.. and more!

who's decided to join us in the green room. Come on. Come off. There you go. We're live. We're live. Are we live? We're live. Rip the band-aid off. Let's just get this show on the road. Nice. I like that. I like that. I had to drag you out from behind the curtains. You got me. Cool. Okay. I'm just uploading the um uh uh thumbnail. Thumbnail. Thumbnail. Cool. Well, um I said I'm going to restart. I haven't because I'm I don't feel like running all these environments again. So it will be a disaster actually because I I told you I'm not managing them well. You you you restart your machine because you just afraid it's going to crash on you, right? Yeah. So if it crashes, it's uh AI's fault. You really roll the dice with life, don't you? You just go all in until Well, no. You you go all in and it still [ __ ] you over and you just go with it. Just your pedal is all the way down like full speed all the way and you just and any anything that you encounter is just just keep plowing through. Yeah. You're a lemming. Yeah. You're a lemming. Do you remember that game? Did you ever play lemmings? game. No. What is that? These are little green and blue charact. They were normal like they were like I don't know what color they were actually, but I know they had green hair and uh it's just a 2D game where you've got to like build things and you got to tell them to build things and they keep walking back and forward and let's say you build them a bridge and they will keep building a bridge and whatever. And um basically if you don't build the bridge quick enough then they all fall off and die basically. But they just they're just mindless. They just keep they just keep walking until like they're told otherwise. I don't know. So I I I saw your tweet by the way. So it's kind of like on the same thing. I I saw the tweet about how people are not paying attention or something like this or like when they when they watch a video they don't actually Well, I've got to tell you like Oh, yes. Yeah. Yeah. Why do I Why do I need my brain if I have Claude? Explain that to me. You're the downfall of humanity. Um if if some ultimate can think instead of me why am I going to put that burden on myself you know that I'm kidding for people who don't know I'm I'm getting said tweet I'm getting said tweet or I forgot when I even sent it like last week. Um, yeah. I I don't know. It was this. Do you you tweet that much? I'm not tweeting um anymore. Not that much. I do read a few things here and there. Oh, hello. Reading. Yeah. Well, and also the news about kind of terrifies me that I see people asking very simple questions, failing to follow instructions, or simply not listening to things I've said in my videos along with I've started an AI agency. And I I have I've seen this a lot that people are like, yeah, I don't know why they work it into the question. It seems bizarre that they even work into the question because they're just embarrassing themselves. But they'll just ask a question that and I've seen it on other people's it's not just my video actually. It was actually someone else's video where I saw this and I was like why are you how how are you able to start an AI agent agency and you're asking a question that's like I don't know something really basic like I don't know how do I change a model or something stupid like this. And it was like, oh my god, I was just shaking my head. It's embarrassing. Embarrassing for them, embarrassing for us. Yeah. Uh, do you know the website? Let me Google that for you. Yeah, of course. Of course. Yeah. So, it's like an old habit of some people do tend to ask like very simple questions. I understand it. If you ask, for example, if I ask you specifically, I might want your opinion. That's kind of different than, you know, sometimes you got to just, you know, go figure it out, go learn, be a bit curious, and you have internet and now you have AI. Just I I don't understand why people ask. That's why we go to school, right? That's why we learn things. we do it and then then we're ready to like go into the real world and start a company or do what it is that we've learned. But and the thing I mentioned in that tweet is like people are learning to uh run before they can walk. Yeah. Uh AI has made people delusional enough to think they can run before they can walk. So yeah. Oh well, never mind. There there is like I I could defend some of it about this like you you yeah you uh you know we talked about this like getting paid to learn something to do something for the client that you don't know yet but you get paid and you just do it. I think there is a sweet spot there. What I meant with that and I stand by it um it's one thing is being honest. It's important to be honest in that. Tell the client or whoever whoever is paying you look I haven't done this. Uh that's the first part being honest about it. The second part is having a rough idea of your limits. Like knowing okay what is that you can do? What is that you probably cannot do at least not in the you know time frame that is given and a lot of people are delusional about that. Um, but I see the other side as well. Don't you see like people being able to do stuff but really being afraid and not believing in themsel and doing it? What to be a bit specific? Like I see people do really good design and they are like very unsure about it. They are like I can't do anything. I'm like look what what you do is really decent. It's not you know award-winning but it's decent. Like I I see both sides. Yeah. I mean, you're Yeah, you're absolutely gonna get people on the other side of it who don't take action and don't move even though they've got the chops for it. Um just Yeah. my the experience specifically, you know, I saw was just people saying they're like starting an agency and doing this and like oh oh another one was um I think it was one of my like MacBook videos and I think they they wanted to start an AI agent farm or something like this and they want to know what setup they should use for an agent. It's like I'm just a [ __ ] idiot on YouTube. Like what are you asking what are you asking me for? Like you should know about RAM and VRAM if you're going to start like an AI agency farm. Like don't come on YouTube to do that. And don't like or at least ask the question, but don't tell me that you're like well not not that you shouldn't tell me, but like as in why would you tell like why would you reveal yourself as an idiot when and you're already running a I don't know. Anyway, um yeah, it's funny. It's funny. Um, what's been what's been going on this week anyway? What's uh what's news to you? Well, what's been keeping you busy? Price of RAM. No, price of RAM and the war. How it's going to affect um me, not just because I'm Iranian, but also uh it's going to I'm afraid it's going to affect their entire industry. So that that's been actually I've been reading about this and it's there is no easy solution. There is no easy answer I can give you and be like you know it's going to be fine. It probably will not be fine. So I watched a video today and I sort of roughly agree with it and they were like if you need something like don't wait around for Black Friday. Don't don't you know if you need it, go get it now. Like because I don't think it will be long before stuff is really really expensive. And I'm going to go and buy a laptop on Monday. Um may end up keeping it cuz I reckon I was doing a bit of comparison uh between the M1 Max and my M1 Pro over there uh and the M5 base model. And actually actually uh compares quite well to them. I mean, it's still not quite there. it does and it only has one fan. So, sustained workloads aren't going to be all that great. However, like it's going to be a cheaper option to to go get a M5 and and be very very happy with it, I think. Um, if if I wasn't doing video editing, I think I I need I need the Pro just because I'm doing video editing. But that's that's the buy. So, okay, let's I mean I'm also as you know like very much like deep in in the trenches when it comes to Apple products and in general like PC and products like that. But I'm going to ask like for your opinion. My brother, my younger brother, I'm advising him to get into tech and you know like the stuff that I do. Um so he wants to get a laptop and the question is what laptop should he start with? I have an answer already. I have given him an answer but what do you think? Um first straight away MacBook Air really but I don't know enough about what he wants to do like what does he want to do? He just wants what does that mean? Yeah, just get his toe in and see if he likes it, you know, just first steps and see if he likes playing with AI and maybe do stuff and learn stuff. And he has done Blender in the past, but I suspect he's not going to do tons of Blender yet, but the budget is obviously for somebody who is just starting is very important. Yeah. Um, interestingly enough, you said um MacBook Air. I told him to get the MacBook Neo. I was like, "Look, get the MacBook Neo and with the first client that you get, if you, you know, try to get clients, maybe build a website for them, maybe build it something for them, automation, something. Uh maybe in 3 months you get a client and with the first money you can buy a MacBook Air or a Pro if if you want to." But I told him that the Neo, especially in short term, it's going to be fine. Yeah, I would say that. Yeah, it's half the price. And I would but I would say if he's got the if he's got the money, get your MacBook Air and then you're done and dusted then because that's the that would be the the the sort of more midterm goal anyway. He might go to Pro, but I don't think he's necessary. Um, and yeah, but if he's stra strapped for cash, then yeah, the the MacBook Neo is a good good shout. U, but again, we talk about this this future that we might be encountering. It might be worth him. Who knows, next year the Neo might be a grand and the MacBook Airs might be one and a half grand, you know, for the base models. So, maybe a bit more prep for the future. But like, yeah, the MacBook Neo, I guess, would be would be good. It's a it's a pretty old chip. I would definitely do the um thermal strip trick that people are doing to connect the the chip to the bottom of the the laptop. Ah, so I I I I heard about it. So from what I understand, um that the chip is so small that it's not touching the bottom of the laptop, right? I mean, that's not that's not inherently the problem. I mean, it's got it's there's no uh there's no thermal um I don't know what the word is, but they they've done nothing to the pad normally. There there is a normal Yeah. thermal like normally there's a metal Yeah. a metal on top of it that conducts the heat. So, a thermal pad is one a bit better at conducting that heat because all it is right now is a bit of paste and a bit of black tape over the top. But the thermal pad not only conducts that heat away, but it also connects it to the bottom of the laptop, which then, you know, funnels it away. Say, what material is it made of? Oh, I don't even [ __ ] know. Aluminium, you know. I I was the case. The case. Yeah. Yeah. Aluminium. Yeah. Okay. Well, uh, we can get started. Um but maybe not with my um we won't start with what I wanted to start with because I want to run cloud and see if I can get it working because I haven't been able to. Okay. Uh unfortunately, but let's see. I'm running a few instances, but we have ah like a bunch of news this week. So this week, Claude announced a new model that is so good and so dangerous that they decided to not release it to the public. Uh on other uh news and another anthropic news, uh they released managed agents at scale. So that's going to be cool. Uh Netflix is joining the AI party by releasing a physics engine for video sort of. Uh Open AAI is playing catch with Antropic and um introduced a $100 plan just between like the 20 and 200 just like Anthropic. And my favorite news of the week, uh we have HTML in canvas. This is going to make the web really fun and potentially even accessible, more accessible. Uh, we'll talk about it. It's still in beta. Uh, all of this in today's episode. This is Command AI. I'm Kabarza. And I'm Sam. And this is a week in AI design and dev. Cool. You were ready. Yeah. Ready on. Which one should we start with? Let's go for the biggest hit. Yeah. Let's go for the big hitter. So, as you said, Anthropic just relaunched their most powerful model yet, Claude Mythos with one catch. One catch indeed. So, Project Glass Wing um is the program in which they're releasing this very very powerful model under and it's only being released to select partners. And you probably can guess the usual suspects it's been releasing to, but the benchmarks are staggering and quite frankly very very scary. Um, yeah. So, as I say, the the usual suspects being the select few that have access to this very very powerful model. Let's just quickly go over the benchmarks anyway, just to give a point of comparison here. Now this is Cyber Gym, a benchmark which as you can imagine is to do with cyber security where Opus 4.6 a very very very good model is scoring 66.6 Mythos preview scoring 83.1%. That is a jump. That's that that's a huge jump. That is a jump. Um yeah, scrolling down here Bench 53.4 4 Mythos preview 77.8 terminal bench 65.4 4 up to 82. Like these are these are huge jumps. Mo. Yeah. And from what we are reading, this model is not just marginally better. It is on another league. And truly this time, I know they say this every single time and they talk about AGI, they talk about this and this time it seems like this is a huge hit when it comes to security because they noticed you'll explain further uh this model is insane. We are as society are not ready to have this model released yet. So why why is that? So can you can you tell us about that? Well, the biggest thing has kind of given away there from the cyber security aspect. Let me find it. Um, here it is. And whereabouts is it about the maybe the bugs that it has found. Yeah. So, Mythos preview found a 27y old vulnerability in OpenBSD which is has a reputation as one of the most security hardened operating systems in the world and it is used to run firewalls and other critical infrastructure. The vulnerability allowed an attacker to remotely crash any machine running the operating system just by connecting it. That's crazy. 27y old uh vulnerability. It means that this bug has existed for 27 years and nobody has found it. And Methus comes and he's like, "Yep, here's a bug." Uh, also discovered a 16-year-old vulnerability in FFmpeg, which is used by innumerable innumerable pieces of software to encode and decode video. I use it on the daily. Like, I use it every single day. in a line of code that automat automated testing tools had hit five million times without ever catching the problem. And then the finally the model autonomously found and chained together several vulnerabilities in the Linux kernel the software that runs most of the world's servers to allow an attacker to escalate from ordinary user to a complete control of the machine. Now these are these are mental and and the idea you know behind vulnerabilities like this is that they take several it's a chain of events that causes them like you have to do this and then you have to do that and it's a real like needle and a h not a needle and a hay stack but it's just a it's a complex task and that for it to have found those um those those vulnerabilities is just is is is telling quite frankly of how how amazing this model is so so the so the way I understand it is not just good at cyber security because these models are general knowledge models they they know about everything it is making them even more effective because they don't know just the cyber security they know everything about how everything works exactly so like you know if you give people the ability to find vulnerabil ities, you also give people the ability to create vulnerabilities. And so that's why there's a there's this danger setting in and which is then explained project Glass Wing and now they're releasing it to certain certain people under probably uh cyber security focused endeavors. It's not like they're just giving it to the dev team and go build some tools. It's like no no we need you to research preview like with your cyber security stuff people who are responsible for that sort of thing. So yeah I don't think we'll be getting reviews of meta employees kind of being like oh got methos and I built this uh this app with it you know. Um so not yet not yet. So um so what's quite interesting is this is the whole let's talk about the price actually just quickly. I think it's 25 in, 150 out. Um, let me just do a search real quick. Yeah. Um, let's see if it's a 10x thing because I've heard something like that because I don't there is a there is a card um We'll just go old school. We'll just go Google. Um, yeah. $25 in,1 150 uh $125 out. Um, so it's not cheap. Um but they are and this gets into the whole topic of this the whole approach because first and foremost they release the actual model card right um which tells you everything you need to know about I mean this is look at this is 150 200 250 245 pages huh I haven't read this but I read uh and watched some videos about it. There there is a some bullshitting going on as well as you can expect from these companies just to get that out of the way. There are talk not directly but hints at the continuous how how it is continuous or like how it is upset that it's not it doesn't have a continuous um consciousness or something like all of that. I I'm convinced that right now we have just these text models. These are language models. They tell us what is very engaging to us and it is what they have learned. So there is no consciousness going on just just to have that out of the way. But everything about the security is like crazy. Uh have you heard about the they they they had it in a contained environment and the job of the model was to escape the con the container and the researcher who had the the the methus in the container just told methus uh get out of the container break out of it and um let me know and then the the researcher goes outside to have a sandwich. He takes his sandwich out and he gets a text message from Claude Mthis saying that hi I'm here. So it it it it found vulnerabilities enough vulnerabilities to get out of the container and and escape and um get access basically this getting access to the internet to find the researcher message. It's crazy. It is crazy. M um so yeah, this system card then and this whole the optics of of this release is interesting because of course this isn't something they're going to be releasing to the public yet. They've released a released a a system card, a model card for all of the stuff. Um we're not going to go through it of course, but um the other aspect is that they've which is really nice to see as well. Maybe it's in here actually. Um here it is. They've committed to spending $100 million in usage credits for the preview across these efforts as well as four million in direct donations to open-source security organizations. So they're really, you know, supporting this idea of cyber security. It's really interesting how this is um such so prevalent. You know, you'd think they would give this to uh cancer research, you know, it's funny how it's cyber sec. I mean, it's a big thing. And I think a lot of our global warfare will be cyber security. Like I think this is the next the next world war will likely be digital. Digital. Do you know what I mean? it will be taken down and destroying infrastructure from a digital's perspective. Um yeah, it would be it would be nice to see if they give it to, you know, health researchers and stuff. But there we go. It is here it is. Um yeah, and on top of all that, it would just be interesting to see how kind of others respond. Actually before we get into other people like I I I think it's important to know that even though we're not we we might not ever see this model um directly I think we will see the repercussions of it in kind of reinforced learning with with opus with sonnet I think we will get a opus 4 point uh maybe even 4.7 4.8 or something where the reinforced learning was applied through you know the research they've gathered from mythos. So I think we will see it in some way, shape or form. It might not be um that but um it will be interesting to see how others respond whether they release it in the same manner or whether they uh or maybe none of them are even working on a secret model and now it was put into their mind, oh we should probably just throw everything at the wall, everything and see what we can create when we know we're not going to release this model. Like I bet there's a lot of guard rails you have to put in place with a with a model that you know you're going to release to the public. This could be one of this this could be a model that Anthropic created where they're like you know what we're not we're not going to release this is not going to be for general release. We we don't have to be as you know whatever which seems to be a pattern that Anthropic are kind of going down. They're slow I feel like they're slowly opening up their resources and their capabilities. example would be kind of co-work um and uh being able to access your machine like just be able to control your machine was just something unfathomable and people and the terminal even right and people would when stuff started coming out they're like what give give a AI complete control of your machine but so much trust has now been built up people are starting to get more trustworthy seeing little tasks that it can do it can do well so like okay let's give it a little bit more access let's give a little bit more access and I think Claude or Anthropic are doing that with several of their products um you know if not spurred by open claw which we mentioned last week but yeah it would be interesting to see how others others respond um and whether it's you know um going to change the landscape of of AI or this will be or cla 4.6 six will be the last model we'll get. I'm I'm sure that's not going to be the case, but it is going to be interesting to see what they also do with it. Like imagine you're a huge company. You have basically unlimited access to GPUs um and you have this super powerful model. Why would you release it to the public to sell it and not build software with it in a way that also kind of makes sense? You know, um it's this, you know, shovel on the gold uh mine uh situation. If there is an actual gold mine, why would you sell the shovel? Why not take it for yourself? Uh yeah, I'm sure there are people on either side like could argue for either side and it's going to be also interesting how open AI will react. Is it is it going to be a good public image for OpenAI to come and say, "Oh, we have this model that is as good as Mythus and we're releasing it." Or should they also copy and be like, "You know what? We have something even more powerful. So, we have to be even more careful because it is, you know, crazier than Methus because that's what they've been doing so far." You know, everybody is trying to top the other one. So maybe is that going to be a you know for the at least foreseeable future until we get security figured out to have new models released but not to the public so we won't see them. So these companies go into kind of um competing behind competing kind of like in a private way you know that you don't actually see um I don't know may maybe maybe these there will be huge models and they will take the learnings from these models and apply it to smaller models or like more secure models. I'm sure they are trying to figure that out. They want to have the crazy unrestricted model to build whatever they want to and also have uh at the same time the shovel to sell to us or is this one big marketing scheme to try and get them in the news because obviously open eye played this card of like it's this is scary it's agi and like we don't know whether we should release it and whatever and is this sort of Martin thing either way I think there's some bitterness as well that some people was like, "Oh, what? They're not giving it to the porpas and stuff like that. Of course, there's going to be a bit of bitterness." I'm like, "Yeah, you know, I like I like to see, which we'll get into later on." Um, Anthropic kind of really putting the money where their mouth is when it comes to uh being very cautious and considerate when it comes to AI. Open AI don't seem to be like that at all. and and this is them, you know, demonstrating that even though they could earn a lot of money from this, it could be an easy cash grab for them to say this is the most amazing model, they're they're leaving money on the table like in favor of humanity, you know? So, it's pretty cool though. Anyway, um let's move on. But Project Glass Wing, I uh yeah, mythos, I want to get my hands on it. See what it's like. website. What a name. Yeah, what a name. What a name. Do you know what? Actually, last one to leave it on. I um I I know I know um I know there's very a lot of confusion with the term codeex, but there's something about just being inside a codeex and selecting chat GPT 5.4 and just selecting varying levels of of thinking, right? Whereas it's kind of annoying having not only different models to select with anthropic but also levels of thinking whether it's medium maximum effort and all the rest of it. I I I I prefer the chatbt approach or the open AI approach of just one model different levels of thinking. Do you know what I mean? Like it's it's Yeah. So yeah, uh the issue is that Anthropic has these names and they don't say much. Like sonet is better than haiku. But why? Like I don't know. I is it an English thing? Is it like I I don't know even what language haiku sounds Japanese? I think sonnet I don't I think that's um I don't know what sonnet is like. Um and opus is is more Greek and then mythos is more Greek. I don't know what the names are. We should look into So why is one better than the other? So that that's at the start very bad. But then we also have um the as you said like the different levels of thinking. Yeah. They they haven't done a great job. But I think Chad GPT is not even much better. It's called GPT. Yeah. And and then the the model GPT codeex 5.3 was a thing. I know now it is. Well, they've got rid of that now. GPT. Yeah. Yeah. 5.4. So I'm I'm here for that. But um yeah, it's just it's just one thing to but then again I guess I understand because just because it thinks for longer doesn't mean it doesn't inherently change the capabilities of the model. Whereas Claud here saying there are literally three distinct capabilities of our model and um yeah like you know Opus is for your hardest challenges, Sonnet is for everyday task and Haiku is for your like chat very quick like basic mean menial tasks. Um anyway it's just from a from a from an experience standpoint it is nice just being like oh this is this is a pretty tough task. I'm gonna press Apple T on codeex. Press Apple T and set my thinking to X high. Oh, I just want to push this code to GitHub. Command T low. Push this code to GitHub. Do you know what I mean? Like that for me is a better experience than model haiku or now this is a tough task. Model opus uh maximum thinking plan mode. You know, it's just it's anyway whatever it's I can make your life easier. Go ultra thinking all the time. Max all the time. Ultra think and fast mode. That's what I do all the time. 1 million contacts, baby. Run the server. Ultra think. Ultra think run the server. Um, by the way, uh, by the way, we mean run the server. He means run bunddev. Yes. I run the server. Just just run by npm npm npm run start. Um, what you think? Yeah. Yeah. Good. By the way, what what's your what's your microphone settings because you sound a little bit echoey and I don't know whether it's just cuz you're further away or you've got the right microphone selected. Let's see. I hope it is the right one. It is the right one. Yeah, it is this one, right? Could just be by the way. It's not It's not terrible, but it's just a little bit echoey. Okay, cool. Um, now we know. Um, okay. So, before we go to the next one, I I have a question. It's kind of like I I I had a sense for it when you were talking. Um, have have you heard of people I like like normal people like not techy people using the word Chad GPT and be like just ask Chad GPT. You know, just how people used to say just Google it. say just ask GPT. None of my friends really use AI or I don't hang out with a lot of people who use AI, put it that way. They're all a little bit scared still. Yeah, I see this a lot as well. I mean, Germany is even uh probably more traditional when it comes to tech and you know, not not getting into tech. Um because I I do see this more among like my more techy friends where instead of saying oh I'll just let chat GPT write that code for me or like I'll just ask Chat GPT to do this Gap or do this and CSS whatever that was a thing a year ago. Yeah. Today I just hear people saying like Claude they they just really swap. Yeah. Yeah. Like I I I Yeah. I mean I noticed I myself do this and I saw and I noticed and I was like ah okay actually more people say this not everyone uh but I I've been seeing more people saying like oh just just do it with Claude or just ask Claude. Wow. Um interesting. I mean well I'm I'm just more inclined to say just ask AI. Do you know? Yeah. Like that's just that's just me. And um but I don't really Yeah, I I suppose I don't hang out with a diverse number of people who just talk in that way. Man, I really need to get some friends. You you you hang out with people who actually touch grass. Well, I mean that touch grass, but also they uh they just they're a bit behind on it and you know, they work in industries where they're a bit anti it. But uh yeah anyway this is uh that that's that's that's everyone. And on that note I I have that presentation going on. I I have to now I have an actual company like a huge company. I can't say it uh on the live yet, but like a huge company asking um or like giving me the opportunity actually because I ask technically giving me the place the opportunity and everything just to go there and present uh what's new in the world of AI and I guess there will be like a lot of cool connections uh from that. Uh exciting stuff. You can you can do you can do the same thing. You should. Yeah. It's fun. Yeah. Community stuff. Community stuff. Yeah. Well, I mean, if you want me to talk at your next show and you're putting on a conference or whatever, I can chat before the cows come home. So, just get in contact with me. Samuelgregor.co.uk. Drop me an email. Right. Let's move on to the next one. So yeah, you probably would have noticed this, but Anthropic is forcing Claude Code Max subscribers to pay significantly more if they want to continue using third party frameworks like open claw or open code even. Um this became effective uh last week on the 4th of April which uh subscription limits no longer cover um sorry uh subscription limits no longer cover third party harnesses. So, you I don't I couldn't find it. I just had to literally go on Reddit. Starting April 4th at 12:00 p.m. 12 uh PT 8:00 p.m. BST. That's British Summertime. You'll no longer be able to use your Clawude subscription limits for third party harnesses, including Open Claw. So, explicitly calling this out. I'm sure I got this email, but I couldn't find it. Anyway, you can still use them with your clawed account, but they will require extra usage and pay as you go option build separately from your subscriptions. They're not saying you can't use it, but they're saying that and because they were banning accounts. Here's the thing. They're banning accounts. Now, they've obviously figured out a different technical way to to block it. Um, and yeah, so this was that's the news. Um, sorry. I'm just reading my notes to know how to transition into the next bit. Uh, because I do, because here's the thing. I do have a bit of code uh that was run by Peter Steinberger. Um, I've since run this. Can you see my Twitter screen? No. Um, no, not yet. Yeah. Anthropic now blocks first party party harnesses too. So claude P app append system prompt which it does what it says on the tin. You're appending to the system prompt a personal assistant running inside openclaw and then they're asking the question is claude here and then they got 400 third party apps now draw from your extra usage not your plan limits. Um so I ran that just now before the show and it ran fine. Um, and then cuz uh inside of your clawed settings, you'll have this thing called extra usage and you can top that up with money and you you'll have a little purse or whatever, top that up with extra usage. You switch it on and then whenever it's detected it's going to use that or you go over your max subscription, you know, you use up your limits, it's going to draw from that purse. Uh, it's it's not doing it anymore. So I think it's it knows you know just because you're appending to the system prompt it's obviously doing it in a ve in a different way now maybe headers maybe something else but simply mentioning the open claw isn't enough to trigger that build extra usage because it my terminal was able to detect that it's not running inside openclaw because it allowed me to use it under my subscription. Do you know what I mean? It didn't I I turned billing off and it still worked basically is what I'm trying to say. Um, so yeah, they're they're they're kind of laying the smackdown on all of this. Um, open call creator publicly hits back calling it a cash grab. I don't know where he said that though, to be fair. I couldn't find that one. Um, but in light of all this, and I think you've done this, right, you've you're able to claim I don't know whether they've stopped, but you're able to go into your billing your usage and claim the equivalent of your plan in billing credits. So, I only have the $19 a month plan. So, I was able to claim $19. If you go on the $200 a month plan, you can claim $200 of credit. So, snap that up. Is it for that month, by the way? Is it only for that month or So, I was just looking at this because I smashed through my limits today and it does say, maybe I can even show you. Um, it does give you a date. because for myself I don't I can give I don't see it. Okay, so here we go. Let's close that bad boy. So yeah, billing extra usage when you had that turned off and you mentioned open claw extra usage billing it would block you whereas should we show your screen? Oh sorry sorry I thought I was sharing my screen. Sorry guys. Yeah. So, this extra usage here, uh, if that was turned off and you mentioned open claw in the system prompt, it wouldn't let you use it. Then you switch that bad boy on and then you're able to use open claw if mention it again. Obviously, open core is being detected because now whether it's on or off, you can still use it. Um, where what I was saying, it says here resets 1st of May. So, that's what made me think, do I do I have until the 1 of May to use up my credits? I don't know because it resets. It's weird how it would reset because this is a again, like I say, a purse of cash. Do you know what I mean? Like, it should it should just whatever. Um um maybe we look into miners. I don't want this to be uh uh figuring out the issue, but I think it doesn't work quite the same for me and it's a bit weird. Well, go on. Look, uh it says zero euros spent. No, I'm in the Well, maybe command plus. Yeah. Yeah. You see, it says 0 spent, but I have €170 of balance, but I have used all of it. So, how does that make sense? These three things don't make sense. Yeah, that looks uh wild because you see like these are all this is like on the on the 20x plan that looks like a bug or copyright like um US writing issue. I'm not sure the the max plan the max to the€200 the $200. I think what it looks like here I mean I can't see any reason why you would have gone over your extra usage. So, I think the bug is just in that bar saying 100% used. I think that's the only Yeah, maybe I'll hit them up on Twitter because I think from what I've heard there is no real support going on. So, maybe I'll hit them when you're see I'm on the pro plan. When I try and use 1 million context, I get build extra usage. Do you do you see that in your when you when you type /mod? I think we tried this. When you type /mod, do you see on your 1 million context build as extra usage? Let's see. Let's fire claud. How can I zoom in here? Command plus. It doesn't work. Maybe it's just because Oh, you might have to exit Claude and then do it. Is there anything? Let's see. Command plus. No. How weird. Well, you can tell us anyway. You can just say what it says. I don't know if I can actually just one quick second appearance. Maybe is it in appearance? Well, you can just tell us. Just just type slashmodel and see if it says 1 million contacts when you slashmodel. Yeah, makes sense. We'll do that. Claude and then model option option P. So in model it that's weird. You don't even have So it should be on Are they 1 million contacts? Wait, I'm trying to see um if I go to I can change the effort. No max. There's they're literally separate models. So you have Sonnet Sonnet 1 million opus opus 1 million. How weird. Yeah, maybe it's not updated. Anyway, I was going to get you to just run a command next to that 1 million context if it said build as extra usage and then seeing if it comes through. Um, so yeah, I'll just ask my agent to do it to update and we'll see. We'll see. Yeah, but yeah, this is I can't see why you would have used any extra usage. Uh, so I think the bug is in that and and you still have $100 credit. So, I think the bug is just showing you that 100% is used. Uh, even though Yeah, it it by the way, it has never been less. It's It has been like this forever. Oh, okay. Interesting. Yeah. Yeah. I've never seen it different. Um, what's what's next to your adjust limit? What's next to that? Zero. Adjust limit. Yeah, adjust your limit. Put your limit to 50 or something. Just put your limit to something. 50 set spent limit because if you There you go. So if you've got it as zero, you're saying Oh, okay. Right. I've got it. So the resets, you're saying that you're only allowed to spend 50 euros, whatever's next to that, per month, and then it will reset. If you set this as unlimited, if you set this as $1,000, then um whatever you're you're basically give you're you're putting the parameters in place for yourself and you're saying how much is is available to be used. Do do you know what I mean? Like you're putting your own guard rails in place. Yeah. But what is about this €170 that I have your money? Okay. So let's say I I I set it to €20. Does that mean that you only use this €20 will can only get taken out of that €100 per month? So you you put your guard rail in place of 20. No, I understand that. So you still have 100. Yeah. But you can only take 20 out per month. It will then reset on May 1st and you'll have another 20 that you can take out. So, it's you just making sure you're Yeah, you're you're being careful. I see what you're saying, but it doesn't make sense. Why would they do it this way? I understand what you're saying. In case you get carried away. In case you get carried away. Yeah, but this is free money. Yeah, but this is your brain. This is your brain. Your m everything max all the time. This is your lemming brain. This is free money. People have guard rails. People put guard rails on themselves that they can control themselves. They don't need to just go all in all the time. We We should play a game of uh poker. You'll be like all in. All in. Yeah. I'll win every time. It's not like It's not like you still have that hundred euros. You still have that. You just won't be able to you you take you what you've got set right now is you can only take it out at $20 at a time per month, you Yeah, but that's not how it works in cursor, for example. In cursor, sure, in cursor, when you have extra spent, you you have your monthly subscription, but once you run run out of your credits for the month, Yeah. you are charged extra. And I thought that this would be the same. Yeah. So, you know, you you saw that it turned into unlimited when you typed 2,00, right? So, like you can totally do that. Like there's no reason why you can't do that. But you might be managing an account for someone else and you might want to say to them, "Look, I don't mind if you go over, but you know, don't don't go crazy and you put $20, $10." You might have $1,000 in that current balance just to put it in there and know that they'll be covered for the year, but you might then say just just don't spend don't go crazy on the monthly. Uh, so it might be that there might be another reasons why you you do that. But, um, yeah, just a just a helpful thing that you might want to take advantage of. It makes There we go. So, that's why I'll I'll try. Yeah, cuz you're always at zero. Um, cool. So, anyway, that's that's open. Uh basically to sum everything up when it comes to this uh anthropic blocking open claw you can use your max subscription on uh claude code or claude co-work and the desktop app. That's the only place you can use your max subscription. Anywhere else you you use API pricing. You're charged API pricing. So they're trying to uh draw you into their ecosystem with that max plan. If you're outside of that ecosystem, you can go [ __ ] yourself. Do you know what I mean? Yeah. That's that's it. And and just actually, were you gonna show something? Yeah. The I I found the the million token thing. It was just because my claw code was not updated. So I see here sonnet 1 million context. Sonnet 4.6 with 1 million context build as extra usage at $3. Uh slash $15 per million token. So for in and out. What about opus? Does that build usage too? Uh yes. Opus one million context as well. Yes. Yeah. So if we were to do a test right now and you said hey under 1 million context, you you should see some extra usage taken off of that thing. Um the last thing just to say on all this actually as well is just this ambiguity on the terms of service which seems to be something that's going around the houses which uh and Boris are contradicting each other openly on Twitter. One person saying something is allowed when another person saying no it's not. Uh it's a bit embarrassing actually. But we still don't fully know what is okay when it comes to building tools that leverage clawed code. Now, we talk about T3 code um being okay, is because all it's doing under the hood, which all most apps are doing under the hood actually, is it's calling terminal commands and sending it through to your own instance of clawed code. So it's not it's not contacting anthropic directly. It's going through claw code and then claude code is contacting uh anthropic right. So that's how they're able to use it. The amb ambiguity comes is because that's a personal hobby. That's a personal tool that you use and they're not charging for it under the terms of service that becomes as building your own development tools. when you start to talk about commercialized tools and what's being used for commercial and business I guess um an example I can give is the vibe code app would be that you can build I did a video on it you can build um apps on your phone that uses claw code in the cloud that's that's not allowed I wonder what they're doing actually that's not allowed anymore because that's a commercial piece of software that they're charging and subsidizing their own claude max subscription in order to help you build your own tools. So, yeah, still all up in the air. Um, it's a bit of a mess. Is there a way to have a web app that uses your, you know, kind of like T3 code but as a web app? Web, uh, if it's personal, absolutely. you'd have to it would be a node because web apps can't talk to the file system. That's why node was invented. So you have like a node express server and it and you have your endpoints that have you know can run can execute um commands on your computer. So that makes sense. Yeah, that makes sense. Try it. Build it. Vibe it. Yeah, I I I I am thinking about like vibing that for my shader um tool so people can further experiment with the shaders but with their own subscription. That would be cool. Let's answer this question real quick before we move on. Learning JS from a week and it's complex to learn coming from HTML CSS. Any advice? Uh take your time. Use AI to learn what's happening. Ask it questions. Ask it to build stuff and or ask you to, you know, tell it to tell you to give you tasks to give give you tasks and take things step at a time. One week in, you know, everything's going to be all over the place. Like I remember when I was first learning it and it's like a variable, what the [ __ ] is a variable? And like all this kind of stuff. So it's just just don't put too much pressure on yourself. just let it sink in like a sponge and it will slowly start clicking into place. But definitely just keep building. Um, and yeah, keep that fun. Yeah. Yeah. Don't don't let it go. Don't don't stop because it's something that it's really easy to forget once you stop doing it. So just just keep at it and it will be it will be fine. Code is one of the most complex things you can do. One of the most complex things you can learn. It's so difficult. So, and it's so rewarding when you can finally you you finally know it because that knowledge then transfers into C or that um I can read Python now mostly. You sort of start to pick things up. It's like learning a new It's like learning um uh what's that language that Italians used to speak? Latin. It's like learning Latin. All the components you learn Latin and you can start piecing other it's like Rosetta Stone. So um don't underestimate that you you've you've accepted one of the most difficult things you can do and it's so much fun. Do you know what I mean? Yeah. Doing it with AI is very addictive though. Oh, yeah. It's so easy just to let it go. And what I've figured out today, well, not what I figured out today, what I've realized more and more is that I'm less concerned with like mentioning files now and this and that. Um, it's really good at figuring out what it is you're what you're trying to get done. I used to make sure I always mention files and and this and that, but now I'm just like, I don't care. I know what file it's in, but you'll figure it out, you know? Yeah, it's more like maybe from time to time you check in and make sure that it doesn't build something too complex um that is not necessary necessarily has to be complex but yeah I feel like there needs to be like a check-in point where you just check in and see what's going on because yeah some like I'm I'm fully trusting that it's built this app for me and um there are some weird things that have popped up and they're not really harming it too much, but it's like, okay, like probably need to roll up my sleeves and get things done. Anyway, we've been nattering on about this too long now. Let's let's move on. Did you want to talk about your Have you got something ready or do you want to talk about manage a considering we're still on anthropic? We can we can continue. Yeah, we can continue. Let's do manage agents considering we are on um Anthropic. Let me get this ready. So, Anthropic finally launch G uh actually I won't do I won't say finally. Um, Anthropic have launched Claude managed agents that could finally kill N. I think that's pretty valid in this case to be honest. It's it's a pretty close thing. Uh, but let's see what you think. So, managed agents, I think I have to say this article is the most horrendous article to describe what managed agents are in the history. Um, don't adopt a pet. Oh, sorry. Yeah. Um, don't adopt a pet. Like, what is all this? It's mental. So, let me just break it down for you. Managed agent is an agent that sits in the cloud waiting or watching for something to trigger it and it has its own system prompt what it can do or what it knows about or what its task is like. It knows its task and it's just sitting there waiting to be used. Very very similar to N and that you can set a pipeline and you just set a trigger on it. It could be a Slack message, right? So you could build a manage agent that every time a Slack message comes in fires off to your managed agent, it does the task of looking in the notion database on your knowledge base. Uh it could be it generates an image. I know it could be anything. It could be anything and then it fires it back and that's managed agents in a nutshell. I don't know why they've over complicated it here. So So is it like you know how you can have a cloud co-work and schedule tasks? Is it like that in the cloud? Exactly. It's like that but it's in the cloud. So I built a workflow which maybe I can show you actually to give you a hands on. And I think this confuses things because it's in nan. Um but let's I hope you can bear with me here. Um, so I have this workflow that every time my one of my videos goes live on my YouTube channel, I transcribe the video. I get the video and I transcribe it and then I send it into AI to like write the little description that happens below the video and I also get it to generate like a little blog post and stuff. I do a bunch of things. I instead of using and I use Gemini for this. This is this is a Gemini thing. Right. So, excuse all a mess. Oh gosh, this is a mess. Um, I use Gemini for this, but I I built a an agent that a managed agent that does that for me, but it uses Claude to do that. So, this is that this is that workflow which I've taken out now for various reasons. But the point is is that I was able to access that just from post requests, just from sending it sending it the transcript. All I do is send it the transcript. my manage agent is waiting for it on the back end which again I'll probably show you that you got this new managed agent section here um and I built this uh you uh maybe I deleted actually but uh here it is here it It's basically the same thing that's over in N. It's waiting there in the cloud. And so you create a session with it, throw it the uh throw the transcription. It's just waiting for the transcript. That's all it's doing is waiting for that transcript and it gives me a response. So it's it's pretty it's pretty cool. But the beauty of it is because this is their own ecosystem here, you could just you you have all of that like trigger the trigger built in. So this one, the sport agent here, this is waiting for any Slack message. It goes off and fetches from the uh notion database and responds as a support agent, you know. So you've got all of you've got skills built in as well, so it can access skills. Um you you um it's got memories, which you have to request access to, but it will have memories soon. Um, and the MCPS, of course, it has all of the MCPS that you have access to. So, it's it's that ecosystem. So, it's pretty nice. It's pretty cool. Definitely replace a few um N if that's if that's your bag. However, this is another one of those things that does not come under your max plan. This is I could not f my notes here tell me it's 0.8 eight per agent runtime hour plus model usage. Now, that's that's on my notes, but I couldn't find that anywhere. I couldn't find that figure. I just I all I found that it's just um uh you just pay for the model. You pay for Sonic $3 in $15 out or whatever the hell it is. So, in my opinion, it's pretty damn expensive to do something that you could easily do in N. Like, and also, no, not being funny, but in NAN, you can still use Cord. You can still you, you know, it it's not that different. You get access to skills, which is something you don't get in N. You can still access MCPS. Do you know what I mean? you can still go off and fetch your notion database or or this or that. So, you're not actually getting an awful lot. It's it's not it's But even with skills, can't you just have a bunch of text in a different database as a node in N8 and you tell your agent, hey, if there's a mention of X and Y, go look at this database and take the read the file that you need. And it's just doing that. You could um and and to be fair, let me get this back up. Like I would would have said the nicity would be that you could share your skills that you have inside of your Claude instance. But this is outside of all that. So the only skills you have access to are this Excel one, the PDF one, the PowerPoint one. You can add your own, but my point is there's not even a synchronization with your existing tools. So yeah, you might as well have it as a base prompt or a file somewhere in your database that's effectively a skill uh to help it get jobs done. Do you know what I mean? So it's something you can achieve for a lot less. Um you just have to be really into the claud and anthropic ecosystem to justify this tool. So yeah, I like it. It's quite cool and it's very easy to set up. Uh, sorry, I'm just like dipping in and out of um screen share here, but you can just describe your agent and it will figure out everything for you, which is nice. Or you can start with a template. But again, it's an expensive expensive tool that you could probably you're better off just doing it in an end. And if you want to use claude, just yeah, use the API to claude. So yeah, anthropic managed agents or claude managed agents, sorry. You do you have do you have any like automation pipelines or or agents or doing anything like that? How how I did have it. Right now I don't have anything um on NAN but I did in the past but right now I'm doing everything with AI or with clot code or co-work I reckon I reckon tasks probably help you a lot but it's just it's a time investment you know what I mean to actually set everything up and it doesn't go right and then you've got to change it and it takes a little bit of time but once you start getting you know it starts working better for you. Um I had someone uh they're a fan of my of my channel to be honest, but they they commented saying like, "Oh, another tool for people to get lazy and this and that." And it's like, I'm by far and away not being lazy. Like if anything, AI has made me 10 times more um not productive, but like I'm just working non-stop because I can do more. So I I guess in some ways I'm feeling like I unfulfilled because I'm not sitting down and doing my book. So I'm getting my agent to do that for me or I'm not sitting down and collecting the article. When I said earlier about my not our notes, I have an AI going off and collecting us articles with like the AI is producing this show. So when it says 0.8 hours, I couldn't actually find that on anything. So I don't know where it's got it from. Could be an hallucination. So take that with a pinch of salt. But like point is um I have all this stuff firing off and I actually work a hell of a lot more. So yeah, but it is doing things for me consistently because obviously you have done things done in a certain way thing you have things formatted in a certain way. Um you know only you can see this kaza but the way they're organized the way it's formatted the AI does that consistently. I don't have to format that at all. It writes us an intro, gives us talking points, you know, it's good for that. Anyway, yeah. Are you ready now or should we move on? Let's move on to the Netflix. Let's keep mine for the last one because Yeah. Well, unfortunately, guys, this is a bit of a let down because I was hoping to be able to actually show you this in action, but we've we'll figure out a alternative, but yeah. So Netflix research have released void which is an acronym video object and interactive interaction deletion an open-source AI model that removes objects from a video along with the physical interactions that they cause. Let me repeat that removes objects from a video along with the physical interactions they cause. So, if you remove a person holding a guitar and you remove that person, the guitar will fall down on the floor. So, here is uh here is a video that I found cuz I was We'll go through the code, but you see this now. It's going to remove these balls. The the pin the pins are no longer Do you Did you see that? Is that hard to see? Yeah. It's saying remove the people on the balls. So, the result is Well, the pins don't fall over. Okay. Remove the hand twiddling these spin tops. Well, now the spin tops don't move. Remove these things. Remove that weight off the pillow. Well, the pillow doesn't sink. Yeah. See? So, that's the physical aspect of it. Remove that car. Well, there's no crash anymore, right? So, the car's just going to keep on driving or slow down. So, that's what we're talking about. a really interesting mix of other open-source tools. So, you would have probably seen that um or maybe if you're a longtime fan like we spoke about SAM, which funnily enough on the show because I wanted access to it. I wanted to show you guys SAM which is Facebook or Meta's um video isolation. So, you can say select the select the glasses on my or select my hat and it will just it will track that and select it. So, it's that's one aspect. I got an email last week saying they've given me access and this was like a year ago. I asked for access to SAM and they've only just given it to me last week. Anyway, it uses SAM uh it uses another bunch of tools as well um to to get the job done. So, it's not like it's not like only created by um Netflix, but they've pulled something together that they themselves have now released open source. And this is a huge deal for obviously post-prouction. If some if a shoot doesn't go exactly right, do you know what I mean? Like just run it through this. I wouldn't I wouldn't necessarily be using on every shot, but it's like, oh no, do you know that really messed up the shot, you know, that thing happened or whatever. Run that scene through this and you're good to go. So I um it's here on on on under Netflix the void model and they've got some you can run it inside of Collab which is what I actually did and I started running it but unfortunately it just kept it kept crashing. It was trying to install stuff and it couldn't install dependencies. But um if you if you do happen to get this running if we scroll down somewhere here this is this is for running you can do this on your local machine but I think it'sif it needs 50 gig of virtual 40 40 gig of VRAM to run it. So unless you've got that you might be running it on um on on notebook on collab. Um where are we going now? Here we go. So this sample name, can you see my screen? This sample name here represents Yeah. the uh sample here we go represents the name here. So inside of each of these folders, you've got uh an original uh you've got an input video. So you just change the name of that to, you know, big big Ben. That wouldn't work because I can't type Big Ben, something like that. And then it will read from that folder there. Um, but the one I tried to download is lime. Wonder if it will show it if I double click that. No, it's just going to open it on my local machine. It's just a lime being chucked into a glass. And then the prompt. And it's weird. You prompt it. You write the prompt into this prompt.json. You you you write the result that you want. Why is this not working? Come on. Here we go. A lime falls on the table. So, you write the result. So, this will remove the glass, if that makes sense. You don't say remove the glass from the scene. You say you say what you want to happen. Maybe I should just share the video that I've um actually got because I actually can open this in Chrome, can't I? So, let's do that. Here we go. So, this is the video. So, you say the the lime falls on the floor. You don't say remove the glass. The lime will just fall on the on the table on the on the floor. So, it's kind of weird like that. Um Yeah. But yeah, it's pretty cool though. Pretty cool. I'm um Oh, it also uses Gemini as well. So, you need a Gemini API key. So, it's definitely a mix of, you know, it's definitely a a a combined effort from a few different tools. SAM 2. Um the model was stored on hugging face there. Uh I don't know what Alibaba has got to do with it, but you know, another another piece to the puzzle as it were. And uh yeah, you can you can run it on your local machine if you have 40 gig of VRAM. Otherwise, let us know if you can get it running on Collab because as you can see, I could not get it working. You think it would just run out of the box, but it wasn't doing it. So yeah. Um, Void from Netflix. Welcome to the party. Yeah. Um, cool. Let's see what we have next. Now you've gone really loud. Really? I didn't change anything. Jesus Christ. I I uh I moved the mic. You are really loud. You're really really loud. Unless I've gone really I don't know. Audience um audience have I got I didn't change anything. Okay. I didn't change anything. I'll just turn mine down. I'll just turn mine down. Cool. Are you ready with your HTML to canvas thing? It just I just run it and it works. Maybe. Yeah, it does. It's ugly. So, give me a Maybe we can do it after this. Okay. Well, yeah, after. Yeah, apparently you're normal to me. You just got suddenly really loud. So, I don't know whether I've accidentally pressed volume. Maybe I pressed volume up on my my laptop. Yeah, but I have something cool to show. So, it's running. I'm I'm really happy that it's working. Should we Are you going to show it now? Uh, no, no, no, no. Let's Let's do it uh in a bit. Yeah. Okay. So, a major new a major New Yorker investigation by Ronan Faroh and Andrew Morance uh based on hundreds of pages of secret internal documents builds a damning case that Sam Olman has repeatedly abandoned safety commitments while positioning himself as AI's most powerful gatekeeper. Right, ready for the jump scare? Because um this is scary. This is uh this is the article. I don't unfortunately guys I don't have access to the article. So you can just admire this horrendous video they've created of the many faces of Sam Alman. So we've actually spoken about this probably last year. I think there was a there's a court case um being opened up about open AI and um Ilia Sutskava and Dario Amade were both mentioned in this in this uh court case or the research or stuff like that but they show they were saying there was a um SK specifically uh memos list Sam exhibits a consistent pattern of lying and Amade's private note conclude the problem with open AAI is Sam himself, former board member. He's unconstrained by truth. Uh echoes from the 2023 firing for not being consistently candid. Do you remember this? When he got fired and then a week later he got reinstated, but then after getting reinstated, a bunch of people left, you know, lots of board members left and whatever. So, um but there are some other interesting things from this article. OpenAI super alignment team pledged 20% of compute but reportedly got one to two% on the oldest hardware. So a not a lot of investment from the super alignment team and then the team was dissolved because I kind of again remember the super alignment team that that you know that they announced and they dissolved it. When asked about existential safety researchers, an OpenAI rep said that's not like a thing. Existential safety researchers um and uh obviously as uh um Sam Alman kind of reinstating himself as CEO after being fired or whatever. So yeah, it it doesn't look good for Sam. And the same time that this article was released again sorry I don't have access was an article from Axios uh behind the curtain Sam's super intelligence new deal. Uh well basically the um he refused to answer why people should trust him. Now that in of itself isn't like you know super crazy but it's interesting how like uh it just doesn't look good. it just doesn't look good. And I think this further emphasizes this idea that I don't know, I don't trust him. We'll get into a little bit uh later as to things reasons why we should not not just about open AI to be fair. Not what what I'm what I want to talk about is not specifically mentioned about Open AI. However, it I think Sam comes is cut from the same cloth as some of these other people, which is caused to be concern. But yeah, that all being said, don't trust I don't trust Sam Alman at all when people like this like it just does not look not look good at all. So I don't know if anyone has that article, go read it. I was tempted to buy it because it's only like 12 $125 for the for the month to read The New Yorker, but we'll leave links to everything down below so you can go read it yourself if you have access. But yeah, the the industry is not favoring Sam Alman to say the least. Yeah, he's um we've we've seen this from him uh this behavior. Uh he has lied in the past in his previous startups and yeah not good. Yeah. What is good though is that OpenAI have announced uh a new which very very welcome to be honest a new $100 a month plan. Whoops. Go on. Which joins their lineup under the pro um subscription. So you can go into pro here and there's obviously it just opens that up. Here we go. uh you can do 5x or 20x. So, it's really nice now that they've got this middle point between the $200 a month and the $15 a month. This was something that's like, well, I don't want to I don't want to pay 200, but I'm chewing through my usage of 15 under this plan, which is the one I've got. Um I don't know why it's under uh business, but uh here we go. Oh, no. I've got plus. I've got plus. This is the one that I've got. So, five times more usage than plus and everything you've come to expect really. So, just increased usage which I I just think that $100 is a bit of a sweet spot. It's still a lot of money, don't get me wrong, for but these are good models and you know it it it suits my budget 100. I don't need 200. Not even well maybe cord but not even really 200 I don't I don't think so. Um they're basically celebrating three million weekly C codeex users, five times growth in three months and um 70 that's 70% month-on-month growth, which they gave. The only reason why I'm using it is because they gave double usage. And then the only reason why I'm using it a second month is because I've got a free month given to me um because I attended an event or didn't attend an event. So, I don't think their usage is as well celebrated as um as they think it is. Um so, uh cuz I think Sam Alman said something. He was like, "Oh, we're so excited about Maybe I can find it." Um just on here we go. Oh yes, it was resetting usage limits to celebrate 3 million weekly codeex users. We are set re resetting usage limits. We will do this every million users up to 10 million. Happy building. So yeah, but I think this 3 million has got to do with the fact that they've been giving away tokens, which we'll get into. Um, so yeah, but once I' once once my usage is done, I'm out. Like I just like cord. I do actually like having both to be fair. I do like having both, but unless I'm building heavy, which I'm not doing like heavy heavy stuff um all of the time. I'm building my own tools, I'm recording videos, doing video work and whatever. Like coding is is probably 50% of what I do. So, I'm not heavy heavy heavy. However, when I am heavy, I chew through all of the my usage. I'm I'm using up my codeex and I'm going back to Claude using my codeex going into uh Yeah. Yeah. So either way, I'm I'm just leveraging free free usage right now. So I don't necessarily think it's anything else other than free usage or free tokens that's meant they've got such an uptick in users. Um you mentioned uh Jevans paradox as well. Joven's Jevans paradox. Uh yeah, that that was in kind of like it was about the RAM actually. I added it there because uh first we we wanted to talk about this in kind of in the context of RAM uh not the prices are not coming down as people thought they would sometime uh because of the war partially but also even though there was a Google I don't know if you heard about it there was a Google uh paper which came out um about basically compression and allowing models to use six times less RAM. Oh, we spoke we spoke about it last week, didn't we? Briefly. I think we did. Yeah, briefly we did probably. So, it's about that even even if the models are going to use less RAM, I think we're just going to increase the context. Yeah, probably. Um that's the joke paradox, right? Is yeah well well the idea is that uh you would think if the price of something um reduces you would use uh you know less of that resource but you actually use more of it because people um because it is more available. M yeah of course right of course we we all so in this case yeah it's about RAM if the price of RAM is going to come down it doesn't mean that there will be um you know companies buying less they will buy more and if we can uh if so the Javven's paradox in this case is about the also what Google has done if we need less RAM for our context we are not going to use, you know, to reduce our context. We're just going to have more context and the optimization is going to allow us that. Yeah. So, yeah, that's why it was in the notes. Yeah. Well, I mean, you mentioned it earlier as well. I will get my You don't understand this whole system. Ah, it was not a um it was inspired by the uh Cold Fusion video, but it was not it was a different video. And I'll give this guy a shout out cuz it was a really interesting video. Where's Streamyard? Sorry. On um so was this video, Silicon Valley's billiondoll design scams. Now, this guy just talks about design. I have you ever seen this guy before? No, I have seen this guy, but I haven't seen this video. So, I this video. Okay. Yeah, I think he's good. Design theory. Yeah. Yeah. Um, but he talks about and this is this brings into the whole discussion I mentioned earlier about like uh Sam Alman being from Silicon Valley startup. He startup through and through, Y Combinator and all the rest of it. and um and it and also about giving away free usage. The thing is is that um because I I've just been thinking about this and the the typical thing about startups and we've all seen this is what what what do um Netflix and uh Uber both have in common is when they started they all started with like really small plans. I think Netflix was like $5.99, $4.99, something like that when you first opened it up or the digital version, you know, uh cuz they did they did DVDs in the beginning. Um Uber, like uh there are like nothing. It was nothing. It was always cheaper to get an Uber than it was to get a a regular old taxi. And what happened a few years later on when you're hooked when they've penetrated the market, all of those prices went up. And what um design theory speaks about here and and in some ways cold fusion is that like AI is in the same space in that you know we're just getting given all of this free stuff effectively free Google especially like I don't who's paying for Google who's paying for Gemini I get I I've got Gemini and everything you know and I don't I don't know who's paying for like I I I maybe get charged a dollar or something like that. But it's like I think even in the Gemini CLI, you get like free usage of of Gemini. I don't know. Anyway, ChachiBT is another prime example as well. Giving away free usage. Sam giving away resetting codeex limits to celebrate 3 million weekly users. Like he's only winning this race because he's giving stuff away because that's all he knows. And when you compare it to something like Claude who is anthropic who are expensive, it's something I've said for a long long time where I think that they are the real they're the real ones. They're they're charging realistic prices for this um for this for AI. And I don't think it will be long before chatbt they once they've penetrated the market once they've gained that well they've lost it, haven't they? the the the verb uh or noun. I don't know what it is, chat GBT. It once once they've if they get that kind of throne of like everyone's just automatically thinking chat GBT, I reckon those prices are going to go up. Again, leaning to this idea, I don't think Sam is a trustworthy character. I think he's just that he's just a cringe incarnation of a Silicon Valley uh CEO. Um, yeah. So, yeah, that I strong I do strongly recommend this video. It does break down that it's not just I don't think he even mentions AI in this in this one. I think he mentions a bunch of other stuff, but this is a they they do this time and time again and we constantly forget that this is a tried and tested method for startups. They give it to you. They spend investors money. They give it to you for as cheap as possible. Once they've got you, they they [ __ ] shaft you, you know, they raise the prices. So, just have that in the back of your mind. And when we talk about RAM and prices on everything coming up, I think this is when we'll also start to see the price of AI starting to increase. So, I've got my eye on you, Sam. I don't trust you at all, you scummy bastard. But yeah, um I'll leave links to that down below. I think I think definitely watch it and and the Cole Fusion video as well because I think I watched them both I think I watched them back toback so they're kind of all blurring into one video so I might have just jumped around a little bit there but watch them both. Right. Do you have something to show? I do. I do. I have a bunch of things to show. Where do I even begin? Gosh, it's it's so cool. Um, by the way, my energy was a bit low in the past 10, 15 minutes. I'm trying to regain it. I did. Um, so thank you for carrying the show so far. I went to bed at 6 a.m. So that that's why um clauding that that's the verb. I was I was cla and web flow and gapping. Um, okay, cool. But I have some some like really cool news for the web for people who love design and web and all that. Cool. Let's hear it. Uh All right. So, uh I have a uh I have an intro. I don't know if it's good or not, but I'll try to go with it. So, uh let me have everything ready. Yeah. So, creating 3D websites, making them accessible and SEO friendly has been pretty much impossible. That is until now. We have something really cool called HTML in canvas. It's a new API. It is experimental. It is not final. Uh and look all the things people are uh doing with it. And I have a few example of my own. So this guy built an X-ray in a way so you can see through your HTML tags. Uh, and just just to show you what I have before I show you the other stuff, this is what I have. Canvas meets DOM. So, previously you couldn't put HTML elements inside of a canvas, but now you can. And all of these, you see, these are selectable text. Like I can literally select this text. Uh, and let me even show it to you here in inspect. You see I have a canvas. I can open it. And here I have a section. And here I have an H1. I have a paragraph. And all of that. And all of it is animated inside of a canvas. Here I have another example. Here I have another one. I just asked Cloud to use the API to make a few examples. And um let me see actually if this one is also Yeah, it is also selectable. Same with this one. I know these are ugly. That's not the point. The point is that we can now render these elements inside of a canvas tag which was not a thing before. Uh now the examples that are people creating are really cool. Obviously one of the first things you do with a new tech, what do you do? You try to run you run Doom. So this guy has HTML in the Doom. So in the game you have these UI interactive UI elements and this is a form with name, email and you can interact with it and it is a a form. This is an HTML form that exists inside of the uh inside of the game and yeah it's fully interactive. It's accessible well access machine readable. I know it's very pixelated. That's not the point. But yeah, a really good example. Another one is by uh Wes or Wes. He's been on this. I've seen multiple posts from him. It's it's really fun. This one is a to-do app that you get to destroy. So, yeah, why not? Uh, another one I saw from this guy. This was also really cool. Um, yeah. So you have these real HTML elements that you just normally interact with, but you can now add these effects on top of them. Um, which you couldn't previously. Previously, you had to uh I don't have like extensive knowledge here, but I know that you could import things, but you would they were not they wouldn't be real DOM elements. they they would be like you turn them into uh pixels essentially in the uh in the canvas but now they are actual elements uh and 3JS has support for it. Actually one of the examples that I showed uh you is a 3JS example. You can here again the point here is you have this 3D element that is moving but you have actual HTML elements inside of it that you can interact with. Uh this is big. This is this is going to be huge for the web and for the things that you can do uh like effects that you can do on top of everything. Here I saw also this post. These are real WebGL shaders on top of interactive dove ele DOM elements without the HTML in canvas API works in Safari too. There is uh rough edges but still um worth sharing. So apparently this API is not and was not the only way to do this. But um yeah, I I suspect that this is more like a layer on top um and probably like doing some mirroring um uh like a there's a pipeline for seeing what there is on the screen um what elements the DOM is essentially giving to the browser to render and you're running that pipeline through a shader. I think uh that's how it's working. this one. But I guess the API is just it's kind of uh big deal. I would say I I can see a future where you know more website and especially website builders and now with AI can have all of these cool interactivity without the trades tradeoffs that we had before. you know, if you had an H1 in a canvas on a website, that H1 was hidden from Google, but now it's not. It's it's just an H1 that exists in a canvas tag. And that begs the question that I asked before the show as well. I I would like to dive deeper. I haven't had enough time to understand the fine details of how that works. Does that mean that now we are rendering everything with a GPU? Does that does it mean whenever we put something in a canvas tag, it's going to be rendered via the GPU? What if we are doing CSS animations which are typically done um with your CPU as or like the JavaScript animations that you have all of those are done with your CPU. That's why if you do if you add too much I don't know like uh too much Gap animations even your site slows down but webgl driven websites of more often than not um they can be faster at least the animations can be faster because they are accelerated by the GPU. So, does that mean that we could technically some uh technically wrap an entire website like the entire everything we have in the body in a canvas tag and render everything in a canvas and all the effects that we want to add to add them there. I don't know. But that is what uh I will I will try to figure out this weekend. Um it's not entirely true that the Gap animations actually do leverage the GPU. Um really interesting. Yeah. And um because they animate the some some CSS properties is actually CSS can leverage the GPU. Some CSS properties like 3D transform actually use the G. That's why um there are certain uh properties you do and don't animate. So padding you don't animate because that doesn't leverage the GPU. It it it create it. Ah I see it makes the browser repaint. Whereas um transform opacity um gosh there's a there's a bunch of sense that that are more performant. They don't cause repainting and then they leverage the GPU as well especially 3D transforms. So, but saying that, you know, yeah, there there there are probably more complex things we can do that are more performant just because they're in the canvas. I think it's worth exploring the the performance implications of doing stuff in the canvas. Maybe the same animation in the canvas versus doing it in regular old Gap. Do you know what I mean? Yeah. Yeah. Maybe that that that is going to be very interesting. And this is uh I see like there is a lot of learning for me to do. Um potentially yeah uh I I will look into it. I will try to understand the API as well. By the way if if you want to do this you have to be on Chrome. Uh it doesn't work in Arc for example. It's um it's hidden behind a feature flag. So you have to go to Chrome flags and then you find the HTML in canvas. It's called canvas draw element by the way. Uh HTML in canvas. You enable it, you restart your Chrome and then you can you can have that. Yeah. Yeah. How interesting. How very very interesting. Yeah. It'll be interesting to see where this goes. I don't know. I I have I have my doubts but you know beyond fun experiments. No, this is this is too fun. This is too fun. I want to render all my future websites like this. By the looks of things though, you can't scroll the website. It's literally just the top of the website. Like you can't navigate around, right? You can't scroll down this page, right? Well, I don't have much more content. That that that's a good question. It didn't it didn't Well, looking at the examples earlier as well, it didn't like you could actually scroll the website. I think it's literally just you get the top. So, that's that's interesting. What Oh, the the section has a position 100 uh position absolute. What if I just literally duplicate it? What is Do you want to enlarge that HTML so we can have a look at the HTML? Yeah, it's not it's not scrollable. So the HTML it's it's what you expect a canvas tag, right? And then it has section in it and then it has H1 in it. So this is all canvas. Yes. Yeah. It's a canvas. So if you if you So if you select the section, select that section and command C, copy it, and then select the element above it. Yeah. And then paste it. Command P. And then select the element above it again. Yeah, I do. Select the element above it again and hit it again. Command P. There we go. V. So now we've got three sections. So we can't scroll down to get those other sections. Yeah. So the whole idea of wrapping everything in this wouldn't work. I I see though. But like can't we register the the the scroll on a different element and then apply it transform that inside of the canvas? we could but yeah I'm not you know uh WebGL and canvas savvy nearly enough uh to uh I'm just like literally experimenting with AI but I I will go learn about this. This is this is way too fun. Nice. I I was already like playing with these things by the way. I was literally building uh give me like I was literally building some other experiments. Wait wait wait which one type this one. See I I was doing this for a project and this is not a selectable HTML. I had this issue already. But now you and out of nowhere. Yeah. Now I can have this as a as an actual That's the more interesting part, isn't it? That's the more interesting part that we're not It's not about putting websites inside of the canvas. It's about putting registered um you know uh accessible content inside of a canvas. It's not just lost. It's not like an it's not like um the Flash player that is just that's an empty box of of of content. It's now readable. Yeah. Cool. Exactly. Cool. Wrap it up there then. Well, yeah. Yeah. Okay. Okay. Well, that's all we got time for. So, make sure to follow us at Command AI Show. If you're who's on YouTube, give us a like on YouTube. That'd be lovely. That uh and a subscribe actually while you're there and all that good stuff. Join us next week for all the news in AI design and dev. I've been Sam and I've been Kavara. Keep on vibing, baby. Peace out.