#68 Rogue AI Agents, Gemini Argon, OpenAI Dev Day, Claude Goodness, Google Launch Chips in Space

Okay, we're going live. We are live. Congratulations. Hello there. We're live. We are live. I am um I'm I'm flustered. I'm flustered. I just I released a video and it was a mess. The edit was a mess and uh there was glitching as well in it. QA. Who's to blame? Um my agent, of course. Yeah, my agents to blame. Where's the QA? Open AI is responsible anthropic but you know ah I'm not I'm not an open I've been using open AI loads for reviews this last week so I chuck it a review code review because I want a different you know I want a different um way to to think about a different brain a different brain that's exactly it so let me anchor myself here free brains everywhere yeah so anyway but it's about Jev. Oh, it feels like it needs to be re-recorded completely, but maybe I can say in the edit. We are already uh we have a few people in the comments on Instagram asking if we have access to Argon. No, not unfortunately. Not yet. I I can't wait to get my hands on it and just burn 1 million tokens in one go. Ah, I really need that for my sloppy gym app. I did not realize that that was going to, you know, get so much attention so quickly. So, yeah, Argon is not available. I I didn't even I thought we were going to read the article together. See? Yeah. Well, people got it from the from the title. So, welcome everybody. Uh, let's let's actually get started. So should I go with the I'll go with the intro that I have not written but I'll just go through the I mean we talked about it. So this week we had Open AAI's dev day. We have 1.2 billion OpenAI weekly users. That's crazy. With a marketplace coming up we'll talk about that. And OpenAI dots dot like you get a dot, you get a dot. everyone get a dots. Um then Google came out of nowhere saying like hey we are alive we have Gemini for Argon but you don't get access to it because it's too good it's too dangerous. That's the game everybody every Yeah. Classic people access it means that your model is shitty. So you better not give people access. We'll we'll see about this but we have some benchmarks to look at. Anthropic launched clot.dev. We'll have a look at that. Uh we had a few other Anthropic updates as always. It's always Anthropic with their updates. Clot sonnet 5.5 uh which is cheaper and faster but I'm not going to use it because it's cool model but I move on. Uh SpaceX actually launched Google chips into space. I I guess they're testing that. will look into that. And then we have some juicy news about how we are going to die uh by 2040. And yeah, uh that's kind of like the list of everything that we will talk about. This is Command AI. I'm Kavarza. And I am Sam and this is a week in AI design and dev. Cool. Subscribe. Subscribe. Join us on YouTube if you haven't already. join us on YouTube. Um, it's a much better experience. You will see our faces in 4K there. Um, compared to Instagram. I'm talking to people on Instagram. Um, yeah. Cool. So, where do we start? We should we should start by I'll I'll show you. Oh, I'll share my screen. Oh, goodness. I'm just clearing up my dashboard. Desktop. I've got my desktop. Should I start a mess? C. See? Yeah. See, OpenAI just had their dev day and they introduced dots. And if you go to dot.com, you get Oh, you get Crockpot. They got bamboozled. This is just so funny. Okay, so we'll get into the Open AI news, but man, when you come up with a name with a brand, the first thing that you do, you buy actually there are jokes about not buying the domain. But typically what people do is that they check the domain and they buy the domain. But apparently all the super intelligence at OpenAI didn't help them to do that. They introduced dots. We'll go through the the rest of the stuff as well. These are really actually cute to be fair. Very cute agents um that will do stuff on your behalf. But apparently the domain uh.com was owned by Elon Musk and the troll he is he just pointed it to x a.ai/bot which I think is really really funny. Uh anyway out of all the people in the world uh this was in the hand of the right person to bamboozle them. Anyway, so there were a lot of stuff that got introduced in the dev day and the only I would say well maybe two interesting things were dots and maybe the last thing that they showed. So dots these are agents that have their own computers. This is the natural next step for agents. You give them their own computer on on in the cloud and they can do stuff on your behalf and they don't wait on instructions apparently. You give them goals and they I don't know maybe maybe once um we try them out we can talk more about it but can I say hey get me more clients I don't know will it work like that? Yes yes yes it will. Yeah, it it it will go uh literally hack people because we'll talk about this, but this is apparently how AI is working. It will hack people and I guess um will make them my client. We had GPT 6.1 soul. Have you tried this? Um I am. Yeah, I've been doing code reviews on it. So, I use it for code reviews on high mode. Um, but nothing for anything more than that. It's uh it's just I just whack a pull request in there and say review this and it comes back with stuff and I put it back in Claude. Yeah. So, Astra being the best model. Now, we have Soul, the second best with a new update that is apparently better. Um, we have the ultra fast mode which is very interesting. Um, and now this comes with a hefty subscription, a new tier for 500 bucks. And apparently some people bought this and they hit their weekly limit in 30 minutes, which is like really crazy. F pay 500 bucks and it's just just like that it's gone. Uh, I I mean that's the thing with really fast models, you are now running out of tokens even faster. But like, can you make decisions that fast? I'm genuinely having a difficult time keeping up with two Opus 5.5s for two different projects. And normally I do like five, six, so many different projects at the same time. And just having two with Opus, I'm having a hard time because it's responding fast. It's really fast. Sometimes I'm like, "Did you actually do it?" Because it's just like this. It does it. It It's really really fast. Um, and maybe partially I feel like that because I'm using a lot of GLM 5.3, which is really slow, but you know, it's cheap. Anyway, uh, we have private intelligence, which is cool for businesses. A lot of businesses have these requirements. I'm actually in the process of like uh buying uh graphic cards. I'm like talking to my client for a specific client because they don't want their data to go anywhere. So businesses actually care about this and it's cool to see that they respond to this. But honestly like I feel like they came up with so many new tools or like new features that it's like slop to me. It feels like they are just throwing something at the wall and trying to see what sticks. Codeex in the uh cloud. Didn't we have this? I guess we had sites but not like codeex running in the cloud. I Yeah, maybe there's a relaunch or a re thing or maybe it was never Codeex that runs in the cloud. I don't know. Yeah. Uh yeah, then we have the codeex CLI refresh uh code review. This is just a view that you can have uh like what you like you can do code review directly inside. I mean previously you could do these things as well. I feel like they are just turning so many things that you could already do or like you could already set it up for yourself. Now they're making it to be a part of the product and it's again to me it's like adding a few new buttons and see what sticks and um in sometime maybe it will be removed. Uh Codex uh security cloud. I don't know what this one is. I haven't checked this one uh out much. They came up with to be honest I'm I'm not I'm not trust we'll go over it later on but I'm not trusting Open AI to give me a secure cloud. Yes. Yes. Yes. Yes. That's actually a really good point. Yep. I mean uh yes we will go through this in a in its own specific video in a bit but they are not just bad at this. They are like really really they are like surprisingly bad. And it's all so funny because now AI is called SI super intelligence. It's not artificial intelligence and anymore by the way. Um buy some domains some SI like actually one of my teammates he messaged me and was like have an eye on the SI domain. Si I was like smart that's smart. I haven't bought anything. Yeah, because super intelligence. Yeah, it might become a thing. I was gonna say is SI a what do they call it? A um I forgot what they call it, but no. Yeah, top level domain. Yeah. Uh I think it is. Yeah, I think it is. Uh and all of these I'm not sure if all all but I know most of these twoletter tldds are from countries. some like from Caribbean countries and sometimes you get issues with uh like IO was dying. I think we covered it. Um either way, have have an eye on SIN. Wasn't it Trump that came up with this? Well, he probably didn't come up with it, but he didn't come up, but he he he literally signed and like I don't know is it like called executive order, whatever. He signed it and the United States of America is uh has a typo. It's just so funny. Oh yeah. Yeah. Yeah. So it's truly What was the name of that movie? Um like Icracy I guess like yeah it's literally that happening. Um yeah, so we have the SI domain and it might become a thing because of also this what do you want to call it like this ladder of they they are trying to just up the game for the sake of upping the game like there is nothing like specific but they need to to go with the momentum. So artificial intelligence is not enough. We go with SI. Anyway, they came up with decisions API and this happened rather fast. Like there was no time between Jeff and this I feel like it was like last week, right? So Jev was a decision in behind. I say was because now open AAI has their own model and I think this is just going to take to have the adoption because it already exists in your subscription. It's uh already with the same API u dashboard you control everything from the same place. The documentation is probably better. So yeah. So, did it say anything about the model? Like, is it a new model or they abstracted that away? I'm not sure if it's a new model. Uh, but what's interesting I I think I have a I think I do have a use case for this. I'm making I I was making an app for my dad for translation and I had issue with formatting. I don't know if I should use that anyway. Um, I'll I'll try it out. So, I'm happy that I have a use case so I can actually see if it's useful to me or not. I don't think this will be good for formatting though. This you you not for formatting for choosing like the the correct ones. Um the yeah like my other AI is coming up with actually a JSON list of uh tests that I design but sometimes they are wrong and I don't know if this model can I could do this through my model as well you would take his natural language response and determine whether he has answered the question correctly like um yeah that that's one of like I have it both ways So I I need to control what the agent output is correct grammar because sometimes they suck at creating tests and I also have to when my dad answers I need to um to know for sure if he answered correctly. Yeah, it depends what questions are because if it's like if it's like multiplechoice like doingo then you don't need AI to to control that, right? you just did they did they select the right thing or not but if if there is like fluidity in the answers if the response can be a long fluid thing and in amongst that interpretation they would have answered the question then that's where the decisions API would come into into play yeah I have both I have both yeah I know like for the deterministic one obviously you don't need that anyway so Next we have agents API with computer use. Did Claude have this? I feel like it had I'd be very surprised just because Claude computer use is so bad. Um but they have an agents API but whether it's got computer use access I don't know or how this all ties in together I don't know. Then they have bedrock managed agent powered by OpenAI. Yeah. So basically you I'm guessing this maybe again I would have assumed that you can you can run open AAI on Bedrock already but clearly not then they announced plug-in extensions which is interesting because there are like 1.2 two billion weekly users. And with the plug-in extensions, there is this idea of creating a plugin and potentially selling it or like having subscribers like a marketplace type of thing. Yeah, I don't know. We'll see if it sticks. They yeah because they they released that thing where you could effectively build like a little web preview inside of chat GPC and that it you know it sounded promising and you'd think that oh you get to basically build apps within chat GBT and then chat GPT becomes this master app but I don't know about you I'm not a massive user of chat GPT but I'm not really seeing anything really take hold in terms of yeah operating within chat GPT I think a lot of these things that they announced and what we will also talk about they require some sort of community. Yeah. Like it only works if they are like actually like if people are hanging out there um pointing to web flow like even with the web flow made in web flow community people were like checking it daily. I was checking it daily. So people were hanging out there they were commenting they were they were like like system. So when there are people there is discovery there is like this care and then you can also have like obviously sales and whatnot but I find it to be difficult to to stick or like to have actual revenue if yeah maybe maybe it also can work without a community maybe just like a tool list that you use but if you can build any tool why would you pay for one I yeah I don't know Maybe is it like going to be that advanced each uh plugin that is not easy to create so people will end up paying for it? I'm not sure. Uh yeah, improved plug-in creation and now sites can host plugins. This was also interesting. So you create a plugin with chat GPT and you can put it inside of the site. Yeah. uh which now which comes also like with automation like the MCP and the the plugins. Then they introduced chat GPD space a new home for your team to collaborate with AI to get work done. Create a dedicated space with where teammates CH GPT and your dot can build a shared knowledge. Chity can so this is notion essentially. Well, didn't we report on something a few weeks ago where was it Spotify that created this this basically where it's like Slack but built with I think it was different a bit but learning interesting it is interesting like they're like again at this point you see why I said this feels like slop to me it feels like they are they they they have unlimited tokens and they were like okay let's build notion in GPT let's build this in GPT let's build it's like let's every app is trying to I don't know maybe next week they will have like Netflix in GPT I see I see the reason for this though like you know if you're yeah like you say notion was the sort of bare bones version of that where if everyone's working in notion everyone's got access to everything there's this shared knowledge database or shared knowledge space experience or whatever. And if you can house that in an in an AI that is doing presumably learning, presumably um assessment on everything within it, plus the ability to say what was that report that was created when we said X Y and Zed like just to have just to have all of that instant access without having to drag and drop files on ChatBT without having to build your own wrapping system or build your own effectively rooting system where you stick an AI in front of your Google Drive workspace as an example. I see I I see the reason why you would want it. It just it will just be weird to to to know if businesses actually adopt this, you know. Yeah. I I think it has to do also a lot with um how dependent is this from the chat GPT account and like organization setup. like do I need to like can I share it easily with you as an outside person? Yeah. Still collaborate with you. So if that is difficult, yeah, we will see about this. But on one hand, notion has AI in their you know in their tool and now OpenAI has notion in theirs. Uh and we'll see what people will use. I'll I'll probably will give it a try, but let's see because notion is quite limited. Yeah. By the way, we we're literally just responding like I didn't I didn't watch the dev day um Oh, yeah. presentation. So, I actually don't know what this is. I'm just guessing. Oh, okay. Yeah. Um it is like notion from what I saw. So, they have pages which is just like Google um what's it called? like Google Docs and you can do like multiple people and chat GPT. Um again available to all pro business and enterprise plans. Uh now you can create slides and you can do it in collaboration. H a lot of these things I I see them. I see them and I'm like this is cool but I I kind of I I'm past this mindset of getting myself stuck there. I see this as actually this can be just a GitHub repo between me and my teammates and we can work on that with whatever AI that we want, right? What what's the difference between there? Like why do I need a tool for that? This looks like real time. This looks like multiplayer they call it nowadays. That's quite cool. Fair. I I think I saw uh I saw a repo, a git repo. Maybe I should give it a try that can add like you can build on top of it and then you will have real time collaboration for anything essentially. I should give that a try. Yeah, I should. But good point. Good point. That that's a good point. Like you have like comments but I'm past that. I'm I'm like uh this is just I don't want to get limited in the mid midway, right? That's my issue with it. Uh with my own slides, I made my own slides. I made a for the Web Flow Conf. I had a controller on my phone. I could control my own slides with my phone. I I I I really loved that. So there was no limit because I was building just on just with HTML, right? Um anyway, you have teams and shared tasks which is also interesting. It's kind of like is it like a bit slack like like you can chat and have uh tasks not very interesting. And obviously now you have HH in Slack just like Claude that you had in uh in Slack a month ago or so. You have a meeting plugin. The meetings plugin takes notes for you and saves personal personalized summaries with action item uh items in chat GPT space. So yeah, this kind of like notion you know notion has a meeting app as well. I feel like every app has an meeting app nowadays. So this is just one of those and sharable profiles. I this is kind of like why I mentioned like the the community aspects of it because I feel like they want to create this this community. I saw this and we talked about this with um a few friends yesterday like I don't hang out there like people are on Instagram and they hang out and I think a really strong thing about Instagram is you send like funny jokes to your friends and you have like these group chats and you have like internal jokes and whatnot and all of that creates this sense of community and it's it just doesn't I don't know This is I guess it's cool but it I don't know why why like why would I need this? What am I going to do with this? Finding clients. What is it exactly then? It's sharable profiles. I guess now you have like a public profile you see like to to show what to show like the tokens that you have used. uh peak tokens, longest task, current streak, longest streak, and then you have showcase, which to me looks like a few uh web projects. Basically, your AI slop. It's an AI slot, a social media for AI slot, basically. Cuz I'm not being funny, boys and girls. I don't give a how many tokens you've burnt. Like, it honestly is not one of it's not it's not impressive. Like if anything, the less tokens you use is better because you the more Yeah, exactly. better at prompting you are and stuff. But so yeah. Okay. The social media for AI slop. Let's go. Yeah. Uh I I don't know what I will use this for like h and I feel like as a designer developer, I feel like oh no, another profile that I have to create and you know take care of and like my link stuff. Yeah, it's just not. Anyway, we have this one which is really cool actually. Sign in with Chad GPT. And it's not just a normal sign in. So, you can build an app, let your users sign in with their Chad GPT account and use their own subscription in your app in opposed to you paying for them. Oh, I see. I've got you. Yeah. Yeah. So your app relies on chatbt as the inference engine. Yeah. So it ties I like the idea from a business standpoint but like what if I want what if I don't want to use we have I guess you have to build that sort of routting in place. If they're signing with chat GPT then use chat GPT as the model otherwise use ah good point. Well, maybe your app needs to, you know, ask the user like, do you want to because this says use your chat plan. So, it's going to be asked uh at the sign in and there should be also probably a setting that allows you to change that after you sign in. Interesting. But also, what happens as like fall back? So, there there are a lot of things that we don't know how it will be. We we'll we'll try it and we'll see if this even stick because like they are desperately trying to to to capture so much of the market. They want to literally just be this allencompassing data consuming machine like Google. It is. Yeah. I mean that's what you do when you have unlimited tokens and no responsibility. Then we have the new pro tiers that I just talked about. Pro- 500 costs 500 bucks a month. I really don't want this to be uh kind of like a norm. I I don't Anthropic don't get any thoughts. Uh or any other lab. I don't know, man. Why? Yeah. I mean it's really expensive and and it's um 25 times the chat GPT plus. I thought the 200 for uh anthropic was 20 times. So, so for Anthropic, the 200 bucks is 20 times the the base pro and Chet GPT Pro 500 is only 25 times the the very base. I don't know how they are trying to justify this. Maybe this is just the sign that we've been all waiting for that these tokens cost way more than we are currently paying for. And now here we are with the marketplace that I talked about like with the plugins. Yeah. Uh eligible enterprise customers can apply part of their existing OpenAI commitment towards approved partner software giving them more flexibility in how they work with AI. Our first 32 partners include Figma for creative work, Adobe uh blah blah blah and yeah I I seem to remember Claude or Anthropic having this for as well. So again there's a I I don't know I don't know whether again I could use Rep I could use my Claude subscription on Replet or something. I don't think we understood it then either to be honest, but I think it feels like this is the same thing. Yeah. Um either way, that's kind of like the next one. And that was it. That was pretty much everything. I I think what we talked about at the start was like the funniest part with the Elon Musk redirecting dots.com and the 1.2 two billion weekly users is genuinely impressive. The $500 a month proteier probably not worth it. Uh I don't know why you would buy this. Yeah. How much of a rush are you in to to burn 500 bucks in 30 minutes? I uh Yeah. Yeah. I mean, I imagine the people who are like thinking they're building the next startup. Do you know what I mean? Some Silicon Valley bros, you know? I, you know, I won't name names. Well, I don't actually know their names, but I do remember be going at the Open AI um replet meetup thing here in London and the sorts of people I was talking to literally thought they were building the next big thing and they thought they were just like, "Yeah, I've got to got a launch like and it's just half of the people that I talk to are like I have this thing I'm working on but I can't tell you my idea because if I tell you I'm like bro if you are operating at that level you're already like 200 years behind like what do you mean you have an idea you can't tell I think it's I think it's introduced AI has introduced so many people to building things They literally could not fathom the thing they're building has probably already been done or someone hand Yeah. But like we had even before AI, we had this before AI, right? I had clients telling me, well, not clients. Typically people like this are not clients. they are not you know advanced enough to to do like serious work but they come to you and like I have this idea to build this app to to do this business like actually one of really good things that you can do is to put out your idea and validate it or see if somebody else has done it and they are successful. you see that as a good point and you copy them actually shamelessly because that's kind of like an you see that it's tried and proved. Um we have so many examples of this uh in tech and I mean we see that every single day with these tools. Anyway, just been charged another 90 quid from anthropic. One more month down. Yeah. Anyway, so that was a long one. That was a very long one. What are we doing? The list was really long. It's just so so many um dumb stuff that they had. Anyway, uh well, I wanted to do I want us to do the I I thought we would do just Gemini for Argan. Well, let's Yeah, because I'm I knew about it and I saw the Do you want me to share my screen or you sharing your screen? I can see it on Yeah, you can share your screen. Um, if you go to deep m deepmind.g Google I think I'm already there. Strangely, it's not on Gemini, is it? News from Google. I'm just here. So, Gemini 4 argan delivers frontier performance in complex workflows. Oh, you can't see my screen. Sorry. Um, Gemini 4 Argon delivers frontier performance in complex workflows across real world software engineering, enterprise knowledge work like legal and finance and cyber security defense. Great. Um, what should let's just go right to these these benchmarks. Yeah, it's so good that we don't have access to it. That that's kind of like the point here, right? It's so scary that they can't give it to us yet. Well, because I mean yeah, I mean I'm knowledge work, Val's index, automation bench, finance agent, and harvey's legal agent bench, right? It's killing it in all those. I've never heard of any of those benchmarks to be fair. What constitutes knowledge work? I don't know. But Deep Sweet, it's already killing Opus 5.5, which is the the the big boy at the moment. the the most impressive model. Um Astra falling slightly behind, but it is killing those and it's smashing uh Fable 5.1 by almost 10 points. So uh where are we here? Astra's the big boy in frontier. SWE Vibe Code Bench cuz that sounds cool. Uh 91.9 whereas these are scoring 90. It's doing good. You know, seemingly it's doing good. I I mean what? So, you're saying, "Oh, I've just been bitten or something." So, you're saying we don't have access to this now? Yes, you're so great. Yes, you're so great. Leading in defense of cyber security. Do we have a price? Do we know what the price is going to be? I mean, I'm just not even not even that interested. Okay, here we go. We built Gemini 4 with frontier level capabilities on coding blah blah blah blah blah. You said all this before. Developers, professional enterprise tackle most difficult problems. We're grateful for the initial cohort of cyber defenders and trusted testers who real world whose real world evaluations and feedback will help us strengthen our systems before we release to developers, enterprise, and consumers. starting with paid API customers on Google AI ultra subscribers. So if you're on their I think this is 200 bucks this this ultra plan. So you do have it if you're on the ultra plan or if you're an API customer. Now I looked on open router. It's not on open router. So I'm guessing it's on their vertex or something like that. I'm just I'm just opening up Vertex now. Uh Google Google um oh sorry Google Gemini Enterprise Platform. I'm just signing in now to see if we can we can have a little bit of access to it and check that I'm not doxing myself or anything. Okay, let's have a look here. So, we're in the platform models. Model garden. I'm guessing this is where we can Can we have access to it? Um, what was it called? Argon, which is Argon. I mean, Argon isn't coming up there. Gemini 4. Search for Gemini 4. Let's just search for Gemini because I might be in the I don't Oh, here we go. Gemini. Gemini. There was a four, but like we need four argon. Yeah. Yeah. No, it's not there. Um but I Yeah, I just don't know whether we are or it maybe because they did say end points. It's annoying. Um they did say end points which I don't have set up. So anyway, I mean I feel like they've had too many chances now. Google I feel like they've had way too many chances. They haven't exactly impressed anyone to be honest. I think they I think they I was always impressed with everyday usage of Gemini 3. I think it was Gemini 3.1 or something like that generally of like just basic chat you know basic just question answer crap but anything outside of that which is what we're tending to move towards which is like again knowledge work and things like that it was never it never was any good at coding you know um yeah but I I'm getting a a taste for bad models such a bad bad way to put it. Um I I'm using GLM 5.3. It's not a bad model, but it's not a frontier and it's it's slow. But again, comparing to this, this is probably going I know it's it says like it's fable level, but I I don't think it's going to be fable. I don't really think that. Maybe at best maybe at best it's sonnet or opus. I guess it will be between opus and sonnet. I'm just like guessing. I don't know. And if it's cheap like it's really cool to use models like this for computer use for like similar tasks where we're like I don't uh I use my claude or chptd like proper models for like proper development and then these cheapers cheaper models they can do some some of the general tasks. I started doing this with um fable writing specific MD files, putting it in a folder. I don't do skills uh really. I don't I don't know. I feel like I like just MD files better anyway. I mean, it's the same. Anyway, I put it there and then I let GLM 5.3 to do something with it. So, uh and I'm getting pretty good results with uh some video editing and some computer use. Yeah. So, um I built an app this week that relies on Gemini on Vertex. Um it is a it's a document scanner. So, it just scans OCR scans documents and gives me back dates and things like that. Um I opted to put that on Gemini, but Gemini Flash like something super cheap, like readily available or that I'd hope it to be. So like I do see a place for like Google Gemini and and all that sort of stuff, but I just don't see them as like frontier. I don't see them as leaders. I see them as a cheap way to get stuff done to be honest. Yeah, it's it's a bit crazy because Google owns all the knowledge and all the Yeah, they own everything and Well, you say that and this is the thing I was thinking of the other day. It's crazy that they are so behind when they've got they were they've got so much data. You'd think that they would literally like Anthropic had to steal from books and you know opener had to like basically crawl Reddit for its knowledge base whereas Google have real people's data like deep knowledge of people's work and how yet they can't train a model. I mean they have we're talking about Google they own YouTube they have Google search console they have analytics so they have like Google ads and like just I'm just thinking like the data of people putting Google ads like data in all the YouTube like everything they have access to the you unpublished YouTube videos they are they have the meta uh data around these things, not just the published data. I it it seems like nobody has as much data as Google. I mean they literally index everything, right? So they are the indexer. I I don't know how they how they are so far behind. But maybe maybe we don't know this. Maybe it won't matter like maybe it is uh not the point right now because nobody is making money. Chachi PT Anthropic none of them are making money. So but you were you were saying that there is a 1 million output token on this model. Oh yes. Um we we almost forgot about this. So something cool I read about Gemini for Argan is it can output 1 million tokens in a go. Normally when you spin up an agent it does something and the output tokens uh for all of these new models is limited at 128 128,000 tokens and then it has to stop. It can't like go for further. it you have to spin up a new agent um or a new conversation. With Gemini uh Argan, it can apparently do 1 million. That's eight times more than all other frontier models. But what do you need this for? I really don't know. Like it's not the same as being smart. But maybe if it's smart and if it can work this long, maybe it can do the one shot of pretty much anything. It can bring it further. Uh that's that's my assumption. I mean, at the end of the day, I want my responses to be concise. I don't want to be reading through shitloads of text in the responses, right? Um, we'll get into that in a bit with the whole Well, we we haven't specifically said we'd cover the webdev stuff, but we should we should cover webdev and Theo, right? We should chat about that. Yeah, we should actually. That's actually a good chat. Yeah, maybe we transition to that now so we don't forget about it. But yeah, I I want to be able to read the response and I don't want to have to sift through a million tokens of response. Yeah, but it's not kind of not the point, right? like the the response can be as short as possible, as short as you want. You can also add that as a but it can work for longer. That's the point. I I was literally thinking of a 3D game because I in my experience take the longest when I do a test or something take the longest. So maybe it can just build and build and build before it has to stop. That's kind of the idea. But we'll get our hands on it. We will test it and we will let you know. Yeah. So, yeah. Okay. Let's quickly go over that webdev video. I'll get it up on my screen now. Oh, yeah. Let's go for some drama. I mean, we're talking about Theo here. Like, what the hell does that even mean? Um, let's go here. Add a stage. So, WebDev Simplified released this video where he calls out Theo specifically. I mean, let me refresh that page. You can There it is. Brain rot with a picture of Theo there. And this has absolutely Hello. This has absolutely captured a lot of discourse. I Let's get Theo's response up. Yeah. He basically um calls out Theo in one of his latest streams or videos uh where Theo is quite brazingly to be honest saying this is how you you're using AI wrong. This is how you use AI. And in this video, um let me get it up. Let me find this thing. And in this video, this is going to be obviously I'm I'm sort of in the middle a little bit, but he he speaks about the responses that he gets from Claude and how you know you you want to basically take the response and feed it back into the AI so that you can understand let's go this so you can understand the responses a little bit better. And this is basically a a a dual kind of point he was making. In fact, let me reshare this. Um where I think his name's Kyle, he's talking a lot about this book, A World Without Email. Mhm. Which talks about uh basically just handing off knowledge work, handing off thinking and how this is just not that we'll lose the ability Wow. Yeah. I mean, or the OG one. The interesting thing is Yeah. Um I'm trying to read the subtitles there of the So, the idea here is that with this email um with what like the analogy he uses. Um you you don't make any decisions. You you you you're supposed to have a meeting and you write an email and you say when do you have time and nobody like gives suggestion. Nobody is doing anything productive. You are just um shifting away the work from yourself. You're pushing it to somebody else to another level and that is creating this chain of unproductive. I I guess this is the OG meat proxy. You you just shift it to send it to the other person. H Yeah. Here we go. Yes. I swear I was I didn't Yeah. I'm sure this video will result in totally rational and chill reactions on Twitter. So, with that being said, this is Theo's response to the to the video. Yeah, starting I think there's some like there's some um reasoning here and some justification here, but we'll get into it. Another day of unemployed Kyle aka WebDev Simplified dunking on me because he has no real career experience. His main complaint here is that I show multiple instances where I didn't know the depth of how part of the code bases work or codebase worked. Sounds bad until you realize it's a code base with over 300 competitor contributors, millions of lines of code. Doesn't matter how much AI is used, nobody knows the implementation details of all the code in a real world codebase like this. The example we highlight is one where I was unfamiliar with the details of how tagged skills are attached when we pass prompts to Claude. So, I'm guessing this was a T3 code. um uh review or feature and they've got to interpret your use of skills inside of T3 code and say well are you using codeex or are you using claude and send that skill to them uh in the in the format that they expect. So there's a bit of a tree branching thing. Um, funny enough, I know a lot, probably too much, about how we tag skills and how messages are passed over SDK. But this one intricacy was new to me. Knowledge gaps like this are the norm in real work. Learning learning to operate with partial understanding of the system is key to long-term success as an engineer, regardless of AI usage. I'm genuinely scared that inexperienced devs like KL are going to trick well-meaning juniors into holding back their careers. Now, um, he didn't understand in the video. Theo did not understand the or or chose not to try and understand, which he admits to here that this was an intricacy that was new to him. Yeah, he still kind of he he asked AI to summarize it and be a meat proxy in reviewing because it was a code review, reviewing the code and and sending that review back to the original developer who was developing that feature. So, he was a meat proxy in in that instance. And I don't know, I I'm sort of in between because I'm like on the one hand like we should absolutely be uh relying on AI to better ourselves, to learn and and it's going to get better. As we said earlier, you're going to get people who don't know how to do stuff. And AI is just going to take care of absolutely everything. right now we're it's still kind of necessary to know how systems are put together and various things like that but until then um until we get to a point we we have to be very specific about this sort of stuff on the other hand you know there's lots of juniors there's lots of um people learning coming to Theo's channel and saying well how do I learn how to do dev how do I learn how to do use AI and and Theo's there promoting this idea of don't worry about it, run it through AI, let it do all the problem solving for you. And it's like you're not you're not going to learn like that. And he sort of kind of goes over it a little bit here. So he's maybe contradicting himself um a little bit. Um I was going to cover one more thing. So if you let me just compose my thoughts a sec. Go on. Sorry. Go on. No, it's fine. So I I feel like it was a little bit of like uh the contrast between you and me. We had a series of videos where you were teaching me how to do a dev with AI but not just just to be clear for the audience not to vive code but actually do dev like from a dev perspective. And the main issue was always me not reading the plan or not reading it thoroughly and not understanding what I was reading and saying like cool go. Uh which is always the case right? If I'm building something where I'm out of my depth uh obviously I can either sit through it and try to understand it or give the go. In my case, I I feel like uh a lot of times is I try to get an understanding but just broadly and it's not always the case. I have to admit it's not always the case. Uh if I build like a landing page, if it's anything like front end or like websites, I I do understand quite enough for that. But when it's when it's a bit more complex, I I might actually let AI to do the decision making. And I think it will come back at some point as you said like if you just say yes yes yes at some point you made some decisions and you have to face them. And these decisions might be structural. It might it might be something that hits you with a bill with a 2K bill actually. like these things can happen if you um be become a complete meat proxy. So I get the point of like I get the idea why you should understand but I don't think Theo was being a meat proxy to be fair here. I know like I I don't think I think he absolutely was because he did a code review but did not understand the the the response from the AI. He literally did, "Let me let me Google that for you or let me chat GPT that for you." He copied the pull request. He read the response, didn't understand it, asked AI to formulate that in two-way response that he copied and pasted to the the other person, you know, the the original developer. So, he was literally a meat proxy. That's in its I see I see actually I remember now that point. Yeah. a layer of my I was I think I was defending um Theo in one specific area in his tweet that you there will be some knowledge gap like if you like even back to the time that I was building with web flow I was doing things that I would not understand but then I understood them eventually and I'm not saying that the same thing will happen to AI That's actually my fear that you do this so much that you you will never understand what you're doing. That that's kind of like the problem. But isn't it normal to be out of your depth and actually try things where you don't understand them, but you still can copy stuff or like put things together without fully understanding how they work. Well, exactly. And I think that there in lies the problem. So, if we just chuck down here, you got Brad from Traversy basically saying the same thing um that I am in that ultimately the message here. I mean, Theo is a little bit butt hurt because he's being targeted and that's quite unfortunate. I I I feel bad for him in that respect. But he is putting himself out there as a leader and then demonstrating a very poor form by not doing what you just said, which is this skill of reading and trying to understand. Theo didn't even try to understand one of them. He was like, you know, there's there's too much text here. And actually, uh, to Kyle's argument, like there wasn't a lot of text there. It would have taken him 30 seconds to read it. Interesting. If he didn't understand it, fair enough. Like, yeah, I agree with you. There's not you're not going to understand every little thing, but at least attempt to do it. Don't just meat proxy yourself and just be that's too much for me. Send it to summarize this and send it to, you know, send it to the to the person. So the focus here really and the focus of Carl's video although it comes across as quite harsh is let's let's let's still try and understand because this is a that can only be a net positive on the world that you at least try to read and understand that you're not just this slave to yes accept except accept trying to be better and you know we we had the chat before the show about What what and maybe we can talk about this now, but what do where's the missing gap of people's understanding and you know they're see they're seeing a lot of people build applications they're seeing a lot of people launch websites and thinking wow they've done that with AI how the hell can they do that and it's because we've learned it's we're building on top of separate individualized learnings over time that allow us to be able to launch an app or a website. It's not just something that people have come to just that's compound knowledge. It's compound knowledge. It's not that's very rarely does someone with zero knowledge hop into clawed code or hop into cursor and build themselves an app right that they pro that most of these people probably should use something like lovable something where there's an an infrastructure around that where you don't have to think about the little intricacies like database or authentication or hosting as simple as hosting or CDNs you don't need to worry about all that but By launching an app or a website within a a system like that, you'll start to pick up on things. You might hear the word CDN, as you say. You might read a response that says about a CDN, but all that's done is plant a seed to when you start seeing it over and over and over again. You might be inquisitive to look into what a CDN is, as an example. So, it all stems from like the willingness to learn and and grow and and I feel like Theo's video was wasn't exactly promoting that, which there's nothing like why would you not promote someone to try and try and learn at least attempt to to read the response, you know? No, I get it. I get it. uh Theo's respond also was a bit around like real world world knowledge that I want to also mention like in a real world scenario you might build something for someone or for yourself and not understand everything at that time but I fully agree and I think it is good to even if you say yes to claude um five times in a row maybe you come back at the final once you have the final solution and you'd be like I actually do this uh sometimes when when it's not too difficult for me to understand like how did you do it actually I do this as a like I want to understand what happened um and not just it like partially because if you get you want to try something design something get somewhere and it takes you some tries it takes you some hours of work and if you learn how you did how you got there. It will help you to get there next time much faster because now you know the at least at the very least the terminology. You at least understand the dare I say the architecture to some degree. So, um yeah, I agree. It's good to learn. Yeah. And and dipping back into Brad from Traversy's uh comment here and I was in agreement to be honest. It it can it it could have just as well been said without attacking another creator. You know, he was using Leo as a launch mechanism. That's Yes. From YouTuber to YouTube. I mean, we we know that how it works. Like I don't love saying Webflow is dead every week, but kind of like works. We're not attacking other creators, though. We're we're Yeah, fair. Talking about another business, right? We're not saying uh you know take Timothy Ricks as an example. We're not saying Timothy Ricks is stupid or you know he was wrong or anything like that. Yeah. True. Yeah. True. It it would it would be like me looking at a video from Timothy Ricks and like because he makes a mistake or something and then I'm being mean to him. That that would be quite me. I mean uh we know in our community to do better. So maybe maybe the broader dev community can learn a little bit um from the Webflow community that doesn't exist anymore. I can't help it. Um I'm literally just going to watch the first few seconds of this discussion because I was going to make a point and I said I was going to come back to it. I knew I'd forget. I've forgotten what that point is. So let two seconds. Yeah. I'm going to I'm listen going to listen. So since he's watching this uh where are you joining us from? Why is YouTube so quiet? So actually not as quiet. Uh what topic today? It's literally in the title. Uh I'm probably a bit too late too late to reply to that comment. But the the topics we mentioned them and right now we are talking about some webdev drama uh with Theo none other than like when we talk about drama Theo is probably involved. Is that mean? I hope it's not mean to say. Anyway, he's close to drama. I'm sort of listening to two people talk at the same time right now. It's very confusing. But I don't think I'm going to I'm going to have to add it in the edit. I'm going to have to remember it in the edit and then add it in the edit, which we do, you know, with all these videos that we think are going to do well. Hey, Claude, remember this. Yeah. Yeah. It's so annoying. I should have I should have put a pin in it. I should have said it out loud and said we're going to come back to that, but never mind. Um Yeah. I am but ultimately you know um I've been getting a lot better at sort of reading the responses and trying to understand the responses if not just to grow my own knowledge because I think we're all again why it's a net positive to promote reading the responses um no harm no worries if you don't but you're going to do yourself a disservice um you know getting back into reading responses and learning something new and trying to learn something extra through AI doing all of this hard work that it's it's coding better than I ever could, you know. Um I'm just able to wrangle it together and put the systems together right now. So there's plenty I can be learning from from it particularly around particularly around GPR actually and data data protection and things like that. But Oh yeah, that's all boring. Pain. Yeah. Yeah. Um anyway, drama. That's what that was. And we're here for it. And we're here for it. Yeah. Please continue with the drama. Um, yeah. I think that's all we've got to say. I think that's all we've got to say on that one. Yeah, just double checking everything. Um, yeah, I thought Theo was pretty I thought Theo was maybe particularly a bit harsher in his response. You know, could be. Another day of unemployed Carl dunking on me because he has no real career experiences. Yeah, that that's Yeah, inexperienced inexperienced devs like KL. So, I should be sharing my screen here so you can read along with me, but genuinely scared that inexperienced devs like Carl are going to trick well-meaning juniors into holding back their careers. It's like now you're getting a little bit personal. A little bit personal. Yeah, I I feel like this is more personal because what Kyle did was uh he pointed out something specific in a video that Theo made. It's very specific. It's yes, calling him out, but calling out something very specific. It's different from like generally attacking somebody like this person bad. Um I mean, we know Theo Theo is a bit emotional with uh his responses. Yeah, there I say. There you go. You must get it every single day. So anyway, yeah. Um, let's let's migrate. Talking of migrations, I don't know. Um, where should we go next? What's next? I can see you're you're on SpaceX, but do we talk about CL? Well, we have the anthropic stuff, but I I I don't know much about I don't know where to start with it. Well, of course, we had Sonet 5.5 release. That's all well and good. Maybe we can start with the the model card there. Yeah. But to be honest, I love all their little intros they've got here. Ah, the their their intros are so good. The videos. These are real videos, right? Can you can you like play it again? It's they are they are just so good. They are so aesthetically pleasing and the the use of these round shapes. See how everyone's going into space at the moment. I mean Chachi BT Open AI specifically, but you know. Well, literally. Yeah. So, introducing Claude Sonet 5.5. The second model in the Claude 5.5 family. It's a clear upgrade over Sonic 5. Yes, absolutely. runs 30% faster and costs 30% less for most work. Now, I mean, I don't know. It's all well and good, but I just I sent a tweet out today saying like with with client work, I need I need to rely on the absolute be. There's no room for error when you're doing client work, right? You you want the absolute best of the best. I don't know whether I want to waste tokens on my max plan even on something like Sonnet. To be fair, I do use Sonnet for one thing and one thing only, and that's the titles to these this podcast and the clips and the the motion graphics I use for my own channel, my own videos, cuz Sonic was enough, right? Opus just felt like I I literally went from Haiku and was like, "No, no, I kept up. So, let's use Sonnet. And it got it right and it was going right. So, I was like, why why am I using Opus for something like that? It's unnecessary. But, um, realistically, I mean, this is a surprise. A Gentic Coding Terminal Bench, it's scoring better than Opus 5.5. This is very surprising. But everywhere else, obviously, it's not scoring as much. It's still impressive, but I don't know why why you would use where would you use it? Yeah, I just don't know. I just don't know. Even Haiku, like in my Jev test for my personal channel video, which will hopefully go out. I don't know when. I've screwed it up. But even that, using Haiku, which is their smallest, fastest model, you know, that's that's potentially where I'd use Haiku and like really easy simple decisions like I'm just I just need you to do this one thing for me and and you know, not think too much about it, but I just can't think of where I would why or where I would use Sonnet to be honest. Would you would it be useful for writing? Uh, I mean I'm I'm just saying something, but like potentially potentially. Um, you don't need many tokens for writing, right? So why would you not use Fable? Well, I really like Opus. I like the tone of Opus. Now, maybe if I'm writing something that doesn't require a lot of research and and whatever, maybe if I'm very speer almost artistically in the writing, maybe Sonet's a good model. And you're right, I might try that actually because I'm trying to explore writing more and writing my thoughts and ideas down more than obviously using AI to express them, but just working hard on a written piece over and over again to like make sure it's done the way that I want to do it. So maybe Sonic would be a good good model for that. But, you know, I don't know, man. I'm I'm seeing these limits and I'm not even coming close. I'm barely using Fable for anything now, to be honest. Um everything's working the way I want. So I don't know. Yeah. The next thing uh they have uh that they launched in the same week was claude.dev. And it's pretty cute. Look at this website. Um it's a new hub for developers. I mean I love these clawed animations. Like look how cute is that? It's really cute. And here I guess you have access to the knowledge. You have the um clawed docs. This is just a clo code.claude.com. So, it's just a claude code documents. And you have this terminal thingy here where I think you can just use the website through a terminal like UI. I can do like theme can switch the just I don't know what I why would this but it's just fun. I How you get into that? Oh, yes. term. It literally says terminal. You just press T uh as well and then you can do Yeah. terminal stuff. It's It's really cool. Like I can do posts. Yeah. Um can go into a post and then I'm out of the terminal. But I wonder why they're pushing people to use the terminal. Like I know Claude, I love the terminal. No, it's just the UI. It's not Yeah, it's just the UI. But I I have stopped using terminal as well. I gotta say, but I I genuinely love their design and their branding. They are um doing such a great job at it. It's just so clean, so beautiful. These animations. Yeah. Um well, that's kind of it. Do you want to move on to mods? Because I really want Do what what have you seen about mods? cuz I've got the claw code docs and we can infer from that but I don't imagine well here are the mods so getting started with claude code mods um you can you can actually look at this do you know about this is it is it worth doing an intro for it I I thought like it we would do it in the same place okay cool if not just record an intro at the end of us going through this so once you know what it's about then you can say this that you Um, I have a difficult time understanding it. Like you can change the look of stuff. Mhm. And for the terminal, that was for the terminal. And here for claw desktop mods, add PES, commands, and tool call rules to claude code with a mod. That's what it's saying on the docs. Add PES. So, is there an example there with a pain? So, I'm going to try and get a release about it as well. What changed there? I'm not sure. I I should have um read through this. I haven't. So maybe that's something else from what I saw on Twitter. Customize UI, swap your own features, write one with a few um or write one with a few lines of TypeScript or have Claude build it for you. Mods ship inside plugin. So you install them with the plug-in. Um so customized UI. I wonder what people have been doing with it. It it can draw its own UI. It says in the terminal or the desktop app that's kind of like views by web flow source, you know, you can create custom UIs for the data that you have. Okay, so here's some examples. Should I shall I share this for some examples? Yeah, of course. Yeah. Um, let's do that. Got it. So, a token weather adds a forecast to your context. So, where is the weather? Can you see it? Oh, there it is down there. See? So, it's there is there was a config file that you can do and I've I've got the git branches and all the rest of it in my config. So, I I'm guessing it's some form of extension right there. Um, but it looks like it's called on demand rather than um um so we probably install it. Yeah, I'm not sure if it's it should be persistent. It just makes sense if it's persistent enough or like at all. Um otherwise, yeah, blast radius catches risky shell commands like remove or get reset or a force push before they run. It shows what the command would touch in a side pane. So they've done that. Claude's going to call this call this and then it's going to bring up this little command. And isn't it like kind of like artifacts but now kind of like artifacts? They're part of the response. Uh Chad GPT had that as well where you get visualization in the reply. It's kind of like UI building run replay. So it's a skill that actually steps back through all of the the commands that it actually ran. That's quite interesting as well. less of a UI, but I wonder what design control you get over these cuz these look a little bit ugly to be honest. Yeah, they they look they're like HTML 1.0. But true. Either way, I don't know. I I'm trying to think of what I might use them for. Claude modes. Uh we don't know what we use them for, but they are here. And let's let's just see. Let's just see what people do with them. Yeah. Uh was there anything else on Claude at all? No, we have two topics more. Okay. Okay. Okay. Two two more topics. Uh we have the I'm not sure about the basics one. I haven't maybe let's go into the FTC the 2040 plan and talk about that for the next 20 minutes because it's it might get a bit meaty. Yeah, that website. Um, let me bring it up. Where to start with it though? Because realistically, I mean, I watched the the Cold Fusion um video this morning, right? And in this video, I can maybe share my share my screen with it. In this video, I mean, there's a face reveal, which is Let me tell you a little story. In late 2026, standing in New from uh what's his I've forgotten his name, but there's a face reveal in there. Well, he has shown his face before, but now he's doing it like at the start of the video. Got it. Got it. lot of newsers are doing this. So the AI industry is a complete mess. And it basically talks about how the sandbox, you know, the hugging face incident that um that OpenAI where the AI broke out of it sandbox and hacked into hugging face and and whatever was just an a a badly controlled experiment. Completely what's the word I'm looking for? um irresponsible way to operate a cyber or internet or tech company. And he quotes here um both the primogen um around uh got it around here primogen here um and the internet of bugs here. So go watch those videos where they effectively uh talk about the same sort of thing like this is a this is not a new problem. This is something we we've been doing since the dawn age of the internet is when we run these tests we we operate them in a sandbox and we watch them like a hawk. And he also talks about generally the Do you want to explain the Australian government website as well that was hacked in in January or something? Yeah. And figured it out. Exactly. Open AAI hacked it and they knew about it and then they sent um an email to a very like public email address uh to the government. basically being like, "Yeah, we hacked you." But not actually like following up and being responsible for it. Uh, and that email never got checked. Well, not never but for month. Uh, I think it was like for four months or something. So almost for half a year that email was not checked and then until got checked and they they they had to see if it's actually for real. Uh all the while Chach was chu OpenAI was playing cool. They were meeting with the government officials and not telling them we just hacked you by accident. So not taking any responsibility for any of that and it's just a very crazy stage that we are in right now. That's kind of like the point. It it's a bit hard to even explain or understand how did we land here where these companies do harm and then they talk about it. They make a press release about the harm that they've done and they are like let's make someone responsible. Like brother you are responsible. You made an agent. You gave your agent a computer. You gave it a prompt to start with it and then it did something harmful. It's crazy that they're not held responsible and they act in a way that they shouldn't be responsible for it. And we as like collective humanity have to be ready for uh rogue AI for their basically product being buggy. That that's kind of like is the actual terminology here, right? The AI is buggy. That's that's the problem that they have to solve before giving it before bringing it to the public or even before giving it a computer with internet access, right? And we see that they are doing none of these things. They are bringing these AI tools to the public and they are also giving it computer access without proper uh tooling around it and they they are not like looking what is this agent is doing in I think in one instant that it was uh shown they left it to run over the weekend without like even checking like you know how you don't ship on Fridays we we know that for websites like you don't ship a website update on Friday because like everybody goes home and if there is a bug there is nobody to solve it over the weekend and you have a bug until Monday. You don't do that. Generally speaking, uh these companies are doing the these uh this exact thing with AI agents running over the weekend without anyone looking over. So, it's crazy that they are not held responsible. And he I think he even calls out um says like OpenAI need to be jailed like they are they're running reckless um whatever. Well, it it was Edward Snowden saying like oh yes it was Edward Snowden, wasn't it? He literally said, "Yeah, Sam Alman needs to be in jail." And in, you know, we we covered the pace in the frontier last week, which is generally a good direction to kind of go in to have people reviewing the AI for people to catch up and various things like that. But um it's it's not good enough from like uh from a from a technical perspective because you had that guy who it was open AI right where he left or no he's anthropic for like six months who has been open AI for years and years and said look I'm out because these companies are absolute charlatans and they're running a mockery they're not doing things safely and all of this is coming to a head right now and it's it needs to be it needs to be handled they're moving so much faster They're they're running before they can walk and they just don't have the expertise. God knows why they're paying them this kind of these kind of salaries that keep getting float around. But these are supposed to be the best in the business yet they can't operate a simple sandbox experiment and and identify when the the the AI has broken out of its container, you know, and just yeah, again set something off for the weekend, you know. Um they they have super intelligence they say. Like can't they even like put another AI in charge of that? Like at the very least like you you could do Mr. Misix's uh type of thing. Yeah. Exactly. Right. It's it's crazy. So then you you saw this Naval tweet with AI 240, right? Uh well, it's kind of like separate things, but yeah. uh the but I feel like they're talking about the same thing though which is which is we need to we need to catch up like if I remember a in that list it's like a similar thing but we can do it separately if you want but I just I do yeah so Neville said it the best I think it's the shortest and the best and super clear the best way to pace the frontier is to hold the labs fully liable for the behavior of their models uh for dangerous or uh weekly guarded open-source models hold the hosts liable. Of course, the end user hold liability as well if they are jailbreaking models to do harmful things. So here we're talking about like uh three levels of uh liability for like three different groups. One is the the labs. They are making these models and they are giving it to people. But before they give it to people, they test it. And if they test it, they should be liable, held liable for, you know, their testing in they they should do a proper testing environment. They should have maybe government agencies um looking over. They should have like maybe third party sources looking over. So that's important. And then the open source models, well, we know they will have hosts. The same thing is happening with pirating movies, right? That the hosts are held liable, right? Uh the same thing is with not safe for work. So we have figured out a lot of these things. It's not like new. These companies are acting like this is a new thing. It's not uh the product is a new product, but the concept is the same. And obviously the end user should also be holdable. Maybe if I use a model and jailbreak it or I find a way to use that model to to do harm to scam people, right? Maybe the model cannot even know that. But if I find a way to scam people, I'm liable there. But you know, it's a chain and it's crazy. Wasn't Wasn't there with the Australian government, wasn't there a public uh page that was made public or whatever that was showing personal health information and was that something that the AI exposed or was that just something he demonstrated where it's not only the AI companies that are being the issue was that even like the government agencies they are not doing a good job. So there were bugs that uh were known and government is not doing anything to solve them. And in by the way in that video he showed like he said two things about like two contracts like a company got what like 70 $80 million for a website for a government agency. Supposed to be 3 million and then when it came to it it bloated to 70 million or 90 or something or 90 Yeah. And then they got a second contract after that for 117 million or something. Crazy. Exactly. It was 117 to fix the things that they got wrong in the first one. It's so if you want to make money, you got to get some contract uh contracts from the government. I mean, that that's kind of like how the whole US system right now works. It's it's a different topic, but that's that's kind of like where the money is right now with uh in the US. Anyway, um should we separate this from the other one or Well, I thought it was relevant and also tangentially the tangentially um the only tangentially is the touring test video uh thing that where does responsibility lie? Is this something responsible? But you know, oh yeah, that that video is like awful. Let's let's again let's go through the next bit and if we think it needs its own video then we'll do an intro at the end with this the AI 2040 I don't know how to transition from what we talked to this don't worry about that for now let's just just go over it and then we'll figure out the transition if it needs it so all right so there is this really cool uh and interesting website with uh some really cool content I haven't read everything here but I'll just show you around there are it's called AI 2040 and there are possible routes that we as collective as humans uh will take and to to basically see what will happen by then there are these different alternatives uh there are uh how many four five plans we have plan A plan B plan C and plan D and as we go through the website it's really nicely animated and you can read through it. So, plan A is a verified slowdown. So, this is kind of like what happened with the cold war uh and the the atomic bomb the the race that we had there. We had some agreements. I know it wasn't perfect, but we had we we came to some agreements. So here it's really nicely also visualized that what will happen um as these agents get more capable and they grow in number. Uh so yeah that's that's the plan A that we probably want that right well we certainly want that a verified slowdown. Uh we have a plan B that is a bit more crazy to fight China. We have a plan C to burn the lead. I haven't read this one to be honest. I'm not sure what this one is. We must minimize the risk and uh maximize the benefits of AI. Uh the president says he will be implementing strong regulations to ensure safety. Burn the lead. Is it like stopping the leaders? I'm not sure actually. Or burn the lead. I I I haven't read through this. I'll let you go through this in your pace because it's very interesting and it's like in a chart but we have a plan D which is what we don't want to do but it seems like we're uh on pace to do this which is race to uh AI super intelligence. So AI will most likely lead to the end of the world but in the meantime there will be great companies. said by uh you're truly Sam Altman the only uh yeah and this idea is that there won't be a slowdown and all of these companies by we see here in the chart by 2030 we start maybe 20 between it's interesting between 2030 and 2031 we get to the AI that self-improves itself and it gets just ridiculously ly um better automated coder, superhuman AI researcher, top expert, um dominating AI, super intelligent AI researcher, artificial super intelligence. So that's kind of like early 2031, which we're not really far away from. And the idea is if we get to this, this is kind of like the end of the world. And we actually this one is probably the hardest to to predict what will happen if we race it to there because that implies that it won't be safe. It will be we will be in the mercy of the AI because we won't by at that point we won't understand how these systems work and there is a plan S shut it all down which is not going to happen. Um but yeah, I I will I learned about this website maybe an hour before joining the meeting call. So I will go through it and read it. Um um yeah. Yeah. So, I I I get it now because I thought this this article was talking about how we um what we need to do in response to slowing AI down, which is why I thought it tied into the previous section. Hearing hearing what you've just spoken about and reading through it now, it's not to do with that. So, maybe it's worth a separate intro to say what is this article trying to tell us? What what is it saying other than being you know what are these plans? It's showing us the plans and the alternatives. Let me uh read alternatives to so yeah go go for it. Go for it. Well the to what we have because right now it seems like we are on plan D. We are racing to super intelligence. Um that that's that's what's happening right now. All right. Um, okay. Come up with a Yeah. So, the the hang on two sec because I'll I'll give you some insight. So plan plan plan a this this website is a scenario-based policy proposal describing how humanity might avoid uncontrolled race to super intelligence. So I'd be interested to know how scenario D is. Oh, sorry. Yes. Yes. So the switching is as you say what is what is um what is happening or what what other alternatives but they want to point you they want to point you towards plan A. This is what they think they we should do. So it not only talks about other scenarios that are happening and that is plan B, plan C, plan D and plan S. Actually no, plan S is a poss. If you go to the little switcher thing, plan S is a possibility from from plan A. But B, C or D are what what what is happening right now. It's I don't really understand why they formatted it like this because I think it's making it really confusing. But ultimately they think plan A is the thing that we need to do to avoid the uncontrollable race to super intelligence. Yeah. So maybe it does tie into the previous video and that we say well this is what the this is what what do they call themselves AI 2287 this is what AI 2027 think we should do to avoid this crazy race. And that's what this article is all about. And it's like what is it? Is it is it a four step the authors argue that AI companies could develop systems smart other than roughly 7 to 20 years a closed plan the scenar like the timeline basically from 2027 to 202040 it's got various um scenario timelines and what what policy measures need to be put in place. So, I don't know. It's a tricky one. What do you think? Um, well, uh, what I wanted to say for like if I go with the intro, I would go with something like, uh, talking about the scenarios. Um, yeah, what are the scenarios with AI? uh we all heard that AI is going to kill us or you know bring the prosperity to the entire world but what are the actual different scenarios has anyone put these scenarios next to one another so we can see them and visualize them actually yes so there is this website AI 20 uh 40 uh which exactly does this we have these scenarios from plan A B C D and plan S being a part of plan A with plan A being a verified uh slowdown. This is us going towards AI super intelligence but in a verifiable uh slowdown manner so we won't end up killing ourself kind of. Yeah. Cool. Yeah. We'll see how that goes. And then you can let's fix it in the edit. No. No. I don't I try to actually. Yeah. Um, yeah, but I I I even thought like I could we could go another round and go for it. Read read through time. No, like to read through it. Okay. But yeah, uh actually yeah, let's let's do it. So just a continuous continuum of what I u said as the intro. We can continue now by So what's in plan A? So in plan A we actually have this slowdown that is happening in steps. Now I'm not going through the whole article. You can do this in your own time. It is very beautiful. It's very visual. Uh it has a lot of uh details and I will be reading through this over the weekend. But the idea is that um we we have to do this slowdown with other nations. So we have compute declaration. Neither country wants um to halt their own AI training unless they can see that the other side is halting it too. This literally happened in the cold war as well. The we had this issue with atomic bombs that at that time we could not verify that if Russia is halting it. I mean, we could do it with testing bombs that were uh tested uh on the ground, but we couldn't do it for underground bombs. Uh we didn't have the equipment for it. And I guess the idea here is that uh would with you will halt this or you can verify it by seeing how much compute a country has. And I I guess there there are ways to figure that out because of like the the production and the the compute that you need. Um then you have pause in training. Obviously this is something to be done and I I I kind of want to go in and right now uh read this but I'll just go through it. Get worldwide buy in. This is the job of the media. Maybe partially we're not the media but this is partially the job of us as well to tell you this is important and uh you should care why you should care. So that's the idea and uh plan A is established uh basically it is to buy time to figure it out because I I saw something interesting here um about safety. So you you want to get safety right that's why you want to pause it controlled explosive growth. Yeah there will be this growth coming but you want to have it uh controlled. Yeah, a lot of interesting um details. I'm not sure what they are talking about like the citizens divided most of this uh new found well is spent addressing soon to be raising unemployment. Yeah, like you know what's going to happen to the jobs? Uh so the there are there are many reasons why we might want to have this slowdown to not lose our all our jobs. uh mutually assured compute destruction. This is interesting. This is literally compared to the cold war and uh all the concepts that um h that yeah uh that happened between the nations there um you know m the mutual destruction that you can have with with atomic bombs. It's kind of like the same stuff are being brought up. I saw this image and it was like kind of like terrifying. But we kind of know that that the type of super intelligence that we are talking about that can take over will probably require like insane amount of um infrastructure uh way more than what we think uh what we have today. So probably that's the idea behind how we can uh assure knowing that when knowing if somebody is doing this or not if they are like cheating or not life after work. Uh yeah I can't get uh I can't wait to actually read this. This is it's quite interesting uh idea. So this is all plan A and it's it's quite detailed but the other plans are not as detailed. The other plants are rather short actually and they are not as as beneficial and they are actually quite destructive. I think the other one the other plans this is why I think it's really confusing they've called them plans because they're not their plans. They are they are eventualities and how they think they're going to play out. So they're not Yeah, they're alternatives to Yeah. what what can happen. So the whole thing is yeah choose a path. So we we we went through the plan A which was a path under right through the intelligent intelligence explosion by having AI's self-improve and putting them in charge of more things. So basically AI's getting better and and doing everything. So what? No, that's crazy. If we don't want this, let's make a deal with China. So, how do we make a deal? We do a verified slowdown or we shut everything down. That's a part of but we know that's not going to happen. But it is an alternative, right? Then um yes, and that's good actually. And or uh we have no choice. So what if you say well we need to pace it but we have no choice. So what are the other plans? Uh slow down at least a bit for safety and uh governance. So if we say yes and slow down China too, you have to fight China. I actually have to go through this. I haven't but yeah uh plan C burn the lead. Plan D, race to AI super intelligence. And this is the terrifying one. This is the one that is likely happening. And we have a quote from Sam Altman saying, "AI will most likely lead to the end of the world, but in the meantime, there will be great companies." H this is just horrifying. And it seems like we are doing that because nobody is slowing down. They're talking about it, but they're not slowing down. China is not slowing down. There are no like actual talks about verified slowdowns. And we know the current administration is not really great at planning and getting along with other countries. So that's the plan D, the explosion of um intelligence when AI gets better at R&D, gets better at making itself smarter. And that's kind of like the the plan that we want to avoid. This is what's happening and we want to avoid this. So yeah, thoughts. I mean, yeah. I um I'll be interested to know what they I I think we're all kind of thinking the same thing. I just don't think Well, I don't have much hope that they're going to Yeah. anything drastic is going to happen. So, you think plan D is going to happen and we lose the jobs, the destruction will happen. Yeah, I do. Yeah. I do. Yeah, but maybe it won't maybe not not necessar I don't think they are going to slow down. I don't know what the consequences of that will be, but I don't think I don't think they'll slow down. They're too driven by money and greed. Like as we've just spoken about the safety aspects of it, like they're just they they're doing whatever it takes to build the best models and they're sacrificing literal literal AI safety because of it. You know, they're stealing books. They're they're breaking laws just to get ahead. But one thing we need to mention it's all with the assumption that we will get to like actual super intelligence and we get there by AI getting better at R&D and improving itself because if we can get there by just you know doing it ourself it's it's going to be different because we can make sure that that we will have alignment we will understand these systems better and we will have a better chance at shutting it down. But all of this is under the assumption that we will have um AI that can improve itself in a rapid sense obviously that that's where the explosion happens. Well, we'll see. I guess we'll see. Um well, that was a somber note to leave it on, I think. Uh oh, we just got a comment in. Hi everyone. I'm Samar and and I'm subscribed to your channel. You cover topics that are very relevant for designers. Nice. Yes, I think we do. Thank you so much for the support. Yeah, for watching. Yeah, thanks for basically. Yeah, basically. Yeah, thanks for watching. We're We're done. Yeah. Um Yeah. Well, that's all we got time for this week. Uh, make sure you like and subscribe and uh, I guess we'll see you same time next week for all the news and AI design and dev I've been saying. Yeah, join us next week. And I've been Kabaro. Keep on buying, baby. Don't be a