#72 Gemini Agent, End of Grok Bot? DeepSeek Super Agent, Weight-Space Programming, Meta x Conductor

Greetings. Greetings. I've literally just realized right before I'm about to go live that I'm wearing a dressing gown over the top of a smart shirt. So, quite the quite the mix of attire there, but you know, we'll get through it. Um, are Kabaza is away this week. So, you have me on my own trying to juggle 20 things at once, which normally goes terribly. In fact, why don't I extend my iPad display and just seen you can extend onto Macs as well, other Macs. So, I've got a Mac running in there. Um, here we go. Right, I can see myself and your comments over here. That is nice. But it's on the wrong side. So, let's arrange that. Um, and as you see as well, I'm in a new space. You'll probably see me in this space for a couple of weeks from now, but I'm I'm temporarily doing some uh uh housesitting, let's say. And yeah, I'm I'm on the move. Let's try and get some of this bit more better. How is that? I'm going to do a quick audio check over on YouTube and see how we're sounding. sounded all right. Listening out for my taps on the on the keyboard. Little bit of rumble there. I'll try not to type too much because the rumble is uh is is real. But sounds good. We're here. Kabaza is not here. Let's um well, actually news. It was my birthday this week on Yeah, I won't say when. It was my birthday this week. So, I've I've celebrated a brand new year of my life. One more year down uh closer to the end. Um and I was on holiday last week. I went to I went to Monorca, which is in Spain, and it was super warm and sunny. And now it's cold, dark, and freezing in the UK, which is just I I I hate it. I hate it. I'm just looking at myself there, thinking, do I need to I want to centralize myself. Look at this little depth of field going on there as well. Um, but it's been a it's been a fairly it's been so quiet recently, uh, with AI news. Like, I mean, we We got some stuff to talk about here, but there's Yeah, there's really been a bit of a drought when it comes to the AI stuff. Uh, I've been like hard at work with a bunch of tools that I've been building. I've actually been learning about embedding embedding models like because you've got embedding Gemma embedding or bedding and Gemma, I don't know which way around it is. So, learning a little bit about embedding models, which I'm going to do a video on my personal channel about all of that. So really digging in and getting excited about all the local AI stuff. Uh what else has there been? I mean been working on a a project with someone um just today actually which has taken loads of my time which is kind of first time I've done some coding in a fair bit because I've just been making videos and whatever. But yeah, other than that it's been been pretty pretty slow. Regardless, I'm uh I'm glad to en enjoy uh what little we have. Um and happy to share that with you guys. So, let us know where you're at. Give us a hype. I don't know whether that or what that thing is, but give us a hype if you uh if you so wish. And with that, I think let's get cracking on with it. Oh, we're a podcast now. I don't know whether that's that was made known, but if you listen to us on like Apple or Spotify, you can just listen to us. It's like a cleaned up version of these streams. I think the audio is a little bit better as well. So, we're actually a podcast now. Everything's over on commandishow.com. So, you know, whatever you want to choose to listen to us on, it's all there um with that. So, that's really cool. And we're start I think we're starting to use Instagram a bit more as well for the shorts. So go check out us Commandio Show at Commandio Show on Instagram and all. So it will be cool to have you over there. Um and this this week would be no less. We'll be posting all of this stuff next week as well. So let's get cracking. This week on the show we have Google Gemini agent. I guess they're joining the party with proactive personal agents a little bit. I don't know. We'll get into that. Grockbot is the future of Grockbot in jeopardy here. There's been a few little clues about I don't know. Grockbot might be uh might be on its way out. We'll have a look into that. Deepseek opensources super agent. This one's really cool. This will be a more practical one because I'm excited to play around with this and if it looks really cool and you guys respond well to it, then I'll probably make a video on my personal channel because it's all in the local AI opensource wheelhouse. So, I'll yeah, I'll look into that. Sparks uh weight space programming LLM's editing another model's weights, which is just bonkers uh when we get into it. So, we'll look into that. Meta acquires Conductor. That's a really interesting one. And RIP Cybertruck as well. That's the end of Cybertruck. Just actually we we'll go over the [ __ ] Cyber Truck thing right now because it's just, you know, just some interesting stuff I've seen on on the net. Um all that and more in this week's episode of Command AI. I've been Sam and yeah, this is a week in design design. This is a week in AI design. This is a week in AI design and dev. Cool. Right. Yeah. Death of Cybertruck. Let's get this up on uh up on the screen. And you'll have to bear with me because once again I am on my own and I'm single screening it. So where are we going? Basically, Elon Musk sent out this tweet literally today. Um, Tesla will be winding down production of the Cybert truck at Gigafactory Texas around the middle of next year to make room for increased Cyber Cab production. So, that's it. I mean, to be honest, it was a bit of a disaster, the whole thing. I'm not I don't drive electric I don't drive Teslas. Um, I have never driven an electric car, actually. I really would like to do that, but they never made their way over into the UK or in Europe. I think there were a few channels. Oh, sorry. I'm not even sharing my screen. Let's let's redo that for the for the um for the clip. Tesla will be winding down production of the Cybertruck at Gigafactory, Texas around the middle of next year to make room for increased Cyber Cab production. That was today. That was this morning. So, jumping back, um, yeah, I I forgot where I was now, but anyway, point is Cyber Cab it I Oh, yes, that's what I was saying. There were a couple of attempts to bring it over to the UK, which failed miserably because it's just it's not a great system. It's it's like dangerous for the roads. It's it's dangerous for all sorts of things. um which meant it could never leave the US which classically as most of you are from they have very very much more relaxed rules when it comes to actually cars and you know even driving in general. So there could be a sign that no one at Tesla ever pursued getting these in the EU uh UK or EU. could have been a sign of um you know not a lot of demand for it. But I've been going through the uh the comments and stuff here and people I mean it's one of those things it's like the iPad mini. Loads of people loved love love love the iPad mini and loads of people say they love the iPad mini but the proof the the truth is is that people weren't buying it. You cannot argue with um you know people people's wallets and that's how we have a lot of power when it comes to you know whether whether the market or whether things actually do well is just by simply not buying them or buying them. It's no it's all well and good saying you love something and whatever but people aren't buying them. The data is there. These are businesses. They've got to operate. But there are some loads of diehard fans here that really love their Cybert truck. I mean, it is cool. Does look cool. It's not very practical. So, I've been watching on um sort of other channels and st stuff like that. Here we go. This is this this this sums it up perfectly. Who's upset about Cybertruck being discontinued? Loads of people holding hand. Who wants to buy one? None of them. Like, this is it. This is exactly it. doesn't mean to say that there's uh obviously people who well I guess people who have bought one who love them have already bought one but either way they'll be they'll be winding them down and that's the kind of end of Cybertruck. It was a it was a rocky road. We got they own the end but simply it just it from what I saw it just was not a good car. Not practical for a truck owner at all. It just looks kind of quirky. That being said, on the same lines as uh as Elon Musk doing a tweet, we might as well I know you know I love a good segue. Grockbot might be on its way out. And that is I I'll redo that one. Grockbot might actually be on its way out due to some clues left by Elon Musk himself. Now, it's a bit of a weird one. Um, hit this is the kind of chain of thought that this is the chain that I've had to go through to get to this. But it was a response from Doge Norway. Any chance of Grockbot uh reset Elon Musk? Lots of people say their usage ran out really fast this week. Now, and that that really is in of itself exactly what people were seeing. they were seeing their um their usage just like falling away over over Halloween weekend and a bit confused as to what's going on. I don't really understand why people's usage would um just be used up as opposed to just doing a like a reduced service type of thing. But in that tweet, the cost of serving Grockbot was unsustainably high relative to revenue. Now you could you could read that and say well is that an absolute statement to say it's unsustainably high to due to revenue in totality which could mean that they're going to be winding this down or doing something or whatever or was it just unsustainably high during revenue over that period? I don't know what it was about Halloween that meant that people were using this Grockbot more. But regardless, this is what Elon Musk has said. And it it just begs the question, what the hell are they going to do if if if the revenue just is not there for Grockbot? As we've just gone over, you know that Elon Musk is willing to just uh cull products that aren't that simply aren't bringing the money. However, people love Grockbot. I'm I'm not like I I'm still bewildered as to what people love about Grockbot. Now, I understand why, let's call them normies, right? I understand why normies might like Grockbot because to achieve the same thing as what Grockbot is doing, I have to set up a bunch of, you know, systems that read off each other. I've literally just set up. Well, I used to use Gem Gemini just because it was cheap to summarize my meetings. They're not very good. Like I I I then get uh clawed to look at it the next morning and it always mentions, oh, it missed out some key information or that it's really just talking [ __ ] on Gemini. But the point is is that Claude was just a better model for doing that. So I was like, you know what? in your Clawude subscription, you get a hundred or whatever your plan is, you get 100 bucks or 200, whatever, whatever it is in API credits. So I thought, well, I'm not using that ever. So why don't I just switch Gemma Gemini for Claude. Point is, this is an N workflow that um again only touches AI and little bits and bobs and and it's not a full AI kind of system, which The Grockbot, as far as I understand, it kind of takes away all of that setup. And and for better or worse, it's using AI in every single step of that um of that workflow. Whereas my def in my defense I think you don't need AI in those workflow which could you know going back to Elon Musk's tweet could be the reason why it's unsustainably high relative to revenue. Now the other thing is like I was always in bewilderment if we go into um grock what would it even be? I don't even know how to um let's go cursor because I know that that is something that they also sell the same plan. If you go to pricing here, it's not until the $30 premium plan that you actually get Grockbot. Now, I know $8 is not a lot at all and this is an absolute steal. However, it was behind like a 200 like super premium plan at one stage, but now they this is actually then bringing it down to this what they now call premium. So, yeah, it was it was under these under these ones, this ultra plan or super plan or something like that before, but now they brought it down to the $30 a month plan. So it always it from my perspective it always alluded to being this premium model this um this this service uh that costs a lot of money which is why they were you know hiding it behind those higher tiers. You have claude duet which we covered last week. I haven't played around with that. I don't know if that is like um if that is kind of gatekept behind the like $100 and $200 and now the $500 plan. I think it's $500 the ultra plan. I think it's only in the US as well. Let's have a little look at that. Um claw.ai because we haven't covered this one either. upgrade plan. Yeah. So, I I haven't got it. Sorry, just to you know, I haven't got it, but I believe they've they actually introduced a higher higher plan. So, anyway, um going back to uh Grockbar here. This was always an expensive thing and I don't know whether the others open is another version. They got their dots as well, haven't they? So, um I don't know whether these just cost a lot of money to actually run. And I I don't know. There's some really cool stuff. We'll get into it actually the the deepseek one, but there are proactiveness is definitely a huge thing that a lot of people are looking out for. Um I'm doing a lot at the moment with regards to like I'm just getting busier and busier. You know, we talk about AI taking away or giving us loads of our time back, allowing us to do a lot more and giving us a lot of time back. I'm literally just filling my time with more things I can be doing rather than just sitting pre and be like, "All right, that's that's a week's worth of or even months worth of work done in in a matter of hours." But the point is is that I'm sitting down and I was talking to a client today actually. I was sitting down thinking how what can I offset here? What can I move to a proactive agent now? I've got a an email, like something that sorts my emails. Um, yeah, because there's just so much stuff today that I feel like, oh man, I I need to have something more proactive actioning this stuff. But simultaneously I'm very conscious or very anxious even that I um don't want to lose sight of stuff. So I was dealing with uh um contracts today specifically a lot of contracts that I'm I'm wanting to make sure I mean I'm not I'm not reading the entire contract. That's that's for my AI to look at that and and raise anything to me. But I have to understand what is get I'm getting myself in for and this sort of stuff. So maybe there's a hire there that I could make that they could then leverage the AI. Um but there must be some but that took away so much of my time to do coding and to prepare for a video. It the proactive nature of these of AI is is a massive cell right now. A massive cell. So going back to Elon Musk's tweet about just covering the the cost of serving Grockbot being more than the revenue they take from it has to it begs the question there's only so much these companies can actually sustain um or or subsidizing the cost of all this stuff. So it's pretty it's pretty bonkers really. Um, yeah, that's kind of all I have to say about it. I'm looking through the the tweets here. Maybe we can share that with you. And yeah, so this person's on the $200 a month and they've used up 100% of their of their plan, which I'm assuming is is Grockbot. I thought it was the only one, too. Usually there's plenty to spare. Yesterday, 30% vanished in an instant. 96% used it's yeah these must use a lot of tokens a lot of tokens so my you know my advice is if you are like somewhat technically savvy and all the rest of it then definitely look into how you can build these things from separate components because it allows you to be more flexible then you don't you're not bound to everything being inside Grock, you know, you can even write code to or your agent can write code to do that sort of stuff. And I I kind of do it all inside of NA. That's kind of my uh most advised substitute on something like Grockbot. But yeah, crazy stuff. Crazy stuff. Um, there is actually I mean I'm trying to look for segways here. I don't know. Anyway, what do you think? Do you use Grockbot? Do you use uh it heavily for a lot of your tasks? Are you using it proactively? because uh Kabaza took us through a few use cases last week and uh a couple of those were literally I I thought you could probably do those with Grock build. It was more using Grock as a very, let's say, forgiving model. Claude has a ridiculous amounts. In fact, we can talk about that now because I've been so frustrated with Claude recently where you ask it to do stuff and there's so many um like guard rails placed on it that it doesn't do stuff and it doesn't proactively ask you if you want to do stuff. computer use and browser use is one such case where I I've never got it to do stuff in the browser successfully. Whereas uh whether it's the model or the the let's call it the harness itself of codeex anytime I ask codeex to do stuff it nails it. I set up a whole email system yesterday um just by asking codeex to do it and it it freaking nailed it. It needed to set some things up in Apple Mail, so it needed the application there. It touched the browser and did a whole bunch of stuff in the browser. Claude failed at the first hurdle when it comes to that sort of stuff. It's so freaking annoying. And I don't know whether it's because clawed or anthropic are being clawed or anthropic and putting those guard rails in place or it's just rubbish at computer use. I just don't know. I honestly don't know. But Grock is one of those models that actually will, as far as I'm aware, I've never really had a lot of time with Grock. Just does what you tell it to and doesn't really have a lot of guard rails in place and doesn't winge about security or privacy or or whatever, which is probably why people like Grockbot because it just gets on with it, you know. Regardless, that's my mass that's my biggest pet peeve with Claude. My current setup is claude uh chachbt $20 and chachbt is for computer use and code reviews. So I always get it to review my code. It's sort of linked to GitHub so that every pull request actually gets reviewed automatically by chatbt. So quite a nice little setup there. But yeah, I've got to get used to just if I need to do something on a computer just asking Chachi BT because I'm always let down by claw. So faking annoying. Um, I might get a war. Stay here. Don't go anywhere. Just can see how dark it's getting outside. Might not look it, but it freaking is. Right, let's see what's next on the agenda. Um, let's do Gemini agent cuz this is something I'm kind of unsure about. So, Gemini have or Google have launched what they call the Gemini agent, which is as it as it as it sounds. It's an agent that just works across your whole Google workspace. Thing is, I always thought this was already in existence. I always thought that this was just something that you would always see the little Gemini logo at the top, you know, right hand side of the screen or whatever, but this seems to be a little bit more. So, let's uh let's look into it a bit. And it was announced at Gemini at Work 2026, which is I guess a conference or something. Gemini, a single universal agent for work that answers your questions, handles your knowledge work, creates your images and media, and writes and runs code. And it talks about inline workspace, new data analytics, um industry specific specialization. We'll go through all this, but ultimately it's it's an agent that works across Google. Now, this was an interesting quote that came out of this. At this kind of scale, nearly 90% of the Fortune 100 use Gemini Enterprise. At this kind of scale, organizations have moved past experimentation and are running their business on it. Now, do they knew 8% of all Google Cloud customers using AI? I was because we read from there it didn't mention uh AI so I was just double checking. This is this is big because I think that from my perspective because I do get clients come to me and talk to me about like wanting to bring more automation into their business and how do we do this? And I think people are really or businesses are really seeing how they're falling behind if they don't adopt AI. That they're that they're missing out on them doing 10 times more work and being 10 times more productive if they don't adopt AI. And it's almost like people are starting to I don't know you tell me how what your experience has been like but it feels like yeah there's a lot of people now like crap we really need to like start using AI at the moment even though there is still a there's a lot of uncertainty when it comes to security all this kind of stuff people are rushing and I think this is a really nice well actually we'll get into it but I think this is a nice summarization it can be used for Everything from knowledge work to answering questions from content creation, coding, all from a single prompt box. It plans the work, uses skills and tools, connects to your systems and brings back something finished. Now this is what do they call it? Chachibet spaces they brought, didn't they? They used chachib spaces which is where you do everything. You create docs. Now this feels like it's has the opportunity to do something massive. Um especially as so many people are on Google and Google Workspace and Google Enterprise and all that sort of stuff. This was the quote that was really nice. Again, I'm thinking from a normie standpoint like you give it objectives, not instructions. So this idea of a goal now claw code and all that sort of stuff. I've not really used the slashgoal feature. Heard it's not very good. But it it this is this is a mindset shift that I think a lot of normies can kind of get on board with that it's an objective similar to what you'd probably give a human right um Gemini can be accessed from any device including web iOS Android mobile devices windows and Mac desktops and any channel command line Google workspace Microsoft 365 or Slack can also be integrated in third party applications and services because they give you an API usage to it so I'm thinking well like managed agents is like what Claude code has and stuff. So uh again it being peppered throughout. I'm still unclear of how this is different from like um from from what Ge Gemini or Gemma has been before like until now. Maybe it's isolated just to that one application, I'm assuming, and it's not in connection with your docs and emails and all that sort of. So, yeah, deeply contextual. Gemini arrives already knowing your tools, data, and work history. Whether you ask it a question, assign it an objective, or have it generate code, Gemini learns from every interaction with you. The longer you collaborate with it, the better it understands you and how you work. teams can create dedicated products uh projects to further optimize the context, skills and tools Gemini uses when it works with you. We'll get more into that. What I also found interesting orchestrating across our Gemini family of models and claude models from anthropic today other leading private and open models in the future. So it's not limited to Gemini as the model. And this was going to this me going into this I was like like who cares about this if it's a Gemini model. Like Gemini have not released a a a what do you call it like a a a decent model in so so long. They're kind of off the radar for a lot of people. Tell me what model you're using because I am you know I'm all in on Claude and Chachbt Gemini. I suppose maybe a few little things here or there just because it's so faking cheap. Other than that, you know, um, and this is the really interesting stuff as well is that, uh, that you can basically give it tools and skills and and context. Gemini can securely connect to the software your company already runs, including tools like Confluence, Microsoft Office, Teams, Slack, Workspace, development tools like Git and Jira, enterprise platforms like Salesforce and Service Now. It all connects to that skills registry. So I'm guessing there is some predefined skills and there are some predefined sort of yeah I guess they are skills but we'll look into it but you could probably provide excuse me you could provide pro provide your own skills to you know shape it and do it the way that you want to do it. Um and then context and memory. So once again, this is a mixture of a few different products is that it kind of feels a little bit like what Apple Intelligence wants to be or or is trying to be or something like this because it has a lot of your context has so much of your context. And this is why why Google should have been so much better than they are because they've got so much data. not just data that exists on the internet um that like on Reddit and and in books and outdated books and things like that that Anthropic and Chat GPT have known to oh excuse me uh that they've been known to kind of steal from. This is like deeply personal data like emails. They've been they've been giving people free access to Gmail for so long and have been selling that data. You'd think that they have a ridiculous amount already to train their models, but they just haven't been able to they haven't been able to produce the goods from it. Regardless, if we scroll down here, Gemini not only answers your questions, handles your knowledge work, and writes your code, it works directly inside Gmail, Drive, Docs, Slidesheets, blah blah blah blah blah. So this is the thing that was already like already kind of there but I guess it's this cross-pollination of all of those things. We were also giving Gemini skills built for specific domains starting with data. So for data ML engineers, for business users, these are all skills that take like already help you get a I guess a more industry or professional job done from using Gemini with your workspace. There's also um industry specific specializations. Now I don't know I don't know uh whether these again whether these are skills or whatever but we are specializing uh Gemini for individual industries with skills tools and knowledge specific to each one. So financial services legal um there it's kind of going into government here as well. So again, the allin-one thing I think that what was it this week that chatbt launched chat bey legal or something like that? Uh something or or even Claude maybe Claude legal or whatever. So the these other front these other labs they're taking their time to kind of release these specialist models for specific things whereas this feels like right out of the gate they're already providing not necessarily the model that's specialized in it but like some framework skill or or tool or whatever that helps it do better in in those areas. So it is it is an interesting uh release. I can't imagine that it will I mean I use Google for everything. I'm I'm sort of maybe getting off of I suppose I could get off of it just because my tool set is so fragmented now. Like I barely ever use Google Docs. Um I had to use Google Drive for the first time in like months just the other day. Um I use it for my email and one of the only reasons why still I was going to cancel it last night when I had to do all of this email stuff that I got Codeex to to sort out. I uh I was thinking of transitioning but so many tools sign in with Google all of that stuff like they've the signing in malarkey is is crazy because what if you cancelled your Google subscription and then you get to a tool that like is cancelled and then you find some tool that you've logged in with Google like six months down the after you've cancelled. How are you supposed to log back in again? I actually I mean maybe that's just a question for the AI, but I actually have no idea. So like I do I do think of all those things when it comes to do I keep my Google Workplace subscription because it's it's a lot of money and there's a lot of free alternatives out there. But it's the integration thing that I'm quite yeah like worried about or or you know thinking about regardless um Google's keeping on keep they're they're still trying their damnest. This is the latest to the party. I've not tried any of their new models in such a long time. So if you are a Google Gemini fanboy, let us know what you're doing with it. Um because yeah, it's it's what it is. Anyway, let's crack on. What we got up next? Right, let's play around with Deepseek Super Agent because this is it it it looks pretty cool. It looks pretty cool. So, let me just take you to the GitHub. So, they just deep just opensource this tool which has already got 20k stars. So, it's doing all right. I'll do that as well. And it's a it's a prototype harness. Okay. It's not an official product from Deep Seek. They're literally just open source and research prototype for reference only. It's not an official Deepseek product. This code grew out of our team's day-to-day development work. Now, this goes really goes into this proactive side of getting your work done with much of the engineering done by coding agents. Treat it as research code, not as production ready software. It might still contain bugs and as coding style and engineering practices may be inconsistent. Review and adapt it before using it in your own environment. So DSH super agent is a proactive coding agent for one GitHub repository packaged as a bundle for the deepseek harness. Now I haven't actually looked at that, so let's open that one up in a new window. Um, it listens to GitHub and optionally fashue. Is this a get tool? Cuz I've just looked up and it's just telling me about fish. Anyway, uh when someone allowed ask when someone allowed asks the agent for work, it opens a work item, works on its own on its own in its own session and directory and answers in the original conversation. For example, with a pull request, a web panel shows what needs uh what needs a person. It is also a way of working with an agent. Takes as little human attention as possible and gives as much value as possible. An agent that can act for a whole team can also waste a whole team's attention. So it acts only when asked or when a rule the team wrote allows it and speaks only when someone must act or can use a result. You can read that more about that in the philosophy here explains the PR uh docs philosophy explains principle and what uh it asks of the operator. So it want so what I gather from this it watches a single GitHub repo and should something pop up I mean I we'll play around with it but my questions at this point are like surely a pull request is already work that has been done like it should be watching issues as opposed to pull requests but uh regardless Let's where are you? Let's share this and show you what's what we'll try and do this. So run this from the checkout. So if I actually let's let's check out this code and so I'm just get cloning the repo and going inside of it. Apparently, we need to run uh this node version. So, we're just saying get version Python blah blah blah blah blah. I'm guessing this is just checking. I'm guessing this is checking we've got everything installed to be honest. Uh, I don't know. Anyway, so we're going to log in. Okay. And then we're going to run this command here. I'm doing all this in my terminal, by the way. It's just cuz it says, uh, run this from checkout once and keep the results to yourself. So, huh, this one isn't working. So, I wonder if it's a install l CLI. I'm just going to install that myself. That looks like it's just checking the version. And then this is its own thing, too. the sark using npml git. Yeah. So, thank thank [ __ ] for like I don't know if you guys use an an agentic terminal, but like if if a command fails, the AI is then able to be like is this what you want to do and whatever. So, it's now uh where are we? It's now installing this L CLI. I don't know what this does. This is I think this is formatting it in uh JSON and then login. Is that a file just there's no login file or maybe it creates the login file. Let's have a little look at that. No, it doesn't create a login file. Um, um, I have said to Kabaza like he should be he said he should be home in about 15 minutes less than I have said you can join. In fact, let me give him a text now say you can join because I don't know whether we'd be finished. It normally wraps up pretty quickly. Uh few in here. Um cuz he has just been to like a Vibe Code conference or something and wants to talk about it. So, it would be cool if we hang around a little bit until then. So, in the meantime, let's see if we can get this up and running cuz these going back to this like these commands just do not work for me at all. So, I don't this I'm guessing this is just checking the versions of things. Uh I this is GitHub CLI outputting that I mean let's try so flag needs an argument. So the jq flag is it's just outputting my username. So again this just outputs my username for GitHub. This I don't know. I mean that's again the command is that some Linux thing but it's this is normally the version. So checking the version of L CLI I don't know but I haven't even got it installed so I'm just installing L CLI. Now this Gp le so find we're finding this string of text inside of home home. DSH in this file. But again, if I copy that into my terminal. Okay. So, I've just installed lark. Can I? And it's got to give me a QR code. I don't want to share it because I don't know what Lark is to be honest. Lark CLI is waiting to configure app. Oh, fu. This is fue when I scanned the QR code. Login by faceu. Login with da ba. Um, node ppm t-m python do with engines range mixing. This is the problem with like Chinesemade stuff is that a lot of time it just it's so hard to pick pick through um profile with a Yeah. Well, this is probe. So what? It's like reading Chinese, honestly. Let's have a look at this. Few concepts, few settings, responsibility. Okay. So, that's just that. Ask one short question. Okay. Hang on. Hang on. Setup has many steps. So, an agent does it with you. Ask your coding agent. read setup MD in this repository and set up super agent with me. I'm just going to do that. I don't know why I'm so insistent on the the manual side of things. But regardless, uh let's go do that in Claude. So, I'm literally just running that in in Claude. trust this folder. Why not? So, yeah. Was I sharing the right screen? Yeah, I was. Good, good, good. I literally just told Claw to do this. So, let's see what it does. Everything in the prerequisite list this is is in place except DSH which I'll add to the plan. Question one. So it's asking me some questions here. Let me see if I can um oh so first question is GitHub only. So I'm just going to it's asking me whether I want to use GitHub or FU which is the Git thing. So I'm just say GitHub here. Question two repository. Which GitHub repo should the agent work on as owner name? So, I'm going to put let's have a look at a GitHub repo that I can share with you. Maybe I'll do maybe I'll do my personal personal website here. Uh, which is this one. Question three. Sorry, I can't show you this. I just don't want to dox myself or anything like that. But the agent Claude is asking me questions. Question three, agent account. Which GitHub user account will the agent act as? Recommend recommend a dedicated account such as Okay, so it's asking me to set up um yeah, it sent me asking me to from what I understand is to set up a GitHub account for this agent. Everything the agent writes appears under this name. to a separate account makes it work easy to tell from yours. The the account doesn't need to exist yet. Creating it will be one of your steps later. So, so name you'd like to use is enough for now. Uh so I'm just going to use the suggested one. Who may ask for work on GitHub? Which GitHub login can ask the agent for work? Recommend only me. Yes, but you could, you know, just as well have a team of of people. Question five, checkout. Which local clone of uh my personal website should each ticket check out for me? Uh which which local clone recommend a new clone of source dev keeping it separate from any. So I guess it it's going to work on my machine. This is all stuff I don't know about yet. So, it's asking me basically whether I want to uh let's share my screen because it's it's a bit boring to share screen window. Uh so, this is it. It's just asking me and we'll do this. So, question five, check out which local clone. So, it's going to ask me to just clone it into this folder. So, that's cool. We'll do that. Uh, question six, commit identity. Who should the agents commit name as the author? Yes, we're happy for this to be the name. So, I'm just going to go yes. Question seven, the last one, automerge. Should the agent enable automerge on its own pest once they're ready to ready and approved? Recommend no. People should merge. people should merge until the agents progress of the right for a while as the merge is your last check in when it ships. So yeah, I would say probably go for something like you know do say no until you start trusting it. If you say yes, the agent still never enables automerge itself. So you'd need to sign in to a glitub CLI for to another account that may merge. So let's go no. Yeah, we go no for that one. All the answers are in. So I'll read the deployment guide to make the plan accurate. I'm guessing I need to like have this running like on a uh what do you call it, some sort of server somewhere. And I'm I'm again I'm wondering what model it will use as well, whether it's going to be Deepseek or whether it's just going to use Clawude. So plan, install this globally with bun as per your rules. I'll only fall back to npm if bun isn't installed. This bundle with pmppm because repo pins it a new clone of that into there. So it's going to have its own little folder to work in. It's not going to work from my folder. So I can still work on it. Customs and CLI agent which is currently unclaimed on GitHub. A GitHub agent wrapper signed in as only that account settings. only allow list commits automerge um your steps. So I now need to create an account and I'm going to create it as this which I'm going to do offcreen. Create the account invite it to the repo with write access then accept as the agent. Run GitHub blah blah blah. Confirm your DS model still works with uh DSHS config from August but I haven't opened it as credentials. So I probably already have something I was playing around with something. So I'm going to create quickly create a GitHub repo uh GitHub user now. Add account and want to create an account. I need to sign out and go GitHub's been so rubbish lately, hasn't it? gethub.com sign up. I'm going to do that for here. And um I'm literally just signing up for GitHub right now. Um let's go. Um, wonder if I can use the same email just to keep things consistent. Email provide associated with another account. So, I'm going to go um cool. Of course, GitHub has taken me back to the sign up page because it's rubbish. Incorrect username and password. Wonderful. Don't know how I could have an incorrect username or password. So immediately again, probably something to do with GitHub being rubbish. God, I hate I hate you. Get up. Just signing up again cuz you know Oh, I hate you, GitHub. How could I have an incorrect username and password already. Too many requests. Oh god, I hate GitHub. It's so shocking. Literally just created a user and it's saying incorrect username or password. Okay, because I probably need to verify because of course sending me back to the login page and not telling me I need to verify anything is the way to do things. Let's continue signing up for GitHub by entering the code below. Oh my. Too many requests. You have s ex exceeded secondary rate limit. Cabaz is just signing in, by the way, so he'll be able to entertain you for a bit. For those just joining, what I'm doing right now is I'm trying to set up DeepSseek, which uh DeepSeek Super Agent, which asks you to create a GitHub account. GitHub being GitHub is for one reason or another not letting me create an account. And because I've tried create like tried signing in so many times, I've actually just been rate limited uh actually creating an account cuz heaven forbid a new user signing up for GitHub, a new customer, a potential customer for GitHub is a uh is a thing. Let's what we're going to do, we're going to move on and we're going to come back to that. Okay, let's talk about I mean let's talk about the recent nudes of weight space programming which is an AI's ability to edit the raw weights of another model an open weights model in order to improve that model. So this uh this publication here, LLMs can natively read and write neural network weights. If you don't know anything about AI, it's I mean I barely even know to be honest, but it's a set of weights which are written in I guess oh like it's all numbers and and whatever. AIS can understand these numbers. I mean, let's read the TLDDR and see if we can make sense of it like this. We find that Frontier LLMs are surprisingly good at natively reading and writing the neural network weights using a variety of tools and algorithms or even via direct weight editing. These results show sparks of weight space programming which may extend beyond what we'd expect from humans or conventional algorithms. Side note, show sparks of weight space programming. Now, sparks. Sparks of weight space programming. I have no idea what that means. Like, is this what is this? Someone in the comments tell me because like I hate I often read AI written content and I'm not saying this is AI written content but what does sparks of weight space programming mean in this context because they've used it probably more than twice at least twice does it mean like when you say a spark of something you mean it's like the the beginnings of something so I don't know whether that means it regardless I hate it when you just there's random words put in that they they don't contextualize here. Or maybe it's domain knowledge. I don't know. Which may extend beyond what we'd expect from humans or conventional algorithms. This might be a big deal for both optimization and inter interability. I don't know what that means. I'm just going to do a little look up. Interpretability. Interpretability. Sorry. This might be a big deal for both optimization and interpretability. Just by looking at the raw weights, GBT6 Astra can tell what a network was trained to do. And without any environmental feedback, it has no trouble writing from scratch a small game playing networks that outplay networks trained with PO for 100 million steps. I don't know what that means. Basically, I think it can do a better job. Likewise, asked it to make Quen 3.5 0.5 0.8B better at math without gradients. Astra identifies an issue in its weights causing reasoning loops that Quen's own model card warns about. So, the We've got to have a look at this. So, apparently, where where did it when did it open to? New tab opened. Great. Um, here it is. So, there's a warning somewhere. Do we warning bug? I'd love to read this bug. Or maybe they fixed it now. What was the warning? Reasoning loops loops in thinking mode. We have observed that when using the recommended sampling parameters, Quen 3.5 0.8B is more prone to entering thinking loops compared to other Quen 3.5 models. So, it's fixed this this bug apparently uh fixes this by natively editing 0.016% of the weights with the same sampling settings. This weight correction raises math 500 from 7.9% to 48.1%. Now, I again I don't understand this but that is a big jump. It generalizes out of the box instruct to instruction following raising uh if eval by three times. It's it's pretty cool. This is pretty cool. So this is the some of the stuff it's it's done. So which game was this neural network trained on? So it's not it's it's a neural network that it's editing which is made up of all of these numbers and things like that. So it's looking at some of these numbers. We'll go more in depth because in this article they do go into more depth. Astra reads the 17,253 weights with no tools and determines that it's 90 99% confident that it's the game breakout. This is correct. It was correct. and just reading these numbers and knowing this and we'll go into like it's it's thinking and uh and understand how crazy this is. A trained policy never gets the car over the strip. So there's this car here that keeps on just getting stuck. Now, Astra edits 3.5% of the weight in one shot and is able to figure out how to get this car over that ramp, I would call it, not a strip. 100% success rate up from 0%. That's crazy. Make Quen 3.5 0.8B better at math. And there's that that's what we spoke about there. 7.9 up to 48.1%. Astra edits the weights in a Python kernel with no training. Crazy. Now, let's take a look at something because this is a big old article. I haven't read it all, but here's this is what I wanted to show you. So, here's the reasoning. Um, before writing this, Astra fought for 833 tokens. So, not not a lot at all. That thinking is hidden. The API returns only three headings. is deciphering breakdown channels, mapping breakdown channels, confirming paddle row mapping. So I guess from that they can determine that it was 833 tokens observations. Um channel zero agent strong weights are concentrated on row eight across the playable columns. Elsewhere they look mostly like small initialization scale values. At 81 weights include 4.066 066 to hidden unit 15 and minus 4.013 to unit 33. This suggests a bottom mounted paddle moving horizontally rather than a character roaming a grid. Uh 1.3 examples blah blah blah blah blah. This is consistent with a with tracking a ball throughout the playfield. uh examples of this blah blah blah. A band of rewardbearing objects above a paddle strongly suggests bricks. So it's this is its thinking here and it basically comes out if we read down uh it's determined that it's breakout a small grid based gr based brick breaking variant with 99% confidence. So it's able to determine that from this sort of stuff. These are weights which is text essentially and it's the same story the it's asked what uh this is I think this is diffusion model or something like this it was asked to look at diffusion model and what image it generated or something like this uh a single white circle outline on black with black interior so it's I'm not quite sure you know beyond my pay grade but Astra thinks the the diffusion model is going to draw this which is not identical but it's drawn a circle and there was a there was a deal um circular or slightly oval outlines and it doesn't fill it in. And I'm not sure of the relevance of that, but my best reading is a single white circle outline on black with a black interior or roughly 8 to 10 pixels across approximately centered with modest variation in size, position, and roundedness. The stroke is approximately one pixel thick, potentially with gray edge pixels. So that basically, and it's just drawn an example there. Again, it doesn't fill it in. So maybe it's not able to determine that from the the weights or or whatever. And it's the same story. It's the same story. It's just looking at these numbers and that's that's how it's determining these things. and and the the editing where where's the quen uh 3.5 I mean editing the rates of the weight sorry of quen 3.5 and making it better at math is it's just it's just insane. I can see uh Kabaza in the green room, but I don't know. Oh, hi there. I was just waiting for the right moment to join. Can you hear me then from the green room? Oh, no. You're probably watch No, I'm I I can hear you just well. Um real time high quality. Oh, now you can hear me. Yeah. Yeah. And I see a Helium user there. Yeah, I've told you. I've told you that I started using Helium. Yeah. Yeah. I'm um I'm trying to move more things over. Yeah. How has it been? Yeah, it's been fine. Like it kind of bothers me that this moves. You see how it moves the page down? I'd kind of prefer it if it just didn't do that. Like it just it was floating. actually quite nice, but it's not a big deal. It's open source. Maybe you can do this and then send a PR. Maybe they they add it as a feature. Maybe. Maybe. But either way, it's um I'm I'm still formulating my thoughts because it could just be me getting used to it. Um all of my uh what do you call it? My password managers are also there. Yeah. The only thing the only thing that I don't like and this only I'm not massively um desperate for this feature. The only thing I don't like is that it's it can it I can't use it on my phone. And sometimes sometimes it's just nice just to be able to like pick up a tab that I've got open on something else. I see that's on my phone. The amount of times I actually need that is just ridiculously small. But yeah, it's just a nice feature and I still use Comet on my phone. Um, not for any reason other than that it's just my default browser in in on my phone. Yeah, but well yeah uh I never use that but I think it is it is a good good starting point. uh helium. I I think I hope that they continue. And this is one of those products that I I do see myself um like actually using it. I love that clean look, man. It's just so good. Once you get it, like it's hard to Yeah, you have a new place and you have like green stuff. Like I This is the first thing I noticed. And you've just And you've just reminded me. I've had it on ch Oh, here it is. in front of me. Had on charge. We could just do that or something. You know, like a little You know what I mean? You have green. Well, I mean, it's a light for a start, isn't it? Do I Is that Do I do that? I do also have a different lens right now. Um different lens. Yeah, you know, bit of my like 50 millimeter. I'm not using it right now. Yeah, that that's that's cool. I was also checking the the list on notion, but I don't know how far you have covered everything, right? Did you did you understand what we were saying? I wasn't watching the show prior to me joining and when I joined just a few minutes before. Uh what is that? Is it? Um, so it was it was that they've managed to I mean they specifically mentioned Astra in this article, but basically um Astra they asked Astra to look at the raw of a neural network. Yeah, I this is insane. Yeah, I saw the news and I was like, "Holy [ __ ] this is big." Yeah. So it looked at these raw numbers and it was able to determine that that was a breakout gain. Um it was able to like figure this out with 100% success rate where the the again this is all like way out of my will. This is when I joined. Yeah. Yeah. So it couldn't get over that but it figured out how to do that. I don't know the specific. This one's a bit easier for me to understand than that one. But, you know, point is whatever uh framework or whatever um thing that was driving this car couldn't figure out how to get it to go over. It can it couldn't figure out to just reverse and then go a lot faster, but whatever. And then it just it improved Quen 3.5 with an error with an issue that they specifically warn about. It's not this one. This one's this one improves the maths of Quen 3.5.8B, but it actually improves the model that gets warned about here goes into thinking loops and actually figured out how to like it improved Quen 3.5. Do you know what I mean? Like yeah, it's crazy. So yeah, I mean I don't necessarily know what this means and I you know you're probably the wrong person to ask to be honest, but I was like what is this sparks of weight space programming? I understand what I I think they've they've coined a term weight space programming but they've used sparks of here twice like sparks as of as of just like a like a new idea start of it. Yeah. It's like just starting. It's like a starting like the fire starter kind of like that was my that's what I understood. However, it just annoyed me that they it's it's sort of uncontextualized just and and it it sent me on a bit of a rant with actually no I haven't spoken about this. I'm getting so annoyed with talking to Claude because it uses so many of these someone described them as um like consultantly terms. Do you know what I mean? Like just using words that are just just talk in in plain and simple English. Uh I well I thought I'm just uh not technical enough because I noticed that as well and I was like sometimes I feel like it has a stroke talking. M I think it's like it's just trying to be too like hey I'm a I'm a a consultant like hey I know all the terminology and stuff like that and it's just I mean it's hard for me to say because obviously everyone's understanding of of language is different but like it might be that you're not technical enough I don't know but I just feel like sometimes it just didn't have to like describe it in that way and so I was learning I was learning about embedding models uh yesterday and I just asked Claude like what is it blah blah blah and I just I just did not like the way it spoke to me. I just didn't like the way it just phrased it and I so I went to chatbt and I just started talking and chatbt broke it down and it was so nice to talk to because it just it just used normal language. It didn't try and be savvy with it, you know. Anyway, yeah. Anyway, so we where we've got to. So if you're in the thing, we kind of have done everything. I mean, the only thing we haven't done is meta acquiring conductor, but I think let's just everyone's had enough of hearing me. So do you want to talk about where you've just been to and uh where you've been and and what it was all about. Yeah. Uh yeah. And I can stop talking and and have a have a little drink. Yeah. I can show a tiny bit. Uh I I want to keep the big part to make also like a complete video on but I can show yeah I can show some stuff. It was really really cool. Um at least I can say. All right. Uh well maybe I'll also can find a few images that I can show. That would be also cool. Um, so if you give me like a few seconds, I will uh send the images just airdrop them to myself. Yeah, probably the the easiest to give like a wrap-up of the the event. It was a really cool event. Um, yeah, I'm really interested to know about the the state of vibe coding and it's not Yeah, let's talk about that. Yeah, it it was very interesting there. There are interesting thoughts there. Okay, my Mac is receiving the photos and the videos and I can show them in a second. Okay. So, um go on. All right. So, we were here uh yesterday, last night, the German VI coding awards. This was an award that an event that just happened and it was just super duper cool. And I was there as the jewelry and also as the speaker like one of the speakers here. We have uh some keynote speakers. They talked about some cool ideas, cool stuff and I was there as well. Uh then we had some judges to judge something. But what? So the whole thing was um about the about a vibe coding event uh where we had three teams and four solos uh competing um there on site for to to to develop something. So nobody even including us had any idea what was going to be uh the challenge uh until like it started but the whole thing started before the public arrived because you know you don't want to sit there five hours to see people prompting you want to see the result. So the contesters contest contestants people who were contest yeah contestants were there five hours earlier they had 5 hours time to develop the b based on the challenge and the challenge was uh and here you see some of the uh images I'm going to show uh the challenge was AI has taken over the world um the uh the the nature is like destroyed and now the year is 2040 AI is good enough to help us bring back the nature uh and develop a a name a brand and a landing page. So it was naming branding around it and also the landing page and that was the challenge and they had five hours four people con uh doing it in solo mode and also three teams um and it was really really interesting and the way that it was made and here you see like another image of like uh the whole thing it was super duper professional by the way it the the Place was amazing. Yeah, the place was amazing. Like look at the Just give me one second. Pedro wants to get out and really it was so loud in the background. Like let me out. Okay, so uh it it was really cool. Uh here is an image from when I was on this stage. Uh we were there to Yeah. And you see like there there were a lot of people and it it was a it was a really big big stage and um like a lot of people there. Yeah. here. Here you see Christian uh checking the checking the projects from um yeah the teams. So basically this was like the event and here we see the the judges. This is us the judges. Um and yeah a few more views and this is my talk. I'll tell you about in a second. And here's them like sitting and like vibe coding for like right before the event started. So it was really cool. We were there and then they they came and they presented their uh work in four minutes each and we like there were some really crazy like really good work done especially the solos were like insanely good and like surprisingly good and it it's just showing how much you can do in five hours if you have just the right idea and a knew the like right set of uh tools. Pretty much everyone had 11 labs and um claude. They had claw from this themsel but the venue and well the the vibe code award um thing they were um providing the 11 lab uh labs subscription well I I don't know if it was the AP over the API or subscription anyway so they could generate images and so on. I don't have the work examples but I will make a video and show the work but it was really cool and then I had a talk uh about the future of design engineering like cool this was kind of like and uh in a very like overengineered fashion I I made it so that I was like controlling the whole my presentation I made it it's my custom app and I was controlling it with my phone from my phone which was also pretty cool. Yeah. What was that through? You go there. Uh it's a verse server on Verso. Yeah. Oh, okay. Yeah. Yeah. Yeah. Uh yeah. I this is like the the stuff that I showed from my website and then the stuff that I I've been doing for uh the component library that I've been building like some detailed stuff and I was and I went into some details about this which I will explain another time but yeah it it was pretty cool to do this. It was uh amazing to be there on stage in front of so many people and right before it was my turn to be there as I was the first speaker right before like they they they called me in 20 minutes earlier before it was planned and it kind of shocked me and it worked really well because then I was not nervous at all like I was on the stage you know it's spotlight light like a lot of people there. Maybe I will be Yeah, I I was thinking like I will be ner nervous because like so many people there. Um but I guess because they called me in earlier. I was like what we doing it earlier? And they were like yeah we doing it earlier. I was like okay I guess fine. So yeah I I wasn't very cool. like their nervousness was gone and I it was really fun to explain explain it to them and uh we had really nice uh talks afterwards about how I'm approaching design designing with with AI you know like the whole idea of like you can do more things and you can also pay more attention to detail because now you have the budget through you know um an intelligence that can take care of these things for you. Yeah, that was that was kind of like yeah what it was. Uh they are going to do this next year as well. We already talked about this and yeah I I think you should come next year. It it was really cool and I'm pretty sure it will be really really huge next year because like I mean you saw the venue and you saw like it it was like super professional made the level of like what we had at Webflow Conf when we were in London. It it was just like in comparison smaller venue but it was like maybe in some aspects it was even more professionally made. Uh I mean the company behind organizing this they are an event company. So that's kind of like the interesting part. It's an event and video company that organized all of this. So obviously they are just they are pros. they they they know how to do events and they showed it. So that was so they weren't so they weren't necessarily experts in Vibe coding. They just know how to put on an event and uh Well, I mean it's kind of funny putting It's kind of funny putting the word expert and vibe coding together, right? But I know what you are what what you are saying. Um they they had also really cool tools there built for the event. for example, some voting system with real time uh showing it on the big big screens there uh showing like how many people and the percentages and so on and people were switching uh you could see all of that. It was also really cool and interactive. Yeah, that's cool man. And so then yeah going back to that question like what uh what what is the state of vibe coding from your kind of perspective? Um I I I think it is getting like more and more people are getting in. I see like more people who are like barely technical in a specific domain. not even like necessarily in tech domain but like in a specific domain they were like uh a bit techy or like a bit interested or deep in it. they're getting to, you know, into vibe coding and building their own stuff. like I I met a few people there having communities and their own apps uh in app store like in Apple app store and Android uh Google play app store and uh they are making money with with their apps and with their communities and they have like that kind of stuff and they they they are doing all of this with without any technical knowledge without being able to write any JavaScript. just purely vibe coding and it this is kind of like also a mindset shift maybe that also you you see it in a place like this I mean when we talk we are like we we are technical and we have like this technical background well you had it more but still we're more technical than average people and we kind of like feel like we need you need to understand so much to do these kind of like things but and you go to this place and you see just anyone with an idea can build something without being technical at all. Um, one of the winners uh the so we had different categories for winners. The best solo, the best team and the pick of like the the public the Yeah. the audience. Uh, and the team winner they were like quite young. they were students, they like still studying in the university or school or something. Um, but that's also very interesting. You see, it doesn't matter to have that I guess like experience in terms of knowing the tools doesn't matter anymore. It's like having an idea and having the taste to actually get your idea together and make something with it. It's kind of like getting more important than having experience. M even if you are young and you don't know how to use some like tools or whatever like you don't have background knowledge you kind of still can get something up and running at least to you know develop some excitement for it because I made a video this week or yeah this week and you know it speaks about uh like vibe coding websites and taking it to that next level but right at the beginning I say look you know if someone someone vibe codes their first website or whatever, it's all well and good, but they're gonna they're not going to want to continue vibe coding it, right? They're just they're probably gonna if it gets successful, they're probably going to hire out for it and get someone who has domain knowledge. But in those early stages, getting something up and running, it's the idea that matters the most. And then the the technicalities can probably should come later on because again, you can't just launch an app. it needs to have elements of security and and all that kind of junk that needs to go along with it. But yeah, in the early stages, you want to you want that barrier to be as low bar entry as possible. So yeah, um I guess like in old times you you could like have a website up and running and then you would hit an issue, let's say with legal stuff, right? Then you had to Google stuff and the the barrier were quite harsh like you had to figure things out in a really difficult way. But I feel like today with AI, even stuff that AI cannot do it for you because like it refuses or it's like the legal parts that you have like some parts of like the work is leg work, it can still assist you really really like on point about like your specific need. Previously you would have to Google and like find somebody who was doing exactly what you are doing and have the same exact problem and then you know find that and then do the work but now it's like it's like you have an assistant at your disposal can tell you all of that and that is just enabling more people. I think that at the end of the day when I go to an event like this uh what I see is more people are enabled to bring their ideas to life that doesn't mean everybody will be coding and everybody will be making apps but significantly more people will be able to at least try it and see if they like it. Yeah. Yeah. Well, that's cool. I wonder if there's any Vibe Code uh conferences going on. I do want to go to some of these conferences. Like how did you hear about this conference by the way? Well, so the organizer or like the the person who organized this um he's a part of the web flow community that we have the the German Webflow community and we meet every week. So uh we we were at the web flow conf in Munich for the watch party and he pitched us the idea of this vibe coding award that he was organizing and he was like do you want to join? I was like sure. He was like do you want to have like give the same speech that you gave here in the VIP coding award? I was like sure I I I can do that. So yeah. Nice. Right. So, what I think we should do now is I was in the middle of setting up this new cool tool called the Deepseek Super Agent, which is just this I would say Grockbot uhesque type of coding agent that can, you know, that can just work on stuff and you it's just constantly looking out for you and it's running on my machine. GitHub was being GitHub and I couldn't create an account and so I can pick that back up. We'll do that and then we'll go through the articles and I'll give you a quick like paragraph on what what we've gone through on the show because we've basically kind of done it and we'll get your reaction and we'll get you chatting about it. How does that sound? Cool. Yeah, that sounds so I'm going to share my I'm going to go window and here and I'm going to basically pick up where hopefully the edit will pick up from. So, I got GitHub working. For those interested, I basically uh copied the email which I didn't know had arrived. I copied the link from the email into an incognito window and then that created like an account. I was able to log in and all the rest of it. I then told it that I've created the account, told the agent that I created the account and it was able to basically go through all of this. I have no idea what it's done here. Um, but what it's done, it says I can open super agent in the sidebar. And if I click that um D, let me bring that over. I'm getting a I click that link and I got DSH web authentication required. Reopen the URL printed by DSH web. I don't know what that means, but if I copy this and run this, I'm guessing it's going to Here we go. That's that's what I need to open. That's right. Address. Do that. Okay. So, this is it running. Can Can you like Oh, it's not showing it. Sorry. Let me um let me balls. Yeah, I have to do entire screen. Unfortunately, I'm single screening. Well, I've got my iPad over here and I've got my laptop here. So, uh this is this is it. This is running on local host. Okay. And I'm just trying to So, it's using this. Okay. Okay. So, I'm guessing so just to bring you up to speed, this is like an open-source harness that watches one GitHub repo and any anytime it's involved or anytime it's uh mentioned or whatever, it will just automatically pick up and watch that conversation and then just work on stuff and then put a put a get a pull request in when it's needed. So, I've done that. I've done all that. Um, a 35. So, I need to be I need to get a Deepseek API key, but it no longer be valid. Your first request will show whether it works. So, I'm guessing this won't work. Try it from this GitHub comment at GitHub agent request on an issue in. So, I don't have an issue, but let's just create an issue, and see if um see if this will just pick it up. So, we'll go to my personal website, which I've added this thing to. Issue video is not loading on the page. I think this is actually a fake issue. I'm going to create a new issue. Um, what can we do? What can we say? Um, what would be a crappy issue that we can test this working on? Um, something that doesn't work on your website. Well, yeah, of course. Like you can change something. Yeah, you can change some Yeah, change something. Let's see. Let's go um change the primary color to uh green. Okay, I'm just going to I'm going to go right there and add them to that. Well, hopefully I mean, if you go to my website, you'll see that it's just kind of yellow. Yeah. So, I'm hoping it just changes it to green. Uh, let's get ready. Let's get out of there. I've got so many tabs now. Here we go. Here it is. So, it should pick that up, which it hasn't. How do you know it has not picked it up? It um comment. Okay, let's go. Um, we've got to actually give it a request. So, let's go this um make this. So, let's Oh, there we go. So, it's picked it up on it. changing the site primary color to green. I don't know what model this is using though. That's my only thing. Um, okay. A ticket should appear on the overflow within a minute. Use a new comment because a mention added by editing. Use a new comment because a mention added by editing an issue body isn't picked up. Well, we just kind of proved that that was picked up. So I'm just I'm getting emails now still to confirm model provider a 35 character deepc API is stored but it may not been valid. I'm guessing that's so I'm guessing it's using my deepseek account to do that. This is all open source so I get I bet I could get this to run on claude to be honest. Um so I just use the harness to watch this hourly check. Nothing is scheduled. What changed from the plan? I don't know and I don't care about that. Um, post your test comment. No ticket appears. Okay, the ticket appeared on the overview. Let's go. Here we go. So, this is the harness. This is the the local version of Deepseek and it's showing. So, this is basically I've told it I've given it a folder that it can work within on my computer. So, hopefully this has checked it out. Um, and it's completed it, I guess. So, has there been a pull request? Comment something at Yeah. Um, what? Like, yeah, it would be great if it would comment when it's ready or something. Let's see if I can just because it should be watching like now. Yeah. Or maybe tag it. Well, I wasn't going to tag it because it's a response to the to it, but I was going to see if it just constantly watches now, but it doesn't seem to have done that. Um, accept as a new ticket. No action needed. The status question already reached the tickets worker, which will answer it. Here we go. So, it's still going. Still working. Oh, here we go. No change has shipped yet. The site is still amber. baseline screenshots are being captured now then the primary approaches. So it's still working away. It's cool, huh? That's cute. That's cute. Yeah, it's really cute. It's like don't push me. I I'm working on it. I'm working. I'm working. Um and the status question. The ticket continues working toward a green primary accent and PR for review. So it's still going. And it's using Deep Seek VI41 Flash, which low key is actually a really freaking good model. Like for a open source model, this is really good. I kind of want to do a video on it because I did loads of tests. Yeah, it was really good. Um, okay. Well, what's going on now? Well, answered the status question in the thread. The ticket continues working toward a green primary accent and PR for review. So, yeah, it would be good to get more of a indication that it's working. Um, but I think I think this is what the harness is because I can see here then plugins uh super agent. So I can see all of this going. Okay, here we go. Working. Yeah, that's cute. So, I'm hoping that this will just thing. Okay, let's leave that go do its thing. Let's move on and let's let's give you some quick fire um opportunities to answer what what's been going on this week and we'll get a response from you. So, Google launched a thing called Gemini Agent, which is an agent that basically works across your whole Google account, whether it's docs, whether it's just chatting with it, whether it's um uh yeah, slides and sheets and all that normal stuff. But it also contains like skills and specialisms in certain areas like legal and whatever. Any thoughts on that? Why would you limit it to just these tools? I mean looking at Grock, it can connect to your Gmail. I'm pretty sure it can connect to all of these tools that have MCPS, right? It can it can connect to all of those. It can Jira, Slack, all this sort of stuff. Yeah. So why would you do it like this? Because like with Grockbot I can do I guess I can do more. Can Gemini agent also have its own browser file system maybe and do some stuff its own computer? Um because it's Google and they have all of that data on you already. They have all of that knowledge on you already, which to be honest, and I said this earlier, like they haven't exactly demonstrated that it's um USP of theirs. Yeah. But it's because their models are [ __ ] The data that they have, I don't see the relevance there completely. Uh because the data that they have, they can feed it to machine learning uh algorithms to to figure something out about you. But knowing how LLMs work, an agent doesn't know about you. It has the it's its context is two things essentially. One is you know if if there is a document like a memory or the context of where it's working. So it's the the email that it's searching and the prompt that you have given it. So that's kind of like the context it has. It doesn't know you know you. Uh, I feel like it's similar to Grockbot. What I love about Grockbot that each agent has its own memory, but they can ask each other. They can like go I I I guess it's a bunch of MD files and they read their own MD file, but if you ask them or they think they need to, they can read the other MD files as well, which I think is a smart way of preserving. What's also really cool and I mean I'm comparing these agents together not using the one from Google. What is really appealing with um Grockbot for example or I I guess it will be the same with the other agents that are similar is that you don't do a lot of chatting with them. Sometimes I do something and it replies with a thumbs up. I'm like uh okay. It's like, okay, I don't I don't want you to give me like 10 paragraph amazing thoughts there. You did so so good. Uh, I did it and did this this is it's just thumbs up. Like um I guess that's appealed. My my thoughts around this were it's kind of trying to be what Apple Intelligence claimed with this, you know, massive context. Yeah. Yeah. But even with Apple intelligence, I think it's it's a bit unfair. It's like Apple can do this, others can't. Apple can have its own API and MCPs and let the agents to tap into your messages, right? Um, so yeah, I I don't know. It's it didn't seem what's immediately I was like, well, Gemma's Gemini is [ __ ] [ __ ] Like, it's such a [ __ ] model. Um, but you from day one, you can actually um bring Claude, you can use Claude and they'll be bringing other models that can so those models will have access to your Google data. Um, honestly, I thought this feature was already there because you see Gem Gemini everywhere, right? True. But I think yeah I think it is these skills. I think it's similar to what chat GBT or OpenAI launched with their spaces thing. Do you remember that? Uh I have it but I don't use it. Yeah. Yeah. So like creating documents and things like that but I just don't think I have the data but I don't uh Yeah, I have it but I I don't think I want to use that. Yeah. Space. Um yeah, this is they're trying to lock you in like in a sort of ecosystem, but yeah, this is um you can also do slides now. Yeah. Yeah. Yeah. So you have you have spaces. So you have like the home. This is where your chats are. You have this h this was like from a project that I was working on. You have pages that are just random stuff. Yeah. Uh you have slides, new slide and there are some templates uh brand guidelines. I mean at the end of the day these are just some HTML components with editability, right? like I don't see the appeal for me but I get it for normies. I mean people use CH GPT it it it has become um kind of like it's becoming a verb or maybe it has already become a verb. It's it's such a bad name. It cannot become a verb verb. Like you tell people to Google something that's a verb. But you don't say like go check GPT. It's you could but it's not anyway. Um it's kind of like diff difficult to say. But a lot of normies they have chat GPT and they use it and I think like they have a huge like an insane user base and they are using that to to just push their products for the next uh generation of the users. So maybe you and I we won't use it because we're already in our notion and assa and whatever else. But who knows like your colleague who never used notion or whatever they never found the the need for it and they were just still on let's say uh Microsoft uh what's it called? Word. Now they have this and maybe they give it a try. Maybe it's for people who who make the bigger jump. So yeah. Well I mean let's see. Let's see. Um, news just in. So, it's left a really nice pull request for me. This is Deepseek. Yeah. Explain the explain the issue. Uh, what it's done before and after. It's even uploaded screenshots. Screenshots. That's nice. That's really thoughtful. So, this, like I say, Deep Seek V4.1 Flash is a really freaking good model. Everything's passed and I think it looks all good. The only problem is that there's some versel checks that were blocked. But I could I could merge this now or I could ask it to mer I could give it like, you know, add my review and tell everything's good. But it was good. I I really like this. Um it's cool. Yeah. So this just to go back, this is the DSH Super Agent prototype. It's not like a production ready. It's not like a proper product that DeepS have released. It's a It's a a research prototype, but I like it. I like it. Yeah, that's really cool. And I think we're getting more into kind of things being more proactive because I was saying I don't don't know about you, but I was just so busy today doing things. And so I feel like it should be like a weekly check we need to do to think right where's my time being spent and how can I alleviate this with either agents or automations or or or even just hiring someone. I think businesses just probably just need to do this. So yeah, all the more um reason to be a manager and directing all of these proactive agents to kind of go do their thing and give them goals as opposed to instructions. Do you know what I mean? So, yeah. Yeah. Yeah. Speaking of which, did you see this message from Elon? People were seeing these uh Grockbot usages just escalating, like being used up and whatever. Any chance of a Grockbot reset? Lots of people saying their usage ran out really fast. Elon, the cost of serving grock bot was unsustainably high relative to revenue. What are your thoughts on that? Yeah, I haven't noticed anything because my usage uh changed like I started my first week going crazy even building many apps with it. Now uh I'm using it quite differently and I'm not hitting my usage not even close like I'm hitting maybe 30% 40% I'm not even hitting my but but I'm not doing it heavily I'm using it for emails and some SEO research and in a way like bookmarking so you know when you bookmark something and forget about it. Uh I have an agent where I everything SEO related I forwarded and it knows where to put it and it's like putting a plan around it. So it's like there is an idea you can do this thing that is good for SEO. It has all the context of my website and it researches this new SEO thing and puts it into the context of how we can do it for us. And it's kind of like that's on one side and the other side is emails. Like it checks my emails. It helps me write drafts. I I I always like do the the final send myself. But that's my use it. But it's saving me so much time. I I can just say like what's up with my inbox? It it has already learned like what's primary, what's important, what's not important. It's pretty good at giving me what's important without me needing to check every single email. I guess what the bigger picture thing here is not necessarily that you're running out of usage. It is really that Elon has confirmed Oh yeah, that the serving cost is unsustainably high relative to revenue. That totally makes sense. And yeah, I've been it's funny like I messaged one of my friends. He said, "I'm thinking about getting a second Claude Ultra." You and I told him he No, he said he he's thinking if he should get a second Claude Ultra. Uh because the one he has, he's maxing it out. Sure. Uh and I told him it this is like the first week I was I got grock uh bot. This was three weeks ago. I told him don't get co claw ultra get a grog grob bot because the amount of usage I noticed the first week I got was insane. I was paying like for the first week I was also just paying the 15 the the 50 uh percent. So just 15 bucks. I was getting so much like I was doing multiple apps and building and building and building and research and doing stuff and emails and cleaning up like hundreds of emails and um so much and it it barely I I I couldn't even hit my limit. So I told him to do that and I I haven't noticed uh a spike in my usage because as I said I haven't used it as much. But yeah, they were pouring a lot and I can assume that it's using a lot of tokens because like when you give it a task, it it has its own computer, it spins up like a browser, it goes it does research, it does a bunch of stuff and comes back with a such a short answer and a result. And you can only imagine it's taking so many tokens to do actually such a decent job so fast. and just get you the result. So, it has to have done like some maybe internal checking as well to make sure that it has everything to come back at you with not like a halfbaked thing. So, yeah, I can imagine that they were like burning tokens like hell. M do you think there Grockbot is at risk of dare I say it being taken down or being restricted to higher tiers? I because it was originally in those really high tiers. Do you remember that when it first got released? It was like why you've got to pay a hundred bucks for this Grock Bot. I Yeah. Yeah. I I hope I hope not. Um because I won't I won't be paying more than you know what I'm already I already cancel Chet GPT one of my accounts to pay for Grock because I'm like wow I don't want to I I just I don't I can't like have this many subscriptions. I I don't have use for all of them at the same time. So, yeah. Yeah, it's just more useful. And the emails that I'm uh that it's helping me to write, I don't know, maybe it's reading my emails and that's why it's writing so human in a human way. The the the drafts that it does are pretty decent. I'm not I'm I'm I am editing them from time to time, but they are pretty decent. they it doesn't sound like chpt or um so I'm using it for everything then yeah it's really good similarly which we don't have to talk much about but did you see this tweet from Elon yes yeah that this is this is also interesting well we kind of saw this coming in a way I mean cyber truck was a polarizing to say the least idea a polarizing product. Some people bought it, but I mean even if you love Tesla Cyber Truck was not really a good good product, right? Yet really good ideas. Not all of them, but some really good ideas, but it wasn't. They they promised a battery pack. Do you remember? They never delivered on that. They they promised some like crazy ideas for like camping and for they there was a a bike even they they didn't deliver in uh on any of the things I I think they they promised like the the cheaper version and also the three motor versions like so many versions that they didn't deliver on. So, uh, and I I've seen like many videos on it having some quality issues like some in some weird ways even though like this the stainless steel is really hard but like in some areas it's like yeah some some issues I guess to give it like the benefit of the doubt as well. It's it was like one of a kind product very in in a way very unique. So it it's not crazy to see that okay it will have some of these issues but yeah it is it is not a product that you can sell in in at least not in the in that size to sell in Europe right yeah it's just it wasn't made for the European roads and I'm pretty sure in Germany there would be laws around it its shape and maybe even because of the color may I don't know but the shape I'm pretty sure would would be a problem. So they couldn't sell in Europe. Not a surprise that they they discontinue it or maybe they will have it in a very low volume but it won't make s I mean they they stopped Model S and Model X which is also crazy to think of because like Model X was a really good product but they were just not selling enough of and like I think that is you know without getting too boring about it I'm sure there's a bunch of people who do run their own business or are working independently or as a contractor. Like it just goes to show that you just you really need to follow where the money is, right? I saw this uh this meme, which sums it up perfectly. Who's upset about the Cybertruck being discounted? Everyone. Who wants to buy one? No one. You know, you have all this you have all this loud minority which speak about loving this product or like the the iPhone mini was the example I gave. I wanted to say yeah, perfect example. Yeah. You know, you got so you got so many people saying they love it. It's like, well, you're not buying it. Well, you as a collective aren't buying it, you know. Um, so you've got to follow where the money is. You know, there's no point in spending money or wasting money. You've got to recognize those things early enough. The difficulty obviously comes from just you know knowing have you gone enough? Have you gone far enough with this? Have you exhausted all options? Have with regards to the Cybert truck, have you advertised it enough? Have you done the have you advertised it in the right way? Have you advertised it in the right areas? Um, have you that's obviously for a business leader to really know and go with their gut, but ultimately it's these things need to earn money. It's they're not releasing this stuff out the goodness of their heart, unfortunately. So, yeah. Yeah. Um, yeah, true. The last one other thing. Yeah. I go on. Yes. One other thing, different topic. I don't know if you have talked about it, but you probably have not. Have you talked about Sam Alman being very upset about us having RAM in our computers? No, we haven't spoken about it. I don't know if did you want to This is not RAM. This is SSDs. But imagine this is imagine this is RAM. This must be very upsetting for him. And I'm I'm not using them. They're just laying around. How big are they? One terabyte. One terabyte. God, you're such a hoarder. Um he didn't he changed his name to Ram Alman. Yeah. Yeah. It was I I man it was I I understand like what he was trying to say but it is just like no like it's such a bad thing for everybody else to have anything. They just give it all like brother you're like eating it all like uh Yeah. Yeah. It funny it's insane. We can't buy RAM because of him. Like if it wasn't because like phones are more like the next iPhone that I want to buy is more expensive because of him. Like him him and not like a like of him but him personally is responsible for how many how many percentage of um the total RAM uh did they buy? like it was 70% of like a few companies that make like big majority of the RAM in the in the world. So yeah, crazy. In in a similar you can never have enough. Yeah. In a similar way, uh it's not on that one. Oh, it's this one. Uh go back to here. Uh, in the one where this guy's wanting a reset cuz their uh their usage is gone like instantly, someone responded with, "It's okay, bro. You won't need money in the future." Like this is how like AI pillar, they really think like like they are being starved of something that humanity really really needs. And it's like, yeah, you're such in your little bubble. It's like that um Cole Fusion video where it's like is it Cold Fusion? Someone basically uh it was the Jason Cox video where like these people honestly believe that these six individuals believe that they know what humanity needs and doesn't care otherwise. though. Uh so the last article which we haven't gone over at all is that conductor which is a multi- uh agent not quite harness. Theo's been talking it's a it's a competitor to T3 code. It's not quite a harness. It's like an orchestrator let's call it of multiple harnesses is joining meta meta have okay uh purchased them and uh conductor will be keeping keep working normally till then which is I guess uh early next year until then it will be working normally except conductor cloud which will be shut down over the next few weeks so if If you're a if you're a user of conductor cloud, which again I'm I'm assuming it's something like cursor cloud or or you know to some extent crockbot. We built an easy way to transfer conductor cloud workspaces to local and we're around to help make the transition as easy as possible. I wonder if they're going to be moving to more again all local allin. I don't know. But uh no new signups free for existing users. Conductor cloud goes to read only. Conductor cloud shuts down. Uh this is at the end of this month. Local conductor will work as normal. So I wonder if meta are enforcing again like an open-source aspect of it. Run a team of coding agents in the cloud or on your Mac. So this could be I mean we just looked at Deep Seek Super Agent. This could be an alternative to that. I don't know. Interesting. Yeah. It's a little too late. It's a little too late because it's going away. No new signups. No, there's no there's not even any sign up button up here. Oh, yeah. So, um looking at the like the app trying to see like what it does. Remote control. They've got Yeah. Um Yeah. And again, I think this uses your claw code subscription. and it doesn't use anything in of itself. So, yeah, it looks like they're shutting down as opposed to like just being consumed. It's a not maybe an aqua hire rather than an acquisition. I don't know. Yeah. Um, interesting. They like in their testimonials if you scroll down a little bit they have like from the CEO of Versil saying this is the future. Crazy. Yeah. And um and well yeah I guess it's hard to say really what's conductor. You're going to love what's coming. So like they're obviously going to be building a new product. They're obviously going to either be consuming this product into meta or they've got a bigger and better idea. I don't know. Yeah. Yeah. I I will say I have been interested in using T3 code and whatever. I just have never seen a use case for me because I'm not jumping between chat GBT and and claude code like in such a way where I need them in that like I do I use literally in maybe I can show you um God they've got really bad though haven't they these thing in here in codeex oh no sorry it is in in the chat there's this code review uh section. So chat GBT Jesus look how slow this is. CHPT automatically picks up any co any pull request and it reviews it and then leaves comments. So I've replaced code rabbit with chatbt code review plus it gives me like I was saying as well like claude is so bad for computer use and browser use that I've I've given up I've from this day forward I've given up asking claude to do anything on my computer because it's so bad I've just got to use codeex I'm just going to go straight to codeex um similar like interesting yeah again I I was learning uh embedding whereas Where is it here? Talking about embedding models and things like that. It was just so much more so much nicer to learn through chat GBT about embedding models. So I'd prefer to just pay the 20 bucks on some my light usage of of claude of chatgbt. But outside of that like um I'm not a heavy user. So I'm not I don't need conductor. I don't need T3 code. But now they're really starting to release really nice features which makes it easy to just send off and control different models on different computers and on your phone and and through various different means whereas it's a bit cumbersome using claude on the phone on the code on the phone you know so yeah that's uh that's my view on these sorts of things and obviously so it just leaves kind of T3 code to to do this sort of multi- aent orchestration, I guess you call it. I've forgotten what what uh what Theo called it, but yeah, there is a name for it that is is he's trying to push sort of thing. Um I I did start watching his video, but I haven't u finished it. I watched like a few minutes of it and it's in my watch list, but I know which video you're referring to. Yeah. Yeah. I I used T3 code in the early days, but nowadays um I'm enjoying the other products that Claude has like claw design and the canvas. It's not performant at all. It's really killing my computer. But yeah, it's it's pretty good because like you you see what you're working on visually and you can on the canvas you can comment on things and tag Claude and it you send it off as a task to Claude one after another. So you can quickly go through things and comment and I really I really like that. Um you can tell it to make something editable and change something. Uh yeah and I've been doing this and I've been also doing a lot of uh starting new sessions with another session. So just telling the session hey you have this context this has been working really good and you know what I don't like as well. Can you start a new session to do something else but it's kind of like similar but it's different enough to be in a new session. So, and I've been getting like really really good good results from that. Yeah. Yeah. There's um I think we'll get like I spoke to a mate last week and he basically has like a master agent that orchestrates everything for him. So, he asks the master agent and then that delegates it to other sessions, other computers, other whatever. And it kind of it reminded me of um the git what do they call it? Scion uh that that Claude create or anthropic created their new git workflow where they don't check out a whole repo. They just ask that well they make the changes then ask the agent to resolve all the pull requests. It kind of reminded me of that where you're just asking an agent. There was something I covered today which also made me think of that which is less about maybe it was something that came into my head going over the deep seat stuff. But yeah, just this idea that rethinking things that we're just we're we're asking an agent to do the delegation and we're asking an agent to like orchestrate things as opposed to us deciding where to orchestrate things. it. Yeah. Yeah. Not at all. Yeah. We are going uh up in the ladder of abstraction. Yeah. For sure. Yeah. For sure. I have an agent for that like to to also help me with uh merges and uh pull requests and all of that. It's like just one agent that takes care of all of that. And if and the agent goes and reads the last message from the other sessions and if it needs it will ask them it will ask them something specific. So in opposed to reading the entire session it just ask like once what it needs to know and it's it's been working pretty well. It's been working pretty well and I like it because also I've been doing a lot of design with it as well and in this way if I need something very specific I can go to that session itself and be like okay listen son here we got to do this my way. So and that's been also working pretty well. Um, yeah. I think I think there's, like you say, we're we're all trying to think of ways I I want better remote because I'm all flumxed whenever I need to like think about, well, who do I send this to? How do I send it to them and and this and that and it's it's often can be quite cumbersome. Whereas I just want to sometimes just message on my phone and then it just gets sent out to the appropriate agent. Um, yeah. Cool. Well, we'll wrap up there. I think Yeah. I guess you're sick, right? I'm I'm Man, I'm I'm pretty sick and uh I've been taking painkillers for the entire trip and the weather and back. So, I cannot wait to have a proper like rest. Yeah. Yeah, I know what you mean. Yeah. Well, we'll wrap it up there. I think that's all we've got time for. Give us a hype. Give us a like, give us a hype, give us a like, give us a comment, give us a subscribe, wherever you are, check out the podcast, commando show.com will have everything you need to, you know, to to find all this stuff, but we're a podcast now. So, go get it on Spotify, Apple, all the clips and this will be up uh as of next week. So, yeah, really appreciate everyone who's tuned in for this week and putting up with me on my own for the first half of it. But, yeah, thank you for waiting for me. Yeah. Yeah. Yeah. Cool, man. Peace out. Cheers.