19:32
Silicon Valley nerds are quite literally offering fresh, brand new computers to AI. Like Mac minis are getting sold out across the US. Anthropic and OpenAI are freaking out. Every other YouTube video in my feed is talking about this thing. Open Claw. It's an AI thing. got my Open Claw brand new home. Like half the stuff that I'm seeing says that this is very dangerous. Half of the stuff that I see is just hype. So, which is it? Like why is Silicon Valley going nuts? And why are AI companies blocking usage? There's only one way to find out. I think we might be onto something here. Like I've been very skeptical about AI lately. But, I did it. I did the work. I did the work so that you guys don't have to. I spent the last month testing.
I spent way more money than I should have in those tokens APIs. I dove into this agentic madness. And bear with me now, but I actually really think this might be an iPhone moment. I know many people are concerned about AI like but this thing is the first AI that I see that weakens these big tech companies instead of making them stronger. Let me tell you about my rabbit hole last month. Like I don't think I've ever seen Silicon Valley so scared over a product launch as they are with this thing. There's a problem. Hold on. So, this is my scale of nerdiness. I ran this little survey last week about who watches videos in our channel. And this
is how you guys responded. So, how do I explain this deep tech mess in a way that is useful and fun for both the nerds and the not so nerds? So, for that, I'm going to use booze. Now first, I'm explaining what agentic AI and Open Claw are. And if you already know, then you can just watch me get a little drunk for your entertainment. Maybe. And then, I'm walking you through some of the biggest stuff that I've discovered. And finally, I'm going to cover why this could take power out of the big AI companies. I hope I'm not going to be drinking any of this today. Okay. So, the challenge is simple. I asked my team to come to today's shoot. Many of them are
on this side of the scale. So, if my explanation for this part of the video gets too technical, they'll use this. And I'm going to drink a shot. Now, let's say that this cube is a chatbot, something like ChatGPT. I assume that everybody with some internet access watching this video has at least played around with ChatGPT at this point. Now, what you have actually been talking about is an LLM. That is a large language model. Now, these things these these large language models let's say they're they're like computer programs that are really good at understanding and writing text. Now, they've been trained on millions and millions of words and books and Reddit threads, everything text that exists.
And so, they actually write very well. And they do that by Okay. It's fine. They just write text very well. Let's move on. Now, at the end of the day, the core trick of these LLMs is actually pretty simple. They predict what word comes next. And they are actually really good at it. And that makes them amazing at stuff that involves text, you know, chatting or summarizing essays or copywriting or coding. All these companies that you keep hearing about, Claude and OpenAI and Meta and Google Gemini, they've all built their own LLM. They're all slightly different. And they keep trying to outdo themselves. But, the point here is that they are just text readers and writers. And by themselves, they just can't do much else. And yet, this thing is what most
people think about when they're thinking about AI. And honestly, they're not that impressive anymore. I mean, all that hype when ChatGPT came out, that was a bit exaggerated. ChatGPT GPT-4 is here and is literally writing me the entire song. Let's see what this song sounds like. That big chat GPT moments, that was in 2023. Now, pretty quickly the nerds realized that these models, they could be made more useful. Well, for example, if you put another model in front of it that maybe translate your voice into text. This is an ear by the way, it's my rendition of an ear. Well, now you can speak and this thing will translate the voice into text. And because this thing is actually really good at text, it comes up with a text
answer. And then you add another model that's just really good at speech. And that's how you get voice assistants and chatbots and basically the movie Her and actually Her. Oh, what do I call you? Do you have a name? Um yes. Samantha. Now, we've had these for a while, right? Siri, Alexa, Stephen Hawking's voice, but they were all missing this part. Not Stephen Hawking. I don't like your tone. If you are looking for trouble, you found it. Yeah. Just try me, you But how do we go from here to getting actual work done? Well, another place where this thing clicked was in coding.
Okay. Which also means that the nerds they were kind of shooting themselves in the foot. I said this was a bad idea. Cuz these models, they had already been trained on massive amounts of text from the internet. And a lot of that text just happened to be code. So, it turned out that these were weirdly good at reading and writing and fixing and explaining code. And there was something even bigger coming. Now, if you ask your chatbot to write you an essay, for example, sure, it can maybe fake that decently well. But let's say that you ask it to write an entire book. Well, now you have a different problem. Cuz for a human, writing a book is not just typing text. Even for a human brain, there's there's stuff that
we would need to do. We still need to plan it out, right? We need to define what the style is going to be and then probably write an outline and then come up with a work plan of, you know, when we're going to write all of these chapters and actually get to writing the chapters and then, of course, review what we worked on. Right? It's It's a process. And that's when normal chatbots start to fall apart with these large projects. You may have noticed this in long conversations with ChatGPT. It kind of forgets what it was talking about earlier. But what if somehow you could get the chatbot to just like you to break down the work into steps, right? To come up with a work plan, to review each of these steps
and to try and write a book. If each of these writing tasks is small enough, simple enough, then maybe you can actually get one of these dumb LLMs to do it. As long as you have some organizer that takes care of it. Now, this planning and then doing idea, this is what OpenClaw is built around. This planning part is something very similar to an LLM except that it's been tweaked. It's been tweaked to make it really good at planning and supervising that all of these tasks get done. Ah. Okay. That's fair. Maybe I went too far. Okay. So, let me rephrase. This thing is an agent. It plans out tasks and then it does these tasks. It works because things are getting split into simple, small tasks that LLMs can do despite these
guys not being that smart. But anyway, once you understand this planning, you're ready to understand most of what we're going to talk about and OpenClaw and what this what people are going crazy about. Now, this planning side started popping up in chatbots around 2024. You probably seen it as a feature called thinking or reasoning. And when you enable it, you can see the chatbot like first taking longer because it's writing its own little plan. Sometimes it tells you about it and then following that plan. Like an agent. Already drank a shot for agents, guys. Okay. The point is that all those tasks they're still text tasks. Reviewing or writing longer texts, maybe brainstorming a bit. Now but the real
problem with those agents that we're getting that these companies are giving us is that I mean they're kind of prisoners. They're prisoners of the browser. We can't really connect them to other things. We maybe we I mean we could eventually maybe have them manage our calendar or filter our email, but they're locked. They're locked in these little windows. They're they're inside these platforms and they're behind all those paywalls. So we depend really on what these companies, you know, want to let us do with them. And more importantly, we don't know what these companies are going to do with everything that the model is reading. So
I would not want one of these guys to be reading my emails. I don't want Open AI to have access to my email or my stuff, right? And that's where this guy actually comes in. Open Claw. Clawbot. Open Claw. Open Claw is an agent, right? Just like the other ones, right? But it's it's it used to be trapped here. Open Claw is an agent, but if you install it when you install it in your computer and runs in your computer, it's it's not in Open AI's servers anymore.
It's not in a jail. It's free. And it's also actually free to use. It's also open source, which means that the guy who made it completely free, made all the code visible for people to read to make sure it's not a scam, but also to edit and to improve. It became the fastest growing open source project in the world. So you install Open Claw in a computer. I actually did it myself. I installed it in this old MacBook Pro that we had laying around. And then you can connect it to whatever you want to message it with you WhatsApp or Telegram or Discord. I actually talked to mine through iMessage and I called him
Good evening, Dave. Hey, doing Hal? Everything's running smoothly and you? Oh, not too bad. Have you been doing some more work? Now, the reason people are buying Mac minis is that they very much don't want to give their main computer to open claw cuz that is objectively dangerous. Like these LLMs, they're all a bit like black boxes. We should have printed black boxes. Okay, they're a bit like black boxes. We don't understand them very well. They sometimes do weird things. So, we don't want this uncaged thing living in a computer where you store your personal stuff and your credit
cards and your passwords. Put it in a clean computer that's that's that's just dedicated for this. Yeah, you can give it access to the computer, too. Maybe to the browser to search and do things and to write files. Now, some people don't want to buy a new laptop or give it or just want to take it for a spin. So, they're using platforms like Hostinger. Hostinger lets you host No. Okay. It's going to be the last one. I'm I'll stop now. Okay. So, how do I explain this easier? All right, so Hostinger basically rents you a computer in their data center.
It's still your computer. You can install open claw in it, which has advantages like being able to access it from anywhere. It doesn't waste electricity in your house. No power outages, no internet hiccups. Now, my first open claw attempt was a mess because my old laptop kept going to sleep. So, Hostinger was really useful to trust that my bot Hal was always going to be running. They rent you a computer in their data center and they built a one-click deployment specifically for open claw and I timed their landing page claim of being live in 60 seconds and it actually worked out. Now, I mentioned that I talked to Hal via Discord, but you can use anything that you want. You can use WhatsApp or Telegram, anything.
Hostinger's implementation has built-in access to both WhatsApp and Telegram. They also have built-in security. They handle the security patches, the updates, and the backups. It's It's like a headache-free version of this stuff, especially if you're trying to level up for the first time. So, I emailed the Hostinger team, mentioned that we were working on this video. We've worked with them before, and they agreed to give us a discount. So, if you go to hostinger.com/slightbean, which I'm going to link here or this QR code, you can just pick your plan, and the Open Claw app is automatically preselected, and you get a 10% off. Now, by itself, how can't do much. Remember that this guy is just the organizer,
it's the planner. You still need the other models to do the final work and those all of those tasks. Even the planning itself needs a model. Like, this actually needs a model. And of course, I'm not going to get into that. But, another thing is this old computer, this is not a giant data center like those that OpenAI and all of those companies have. This is just an old computer, right? So, it can't run these powerful LLMs like ChatGPT. So, you need to connect to them. There are others, right? There are these small models, maybe they're called SLMs. That's not what they're called. But, those can run. You can actually run those little models in a home computer.
They're not very smart, but if you install them here, and they run here, that would mean 100% complete privacy. Then, you don't even need an internet connection to do it. What? Okay. You're right. I went too far. Anyway, these small models are not very smart yet. So, what most people do is connect Open Claw to one of the big ones, right? To the OpenAIs and such. The brain part, the planning part, is still in this computer. And because this doesn't belong to any of those companies, you can even mix and match.
You can use Google's LLM for one thing that it's really good at, and you can use Claude for another thing, if you want. Okay, okay, okay. I'll stop. Okay. Now, that connection is a bit messy, cuz you have to sign up for each of these LLMs, and you have to connect them to your open claw using some terminal stuff that again I wasn't too familiar with. But Hostinger can also help with that if you're going to use that. They have built in AI credits and you just add credits to that Hostinger account and then your system has to feel to operate. They also have this built-in web scraping so your agent can browse the live web immediately.
Okay. What have I done? I mean I what have I done with open claw and just with the miscalls. Okay, let's start with some simple examples. There's this software engineer in Brazil. Uh so he built an agent to rewrite his CV and have the agent directly apply for jobs. He had the agent navigate the browser on the computer, this browser that was logged into his LinkedIn and then the agent wrote a CV from the LinkedIn stuff and started signing up for job application sites. He got 100 job applications, six different interviews secured by the agent and he actually took a new job. Now, you could not have done this with a regular chatbot window cuz you would not want to give ChatGPT access to your LinkedIn. So in this case, the
agent that was at the helm not just does the things cuz if something goes off with the sign-up process or something doesn't work, the brain agent actually acts on that and it often fixes the problem without needing any human intervention. In his Reddit post he said that the model doing the work was Claude Opus. That's the LLM that took the task. Now, I have thoughts on which model you should be using I've I've tested a few. But first let me give you some of my own examples. Hal. Hal is Hal exists. He actually check it out like I have his contact and everything. Hal lives in my house and controls my smart home. We're actually producing a full video about smart homes. It's a big rabbit
hole. You should subscribe if you want to watch that. But for the office we have another local open claw. It's called Megazord. It has its own desk now. And Megazord has different agents that do different things. Now I trained an agent that I called Carl to work as sort of like a YouTube police for us. You are Carl. I am Carl. You will get us a million subscribers. I will get you a million subscribers. To clarify what an agent is inside Open Claw. Open Claw has this main agent. It's cute. When you put it for the first time, it kind of asks you to define its own personality. So, you can make agents inside Open Claw and you can make agents that are specialized in different things.
Now, that's very useful for this thing called context windows. Now, context window is how much a model can remember about what they're working on and what they were trained on. Ideally, you want a model, of course, to know everything, right? So, so it's better at its job. But, if it doesn't fit in his context window, you kind of have to pick what you want to train it on. So, if you give a model too much context, it'll just forget, right? It'll become less efficient at using that context and that training. Now, Gemini, for example, now has a 1 million token context window, which is crazy. That's the equivalent of like 2,000 or 3,000 pages of text. This is an expensive This
is a frontier model. Every time you run it with all that information into the prompt, you're going to get billed for that. So, you can separate the context between agents. You can have them be really good at one thing while limiting how many tokens you use. I certainly didn't know this when I started Open Claw, and I am not proud of how much the bill for all these APIs was the first month. Now, we've been making videos for this lightning channel for almost 8 years now. We still haven't gotten a golden plaque, and we hope this is going to be the year. But, but if there's one thing that we suck at YouTube, is packaging videos.
Make making like clickbait titles. That's I think that our videos aren't that bad, but I think our titles suck. They just don't get people to click. And we're sometimes picking the wrong topics. So, we wanted a bot that we could use to like run ideas, to get a second opinion about them, to look at our audience stats, and like what else they watch, right? And to bounce ideas when we're brainstorming. Again, connected to real data from our channel. But, anyway, Carl is connected to our YouTube API, knows the performance of our videos, which we do using the Google Cloud Console, and we also trained Carl on the stats of our audience, right? We trained Carl on the related channels that our audience watches. We trained him on some of the lessons from a few YouTube
consultants that we worked on. And we literally gave Carl the transcripts of our calls with them. So, so when we tell Carl a video idea, it's going to run it through this checklist from all this context, all these lessons, and then he provides feedback and tells us if he thinks the idea is greenlit. Often spots things that we haven't seen. Like, all this stuff, right? All this stuff that we trained Carl on, that's our secret YouTube sauce. We wouldn't want anybody to look at these numbers. We don't want to upload these numbers anywhere. But, thanks to Open Claw, they run here on the laptop. Now, if you use Hostinger on this, even though it's in the cloud, your chat logs and your data, it all stays in your private isolated vault
computer. No third-party AI company is reading your data. So, Carl uses Gemini Pro 3.1, which for now has been my favorite LLM for this task. I've tested quite a few, not sponsored, that's been for me the best so far. Now, I had tried to do this with the ChatGPT assistants, but it's just really hard to get them to do any work. Again, they're really trapped in their little cells. And also, again, I don't want my data in ChatGPT. So, oh, to connect Open Claw to things, uh you skills. And this is why I said that this could be a really a potential iPhone moment. Because all these connections to LinkedIn, to the YouTube API, the best way to do this is using a skill. Think of a skill kind
of like an app or like a plug-in. And, there's a whole skill app store for Open Claw. There are skills to connect it to Notion, to connect text-to-speech, or to connect your calendar. There's a self-improvement skill, which is actually the most popular in the platform. And, this is what Remember, this is what made the iPhone back in the day. It wasn't the device itself, it was the apps, the ecosystem of developers. Thousands of builders, founders who found the iPhone as a platform to build things, and they've now become billion-dollar businesses for good or worse, I guess. Now, the danger with skills, at least today, is that they are
community-made and community-reviewed. And, many haven't been checked for malware or backdoors or code injections. So, there is a real effort to get all of these vetted, but again, because this is open source and because these things are really good at coding, you can just ask it to build your own skills. So, I had it look at this skill to download YouTube videos, which we sometimes reference in our own videos. So, I'm by no means qualified to review the code this code for any malware. So, I just asked Megazord to build its own skills, and it did, and we created our own version. Now, there are tons of other examples out there. Uh there's this other guy that built this pocket-sized
personal assistant on a Raspberry Pi Zero W. It's like the size of a credit card. Cost him like 15 bucks in parts. And by the way, all of the stuff that I built All of the stuff I built was on just on Google searches, on YouTube tutorials, which I can I'm going to link a few below, or just asking Megazord how to do it. You can find infinite tutorials online. That, I think, is the very most important part of today's video. Guys, we just we can't stop AI. We can't cover the sun with a finger. It doesn't matter where in my nerd scale you are. I think that suddenly all of us, all of our jobs, including my job, honestly, as a writer, YouTube host, all of our jobs are in, one way or another, threatened by AI. We can't deny
this. We can pretend that it doesn't exist, or instead, we could try and leverage it. I think for the past 3 years leveraging AI meant giving AI your data, giving AI your conversations to someone else to use to train and replace you quicker. But I think that not anymore. I really think that Open Cloak could solve that problem. Because now we have thousands of builders that are using this platform to build things. It's the App Store moment all over again, maybe. And the moment that you realize that Open Cloak is basically this new operating system, you instantly, hopefully, understand why all these AI giants are absolutely terrified about it.
There's this Silicon Valley story that when Steve Jobs introduced the iPhone for the first time, the team at Google they were working on their new phone. And they looked at the iPhone, they looked at what they were building, they realized that the world had just changed. What we're going to do is get rid of all these buttons and just make a giant screen. And they actually threw their work away and started over. And it kind of happened again just a few weeks ago. There was this team at Google that was working on adding Gemini, which is Google's AI, to Chrome so that the browser could do things, could help you do things like book stuff or shop online. And they also threw away their
work. You can kind of start judging why these all these AI companies, they're really scared about this. Cuz the for the first time ever, for the first time ever, they've they've increased their pricing, they've put much harder limits on you using their models when they're outside of their cages. All this time they've been operating these models at a loss because as long as you remained in these little cages, they could trap you into their ecosystem. And they knew that they were going to make money off of you eventually. There's this big spike in use that's coming from Open Cloak is so large that they had to put a stop on it. But once again, Open Cloak nerds, they can just switch. They can just
choose a different model that offers a better deal. These companies, they're truly scared. Their entire business model breaks if you are not locked into their systems anymore. Now, if you're wondering which models we're using in Megazord, by the way, our standard model is Gemini 3.0 Pro. It's been great at advanced reasoning tasks. Now, I also have an agent called Was that runs on Opus, which I found has been the best for coding and for building. It's just really expensive though. And smaller tasks, we have like a short performance report and the video downloader tool we built, like those all work on GPT Mini because pretty simple tasks. But we are not taking power away from all these big tech companies, which
I'm always going to be in favor of, honestly. I did Google Translate, so Sahasya is that the right is that the I don't know. And for us, right? So, for us, for you and me, what this means, like what this means is that the next iPhone, the next technology that really changes the way we work might be here already, and we actually, this time, we don't depend on anybody to make it useful. Like my favorite example of this is this researcher at Meta, who works with AI alignment, someone whose actual job is to test these things and make sure that AI isn't doing things that it shouldn't. Well, she gave Open Claw access to her inbox to help her manage her email, and then the agent decided to manage it by deleting
everything. So, it actually wiped out her entire inbox. So, you have to actually pull the physical plug on it. Um but, you know, at least for now, all these agents, they have an off switch, right? It's our computers, you can turn them off. That's much better than these big AGI things that big tech are building. Anyway, I hope this is a good summary to level up the playing field here. If you didn't get lost by these explainers, you should level up by looking at our Startup Club channel. That's where we really nerd about building tech stuff. Catch you on the next one.
