Hello friends. How are y'all doing? Hello friends. Happy Valentine's Day. We have started. Can I get a mic check? We good? Wow. Okay. Wait, let me like quit my models. I think I'm running too many right now. Okay. Okay, that should be better. Okay, stream rate should increase now. Please increase. Is it increasing? No, it's not. Okay, I'm going to kill all my models. I hope. Okay. Yeah. Okay. Is it? Be merciful to your models. Dang it. Okay. It should be good now. Sorry guys. Yes, that was a great demo of what just happens today. Of what happens when you run try to run too many local models and live stream at the same time. There you go. Hello friends. Well, very good to see all of you guys. Let me get started now after that um that beginning. I hope you guys are all doing well. Okay, so let me go to the slides. Okay, so let's talk about what happened really like 2025 just like even beginning of 2026 there has been a lot of things that has changed. So, I wanted to like cover that um for you guys and [sighs and gasps] Okay, just want to get like a check. We good? Okay, we good. All right. Yeah. So, I wanted to cover that for you guys. Um with all the new things that are that have just come out, it's really shifted the way that we're thinking about AI and especially for AI agents. like it's an understatement to say that it's become like so much more powerful now uh because of the changes that have happened in just the past few months. Does anybody want to make a guess into the chat? What do you think has been the singular biggest change that has dramatically changed um how powerful AI agents can be? Put into the chat. Okay, I guess I shouldn't show you that. I want to see what you guys say. It should be working now. Yes. Coding. [sighs] Any other? It's not hardware. OpenAI models MCPS platforms and tools orchestration. Interesting. Agents open source. Yes. Alias. Exactly. Open source, massive changes on open source and it's really like unlocked so much just within like the past two to three months so crazy right open source models yes that's it open source models thank you guys I am good I think with my audio now right so yeah I do want to cover some major shifts so flagship models uh there's been developments there and then the open source explosion so I really want to focus a lot on this part and then this and this lead to things like this like openclaw. Um some of you guys may have heard of that. Uh drop in the comments if you have heard of or used openclaw. Just write openclaw into chat. So I do want to address some of these as well. And then why all of this leads to a lot of this which is very important um when we're building AI agents. So yes, we have a lot to get through today. So I shall get started. So first thing is that there has been continuous improvements of the flagship models. So the past year has brought a lot of improvements in AI model uh capabilities like some the most common one like the most salient examples right now are GPD 5. 2 from open AAI and also cloud 4. 6 from anthropic that just got released and they did it at the same time you know um and compared to the previous releases there has been really good increases in the metrics like in the benchmarks um for these models. So I I'm not going to go into too much detail about this. Uh you can try it out yourself using like a variety of tools uh either directly through open anthropic or like perplexity or whatever like you know wherever it is that you want to try. You can also try arena um models. You'll notice that they really are much better. So we're seeing like good development of the actual flagship large language models themselves um and getting those released. And because of that kind of like the foundational layer, right? The foundational layer itself is getting better. Um it's smarter. There's more context. So up to a million tokens and then also they're more autonomous can work independently a lot of the time uh now and can do like a lot of things by itself. It can do things like code applications and control your computer um and it's able to retain enough memory so it's able to continuously have that conversation and be able to continuously work on projects. So if you do try them out you'll notice that there really are is dramatic improvement in the base models themselves. So that's number one. Um yeah so there's in terms of coding excellence so it's really just across the benchmarks there's been like massive development so coding excellence longer running tasks advanced tool use and also better reasoning so these are great because these are the capabilities that will unlock a lot more possibilities for AI agents themselves okay so keep that in mind yeah base models are developing but really the really big thing that has changed a lot for the AI agents world um is the rise of open source. So the open source ecosystem has really exploded. I would say it's like past year. It's even less than the past year. I really want to say like the past 3 months really. Um and here are some examples of top open source models like you have your llama models GLMAX and large uh Kimmy K2 Mistro Quen and some frameworks as well. Can you guys put into the comment have you tried any of these open source models before? Just type in open source if you have tried open source models before. I want to see how many of you guys have been following. I think my audio should be fine right now. Heard of openclaw. Okay. So some of you guys have heard about openclaw before. Okay. Uh the fifth name open source. Yep. So some of you guys have heard of have tried open source as well. only open AI. GLM is awesome. Quen quantize open source. Okay, cool. So, some of you guys have played around with open source models. Uh, I think that's great. And I think if you did, you would have noticed that you would have really noticed how big of a difference it has been just within the past few months. Like the development is massive. Like previously, people were just like, oh, you know, open source whatever, right? Like it's that's nice. But it was not even close to performing at the quality of closed source models. But just like in the past few months, it's like really getting to be the equivalent like around the same as closed source models except open source also has really big advantages as well. Um, so yeah, so here are some of the like numbers here. Oh, also like what's really crazy is that these models, all of these except for llama, these are all Chinese models. That's like really interesting as well from like a geopolitical standpoint. there was like this really big shift between the US uh leading open source model development to a lot of Chinese models being released that are open source as well. So um the good thing for consumers right is that for most of us like it doesn't really matter that much um in terms of like what's happening from that side but if from the consumer side it's actually great cuz what is allowed is it allows like a lot of more competition that's there and because these China Chinese models are like just as good um they're so much cheaper and they're actually free to use if you host them and things like that. It's also forcing other companies to start releasing open source frameworks and open source models as well. So overall things are just becoming more open source which is a good thing for all the consumer side and the reason why open source models are really great is that it allows things like full control. If you're using a closed source model like open AI or anthropic or something like that it's great but you don't have complete ownership of these models and the data. um you don't have like privacy like on the privacy side you'll be sending data to these models and they'll be like training their models or doing whatever it is they really want with the data that you're sending them right like you don't actually know what they're doing um open source models are not like that because you are hosting them locally like whether on your computer which is why my computer was messing up earlier because I was literally doing it like locally um and also uh or you can do it like hosted on your own uh wherever it is that you want to be hosting it, right? But you actually have full control over this. This is very different um compared to using a model that's hosted by a company like Anthropic or OpenAI. It also allows for customizability. This means that it's customized. You can fine-tune on your data. You can modify the architecture and optimize it for specific uses use cases as well. It's completely flexible. You're not dependent on a third party now. And there's no vendor lock in. A lot of times like you're paying for something especially on the enterprise level um like open AAI for example you're kind of locked into the open AI ecosystem so it's hard to like switch models switch platforms and infrastructure as well their incentive is obviously to like lock you in as much as possible into their infrastructure right um so you don't have that problem with open source models are also significantly cheaper so 50 to 90% cost reduction so the running of the models if you host them they're actually completely free so the cost the minimum cost that you're paying is just like if you are hosting them on something else, you're hosting like other infrastructure around it as well. But the models themselves are like you they are free to use which is so crazy, right? Especially if you have like very high volume use cases. Um just it's just so much cheaper to be able to use local models. That to me is like one of the most attractive things about open source. So this and also the full control and privacy reasons, right? It's also auditable. So this is a really big thing on the enterprise side. um for companies a lot of times like when you're in a regulated place like health care for example or like finances a lot of our clients that we work with they're in those sectors and like a really big concern they always have is like I need to be able to pass the audit right from because they have like these restrictions in their domains and the audits would have specific things that would not allow them to use closed source models because that would be sending client information or patient information into a third party which is not allowed. out, but because of open source models, they're able to have full control of that data as well as the models themselves. So, it becomes auditable. They're able to still build stuff despite being in industries that are really heavily regulated. So, this is really, really massive. There used to be like so much money that's spent on figuring out how to make closed source models auditable, but with open source, like all of that just becomes not important anymore. Um, and finally, there's like fosters innovation. So communitydriven improvements because it is open source, everybody can access it and develop on top of it. We can just see there's so much innovation that's going on top of this now. Um like stuff like Open Claw for example, which we'll talk a little bit about later and just in general. Um there's so many developers that are able to build on top of these open source models and make them better, fine-tune them, build infrastructure around them as opposed to just relying on the closed source releases from the big tech companies themselves. Of course, there still are cons to open source models like it's higher setup complexity. I I'll actually show you guys like how to set it up in just a bit. It's not that much higher, but technically it is higher because if you're using a closed source model, you just go like openai. com, right? or like chat GPT. Yeah. And then or like anthropic. Um you can just like literally type it in and start using it. But there is a slightly more complexity because you actually have to like download it and then set it up. And it does require good hardware. For example, I'm on a MacBook Air right now trying to run like a bunch of local models and live streaming at the same time. Not good enough. So you do need to have decent hardware um to be able to run these models and continuously do so as well. They're usually also a little bit weaker out of the box um in terms of complex task um and does involve you like configuring it and adding on other add-ons as well yourself um as opposed to closed source where you can just directly go into it and also self-management is required. So you do need to handle your own security scalability uptimes updates and things like that because there's no service support. So the good news is that a lot of these drawbacks are decreasing a lot um because of the ongoing innovation that's going on over here. So the ecosystem is maturing really quickly. Now it's actually really not that hard to start running local models like you don't even need to know how to code like at all. Um and the best use case is with open source models is high volume applications with privacy requirements where you have techno expertise start with cloud APIs migrate to open source where you hit scale privacy constraints. So it's like this is the way that you can um start developing um when you start building out things on the enterprise level. Let me see if anybody has any questions. I'm not a huge Okay, so from Lone Wolf and Cup, I'm not a huge fan of open source. It's open source till it's not and you have to pay monthly after you work for free. That's not true because you are downloading these models like they're literally on your local computer. So it's not possible for them to get you to pay monthly after that because the entire infrastructure is just there. It's sort of like if you've used coding tools like Python and stuff like that, that's like open source as well. Like you don't pay for Python. Like you literally just download it directly and use it directly. That is true. Mistro is French. Yes. Missed that one. Yes. People underestimated MacBook Air. I mean MacBook Air is good enough to run two local models um 8b like 8 billion parameter local models at the same time. That is like the maximum that I've managed to test. But I feel like I need to get it better. I used to have a PC and then when I left San Francisco, I gave it away. So I'm like, I should have kept it. Oh well. Um, will this video stay on YouTube? Yes, it will. It will. Yes. Heard about Kimmy. Haven't tried open source yet, but have read about it. Uh, hello Tina. Hello. Nice to meet you. uh the Chinese take the already created US models and then retrain them to produce them. Thus the large I'm not going to go into this that is not the case like if we could actually do that and just like steal OpenAI's models it would have gotten leaked way prior to that. No it's not you can they they've been publishing papers on this if you actually follow these companies they have been publishing papers on how exactly they've developed the models that they're doing. So no, that's that is not the case. Uh you cannot trust any of them. I mean, okay, like if you're going to be super paranoid about this, it's like I don't know what I can say about that. Like though that paranoid can be there, but I think the major concern that people have is like, oh, like uh I don't want to use Chinese models. I'm so scared. servers. Right? The whole point of open source is that you're not using anybody's servers at all. So that's the key thing that I want to emphasize here. The key thing is that you're literally taking these models, downloading them, and doing stuff to them. So, it doesn't actually matter if it's like a French model, a Chinese model, like a US model, wherever it actually is. Um, these are just models that you're downloading directly or you can host them directly yourself. So, that's really not the issue here. Um, are you going to teach how to make AI agents in Python? So I'm going to show you an AI agent that I did make in Python. Yes, I will do that. I'll be running locally, but yeah, it's if I find myself use still using online models, shifting in between. So there that's the setup part of it, right? Like I think if you really want to have something that's continuously running, you do want to be ho hosting it yourself um in in like a VPS. I trust my mom. True. Me too. How do I actually update the model? You download the newest version. Yes, So, there's this thing called OAMA, which I'll actually just show you guys in a bit. But like this is how it is that you can take it and then that's how you get a lot of like open source models to be able to be used as well. Yes. Okay. Any other questions? Oh, by the way, if you want to get these slides as well, um if you want to go into more detail about there's like links and stuff that you want to check out like stuff like that, I will put the link here. If you sign up for this mailing list, then we will send you slides. It is a free mailing list. One second. Okay. Too many tabs open. So here also we are we our workshop uh our lonely artist boot camp is also going to be launching on the 17th. So I also want to just put this here if you want to sign up for the wait list we sold out within an hour um for the past few cohorts. So if you are interested in learning more about not open source specifically there's open source as well of learning about developing like make creating agents. We also cover things like open source how to use those with models and things like that. If you're interested, you can check it out at this link over here if you like. So, check out this link to sign up for upcoming agents boot camp and this link to get the slides for this specific live stream and upcoming live streams as well for upcoming four slides. There you go. if you want. Okay. Uh I see we have some rage baiters today, huh? I'm going to ignore the rage baiters. Um will you cover rag with local models? Yeah, I actually will. All right, you know what? Let's move on because I'm going to show you guys some demos of how to like set these things up in a bit and I think that will answer a lot of the questions of how do I actually build these things? Nothing other 70 bill 70 billion parameters is useful to run. I really feel there's like rage baiters today. That is not true. I'm really literally running like a 8 billion parameter and it's great. You do not need 70 billion to to be running right now. I don't know if we're just rage baiting today or we just have a lot of misconceptions. If it's misconceptions, totally welcome to that. Like please ask these questions. I really want you guys to know this because the open source community like it's I really want people to start building with this, right? it is free and it's private. So, it's like there's on most computers can handle at least some of the smaller local models that are still really good. So, I feel like now is like the time if you ever want to start building agents um and you're like, "Oh, like but I don't want to pay for like OpenAI credits or like anthropic credits because like because it's so expensive when it scales up. " Well, now you can use open source models. So, that's not an issue anymore. So, let's talk about the open source AI stack, right? So you get your like popular open source models like Kim, GLM, Quent, whatever like all these things. What you need to do is that you need to download something called Olama. So it's like a it's a model manager. It's also a developer tool. So this is like the software which is over here. I will show you this thing over here. Click the button download. Um this is going to be able to download and allow you to manage your models locally. So, you can after you download this, you'll be able to switch between different models um and download different things. I'm not going to show you this right now because if I try to download some models right now, I think I'm going to crash this entire live stream because I already I don't think live streaming and running models locally on a MacBook Air, as you can see, that does not that's not possible right now. But really, it's super simple. Just click download and then you can just decide on which models that you want to download. Um I would say like if you want to get started, uh try some of the Quen 3 series. I think that's a pretty good like general all-around pretty decent models. Quen 38B um if you have like a MacBook Air type computer, you should be able to handle it. Or you can do the 4B the 4 billion parameter one um for like other pretty much all computers can handle the 4 billion uh parameter model and it's still pretty good. So after you get these models, it's really not that different than building agents that we've I've talked about made a lot of videos about that. um building agents previously, you're still having the same infrastructure here. It's not like open source is like some whole like magical like situation. It's not it's just the models themselves are open source as opposed to closed source models. Instead of this saying like GPT5, it says quen 3, right? That's it. Um and you have your local your package manager. So you still have models and in order to build agents, you need to give it tools, knowledge and memory, audio and speech, guardrails and orchestration. And that's how it is that you are able to build um all sorts of different agents out there like things like personal assistant agents um is stuff like booking agents, email agents, credit card agents, financial agents, all these things. So um that is how it is that you build agents. Uh not going to go into too much detail about this. We literally have a 28 day like boot camp out about this. So we teach you all the fundamentals of how all this works. But I do want to like in this live stream make it really clear to you guys like the building agents like the fundamentals of how to build agents keeps remaining the same. So I just think that there's so much like hype and new things that are coming out all the time, right? But in my mind if you just focus on the fundamentals of things, it will help you understand what is actually important to keep up with and what it is that's actually changing in context, not just like the new hype thing coming out. So in this case when we say uh if you just remember this like literally this entire presentation there's like one thing that you can remember just remember this agentic systems are composed of these six different components and we have new developments in terms of all of these different components each time we have a development it improves our ability to build agentic systems in this case because of the open source revolution um as well as other models that are getting better this is going to allow us to have better agentic systems make sense. that made sense to you, put agents into [clears throat] the chat. Uh, and some of the popular frameworks and tools that we use to build agents would be stuff like Lang Chain, Langraph, Crew AI, Autogen, OpenAI agents SDK, NA10, and Google ADK. [clears throat] What's cool also is like because so this is a great example. um open AAI like agents SDK used to be such that you can only build using open AI uh models right but because of the open source pressure they've made it so that the SDK also supports other models now too so open source models included so it really shows how this like whole like open-source um shift is allowing more and more it's pushing like more and more companies to also release things open source as well which I think is great for the consumer Any comments? Agents, agents. Amazing. Thank you guys. I'm really happy. Just remember that nothing has fundamentally changed. Don't freak out about any new things that are coming out and new hype tools everybody is like talking about. Just remember these fundamentals. That's it. Looks like Tina has a gun on her head. What? Oh no. I hope Oh my gosh. I hope I don't have this. No, I shouldn't have said that. Okay. If this stream goes down, it's on you. It's on you, bro. No. Okay. Can I do a develop mobile app? Yes, you can. You totally can. If you're hosting it, you can then connect to a mobile app. That's completely doable. Yes. Okay, great. Um, 28 days. Yes, the boot camp is 28 days because we're teaching you everything from the ground up. I really appreciate your boot camp. The fundamentals are proving to hold true. Thank you so much. Yes, we are very fundamentals driven. See any other questions? Don't waste your RAM on rage baiters. Thank you. One minute fats. No more rage baiting. Why people keep asking weird questions? I know, right guys? No. Stop rage baiting me. No, no. Um, okay. I think you guys understood all of that. So, very happy to hear that. Okay, great. Moving on. Okay, so I want to show you guys how it is that you can do stuff and um actually run these models, right? So, the demo I'm going to show you is going to be Oama plus NA10. So, again, I can't I'm not going to show you the actual downloading of it just because I don't want to like kill my computer right now, but super easy. I promise. Like what you do is you go to Oama over here, you click download, right? Um and then after you do like the installations and things like that, you are going to be met with let me actually show you guys what that looks like. Not that. Ignore that. So, what it's going to show look like is this. So, Olama, right? This is on my um it's just a software that's running locally right now. It's going to be here and then there's like a bunch of models that you can choose from where you can search other models as well. Um I would recommend trying out some of the Quen series just because there's a really large amount of choices that you can have. Like almost everybody can run Quen 3 4B like 4 billion parameters. This one's really quite good. If you want to get like really good, you also get the Quen 3 coder. Very good as well. Yeah. So, you can try out like a bunch of these ones. There's DeepSseek, Gemma, GPT, Mini Max is really good as well. GM GLM models are really good, too. So, you can just try like whatever it is that you want to do. You can download it. You can also just like talk to it and play around with it right now. Again, not going to run it right now just cuz I don't want to kill my computer because this is actually running on your local machine, right? So, you have that part. Um, and yeah, so you can already play around with it directly here. But what actually makes it like super powerful is when you start building with it. So for a noode, I think the easiest way that you can start building agent systems with a noode approach um is by using NA10. Um I'm sure many of you guys have heard of that already. So you can also just download NA. You want the self-hosted version of NA10. So you go to NA10 um self-hosting NA10 and then you click the Docker installation guide. So this is the one that I would recommend. I think it's pretty easy to do and just follow these instructions here. Or if you want like an even easier version of this if you're okay running a little bit of code, just go to this um self-hosted AI starter kit. I think it's super easy uh because it includes a bunch of things in it already. So it includes the NA10 self-hosted NA10. It also includes OAMA so you don't have to download it yourself. And it also includes quadrant which is for vector store and then Postgress SQL stuff for like your data safety and stuff like that. So these two also come in really handy when you want to start working with data. um that's being stored. So yeah, you can just go through the installation process here based upon whatever it is that your hardware is. So in my case, I would for Mac and Apple silicone users, I just do this. You just open up your terminal and then you type these things in and then it should be able to get started very quickly. So I personally I this is what I did like I think this is the easiest approach because it has everything just like bundled together already and all you have to do is run it. Um otherwise you can also just download these two things separately. Super easy. Then all you have to do is go to localhost. Um, so here's an example of one of the models that this is like a really simple example of like a small agentic system that I built. Um, so what it does is you can execute this. I have some credit card statements that's stored locally on my computer right now. Like on my computer, I do not want to be sending this type of information online to other people, right? bad. Not going to do that. Um, I only want everything to be completely local. So, it's able to read these credit card statements and then it's able to extract that information and send it to an AI agent in order to analyze those credit card statements. I'm actually not going to show you these credit card statements because I'm being very serious. Like, these are like my actual credit card statements that are here. Uh, but believe me, they are like actual credit card statements and it's just like stuff that this is my corporate uh credit cards. So, it's just information um about these statements. And just to like prove it to you guys, I will show you where they're stored on my local computer. Don't come and hack me, okay? And steal my credit card statements. That would be sad. Okay, let's see. Let's see. Okay, never mind. You're just going to have to believe me on this one. I don't know where my finder window is, but just believe me on this one. It is stored in my credit cards. Uh my statements are stored locally. It would extract that information. And then the lama here, um it connects to Lama, which the agent specifically that we're using here is oops, let me check. Let me go here instead. The agent that we're using uh sorry, the model that we're using is the Quinn 8 model. Sorry, Quen 38B model. Um, and the prompt that I gave it is analyze his credit card statements and category spending for each month, including a total for each category itemized breakdown blah blah. Just like summarizing these trends and also providing specific advice on how to reduce spending, right? And the output. So, I'm going to show you something that was run previously um is that it would add so I had three statements that were there. It would analyze these statements and actually go into detail. Oh my god, like this is my real statements. Look how all the money that we're spending corporate on the corporate side. Um, so yeah, it would actually go through this information and then it would categorize it like subscriptions and digital services over $1,000 on these. Travel and accommodation for work, shopping and merchandise, um, $1,616, dining, entertainment, this is mostly with clients, education and courses, $454, $445, other services, and foreign transaction fees as well. Right? So it would go through um the information, spending trends over time, and it would give you like specific suggestions. So, this is a very simple example of what you can do, but I hope that you can that you guys can see like how good this is, right? This entire thing that I just ran and I can keep running this continuously. It cost me zero money and it's completely private because it's on my local computer. That that is such a big deal. Does that make sense? It's like such a big deal cuz say like you want to do this and you want to have a continuously monitoring which by the way like I actually do have um not going to show you guys that version of it cuz it is my private stuff. It you can continuously be calling APIs um and getting or having these files like credit card statements, financial documents, bank statements, all these things are being downloaded on your local computer. It stays on your machine. You're analyzing it. You're able to do you're able to figure out like what the trends are, what you should be doing. You can also give your agent tools such that it's able to like unsubscribe for some from things for you. Um it's able to dig into potential issues that you may be having. It can invest for you. There's so much that you can do and all of this is completely um within your control. And this does not actually involve code. Like you don't need to code in order to build this, which is so crazy. It's so crazy. I want to see if you guys like have any comments about that cuz like it blew my mind when I got this to work. It absolutely blew my mind. Can you like scale a 1 to 10 with 10 being mind is blown, one being I don't care. How much did this blow your mind? Big spender. I know. We spend so much money on um subscriptions, right? Crazy. Be careful. I know, right? I know. We spend way too much money on subscriptions. This is also like a previous statement. I pretty sure we spend even more money on this now. So, not good. I need to be canceling some subscriptions. Eight. Okay. Good. I am glad if somebody goes like, "Oh, one. " I'll be like, "What? What are you on, bro? " Good. I'm glad you guys like really see like wow the extent that this changes things. Six seven. Okay. 67. Okay. Ah yes. My rage baiter is coming at me. I don't even care right now. Nine. Especially doing it without code. Um which hardware do you use? Well, right now I'm using my my MacBook Air. I actually do use my MacBook Air for my financial statement documents thing. Uh what it does is just it monitors it. It just monitors my um statement. So every time there's a new files that get pulled into my machine, it runs this agent and it's able to keep updating it and it alerts me and things like that. So yeah, I'm just running on a MacBook Pro. You probably like want to host it like I probably will. Um we I do host like some other things. Uh but this one right now it's like completely on my machine. If you're hosting it's going to be like I don't know like $5 a month, something like that. It's really not expensive. Um, you can try out something like there's a lot of things that you can host it on. Like Hosting Jer is one example. I can't really think of any ones on top of my mind right now um off top of my head right now, but yeah, there's it's like the hosting space. It's there's so many options that you can host it and that's still technically private as well. Just want to make that clear. Just because you're hosting it doesn't mean that you're losing privacy because you're the one hosting it, right? Like you have complete control over the virtual machine. So it's that's totally fine um to do as well. How can you be sure your statements stay private? because everything is on my local machine. Like where could it possibly go? The model is on my local machine. This entire like thing agent that I built is on my local machine. As you can see, it's like literally localhost 5678, right? And then the files themselves are also on my local machine. So it's like there's literally nowhere possible that these files can go. So it is local. I am certain of that. Yeah. How do you do that? Okay. Um, if you check out my most recent video, which I'll give to you guys, I also give you the JSON for it. You can just do what I just said just now, like literally just do this. Go here, download it, right? And then you can just start building stuff. Um, I'll also just like give you guys I'll link the JSON as well. The JSON is specifically for this workflow. This workflow is not very difficult to build. It's very simple example, but you can build like much more complex workflows. Um, yeah. Like, but it's a good start. So, I'll link it for you guys as well. Give me a second. Okay, here you go. Yeah, just download that and then you can upload it into your um NAT your locally hosted NAT and then you can do this crazy, right? Amazing. All right, that is a big cursor. I know, right? So you guys can see what I'm pointing at. All right, I guess I want to show you the code example as well. So there's a code version of this. Um what I did to build this, I used a let me let me share my screen. Okay, first of all, we good with um if we're good with the NA10 version and no code version. Let me show you the code version now. Let's see. Ah, yes. Let me go and open this up. Okay, great. All right, I think you guys should be able to see that. I guess I should show my face so I am real. Yay, there's me again. Okay. Um, so this is warp. Um, this in itself like it's just like a terminal manager. It's also it's for like AI coding. Um, because I was I just made everything using warp directly. You can also use like other stuff um if you want. Uh you can use like cloud code or whatever or you can just like not use a AI coding agent too. You can just build it yourself. Um but the way that this one works is you just run it. So what you do is I would just literally type like run this agent uh or just have it running continuously. It's usually running continuously, but because I'm doing this live stream, that's why I killed it just now. So I just write like run this agent and then it would actually run it and it would go through my emails um it and then it would decide whether things are supposed to be ignored. It can flag it can ignore it, flag it as spam or it can decide that something should be responded to and it actually create a draft for you um in your drafts folder so you can decide if you want to send that email or not or another one is ignored. The other option is that if it feels like it's something that's not um that is like dangerous where it needs human review, it would also flag it too. So, sorry I can't show you this like live. I think it's so much cooler when it's actually live. Um but let me just show you what it looks like. Um the way that the agent works is that a new email will arrive at an email here. This just happens to be my email. You can do whatever email you want. You have one agent that a try out agent that analyzes the email and in terms of importance or urgency and it decides that um to respond to ignore it or flag it for human. It then passes on to a draft agent. If we decide that we're going to respond to it, it would generate a reply draft and then pass it to agent number three uh which is a quality checker that reviews the draft for tone and accuracy. And if it needs to flag something for human review, it would send an email with a title flagged email. So, let me show you my actual email to see what that looks like. Cool. So, here is my email here. Oh my gosh, what do I have on my emails? Okay. Um, the example I showed you guys were these three emails that if it went through. So, this one it just decided like, oh, you know, oh wow, like severe forecast, not that interesting. Like, you can just ignore it. So, it didn't do anything. Here's an example. So I sent this email from uh here's an example in which it would draft something so that I can reply to it right say hi Tina I'm pleased to inform you that you've passed the initial interview stage for a software engineering position blah blah so please reply by Wednesday end of day with your availability and it would actually drafted this email for me um so it's like dear Octopus team thank you for the invitation I'm delighted to proceed to the final round appreciate opportunity I'm available next from Thursday to Friday and can adjust my schedule to accommodate your preferred time looking forward to your response right so obviously I might need to change this depending on what actually works best but it would actually draft this email for me. Um, and I would know that it's something that I need to send out. And I'll give you another example of something that gets flagged because nothing got flagged recently. So, this is an example where it's like basically um from a career notice and then I get an email that something gets flagged um because it the AI here decides that oh like this is this could be a sus suspicious job posting with HTML code and unclear sender legitimacy. So flag for human review to verify authenticity and potential fishing risk. So also crazy like the impact of this is massive. It's screening through all of my emails right now, right? and it's being able to like flag things. It's able to create drafts for it. I I'm not going to let it like send it. I don't let it send it directly because I still want to like review everything that I do send. But it's able to do this. And for personally from personal use, this is already like really great, right? But also like for those of you who are using who get a lot of emails like for my business for example we get a lot of emails that are coming in. It's really difficult to actually monitor all of these emails and make sure that we're replying to everybody in a timely fashion. So this is a way for us to be able to um manage that like all the questions that we're getting. All of the collaboration requests um all the people that are asking questions about videos asking questions about like um the boot camp the workshops everything that we're doing were able to go through the system. And again, this is completely private because this whole thing is running locally, right? Like I'm not going to be sending my email, my personal emails or like my corporate emails or anything like that to I don't want to send that to like close source AI and have it like be used as training data. Like that is not what I would want to do. But these emails I'm just directly getting it through the Gmail API. Um, and it's processing all of this information locally. So I know that this data is not going anywhere. It's still in my Gmail. All it did was get processed. Um, and when it drafts these emails as well, like nothing is leaving this closed loop system. So this the implication of this is also really huge. It's like we're able to monitor so much of this now because of a agent like this and this agent like yeah like as you can see um how it works. This is using the OpenAI agents SDK and you can just build it with this. Yes. And this was built using Python. Does anybody have any questions? I want to see if anybody has any questions questions. Oh, in the meantime, I'm also and let me just show you guys like what the code actually looks like, you know, while I'm at it. Um, okay. So, here's the code for it. That's the read me. Um I have this on GitHub as well. So uh you can check it out if you want to just directly use it too. That works too. So yeah, it has like the agents that are here. It's configured using the agents SDK. Uh different tools. It has credentials which I'm not going to open right now. Uh the agent workflow. So here's the code for the agent workflow like defining how these are being created. And then here's like a test sample here as well. Um yeah. And this is the email agent workflow Python agent. Any questions that people have? Let's see. Oh my god. Obama emailed you. I know, right? Wow. Did you make this? I'm so I'm Yes. Yes, I did make this. And by the way, if you're if for our agents boot camp as well, like these are all things that you would learn to make within a 28 days. you would be making a minimum of four agents that are going to be a lot more complex than this because they would actually, you know, it would be processing everything. Um, and yeah, A lot of this like I took out a lot of the complexity here cuz I wanted to show you guys like demo version of how these things work. Also, I don't want to leak my financial statements to you guys online. No offense. I don't know if I trust you guys with my financial statements yet, but yeah, these are like all agents that you can be building. Um, and it's just like I can't even express like Okay, one to 10 again from the code version. Um, how mind-b blown are you? I'm like 10 out of 10 mind-b blown. Isn't Warp a paid service? Yeah, it is a paid service. You don't need to use that though. Like I'm just using Warp because I find it to be pretty easy to build like agents um using Warp because I just give it documentation. It's able to build it for me. You can use stuff like Cloud Code. Uh you use cursor, you can just like vanilla code it yourself if you want as well. Like yeah, do whatever it is that you would like to do. But the actual agent itself is completely free. Crazy, right? No matter how many times you run it, it's completely free. Okay. Can you create a multi- aent system where different agents run on your different machines like one on a Mac, another on PC, third on Raspberry Pi? Yes, absolutely you can absolutely do that. Yes. Um, can you build agents of beginner and non techchnical course? Yes, I just showed you the end to end version. So, yes. Okay, let me go on because I know there's still like other things I want to cover right now. But yeah, I hope you guys um are like, "Wow. " Because I was definitely like, "Wow, that's crazy. I can't believe that this is possible right now. Blows my mind. Okay, anyways, blah blah. Enough yapping. Okay, let's talk about open claw. Um, okay. I don't have too much time. So, um I'm time right now, but I'm going to send you guys like a little walk through um as part of the slides. So, about like how to install open call if you would like um to sign up for the so when you get the slides, we'll also send that over. So, I don't really have that much time to go over it right now. Sorry. But I do want to mention it. So, OpenClaw, the viral AI agent, right? I just want to say like exactly things like this when somebody goes like, "Oh my gosh, OpenClaw, right? That's so cool. " Um, what is OpenClaw? OpenClaw is an AI agent. Like these are this is the type of thing that you're able to build. Open Claw is able to be built because of the new open source developments that are happening right now because of the infrastructure that's available right now. Like you can build agents like OpenClaw. Like yeah, you got to be like a lot more technical if you want to build something as good as OpenClaw. But the whole point is like Golden Claw isn't like some new technology, some like magical type of like software that came out. If you understand that basically everything is an agent these days, it's just another type of AI agent. It's a [clears throat] self-hosted autonomous AI agent that runs 24/7 on your hardware. It doesn't answer question, also does things as well. What makes it special is that it has a lot of integrations in it. Like it's just able to integrate with a lot of different things. Um, it has really good memory and it's also running locally. So, similar to what I just showed you guys on the code side, also for like um yeah, like similar what I showed you guys with NA10, also with like the agents SDK agent that we that I just showed you. It's the same thing, right? It's just running locally. It's an agent um and it has persistent memory remember stuff over time and is able to proactively automate stuff. You can think about it as just like a really intense version of the agents that we just saw. So, it can do stuff like read and write uh files on your computer, manage calendars and send emails, browse the web autonomously. You control it via Telegram or WhatsApp and extends it's extensible with agent skills plug-in. So the way it works is that you just need to install and configure it. Set up your server with LM API key like whatever model that you want connected to Telegram or WhatsApp and then just start asking it to do stuff. Really simple um to do that. Yeah. So I think like be careful with OpenClaw because it does have a lot of privacy issues right now. Um, and especially if you're not a developer, you might not really be knowing what it is that you're doing, but I think you can definitely try it out. Um, I think it's a good thing to try out and then just experiment a little bit with. It's also going to be running completely on your computer um, locally here. So, if you just go to opencloud. ai and just literally follow the instructions of how to do this um, super simple to to run and start working with this as well. Yeah. So, do recommend that you try it out. Anybody have any comments about what they think? Like, this is the power of open source, guys. Crazy, right? Let's see if anybody has any questions. Does that help clarify like what I mean when I say like it's not some type of like super new innovative thing that everybody's freaking out about? I mean, everybody is freaking out about it, but I hope like this kind of helps you understand that it's just another agent. It's just a really good one and it's like using the power of open source to do it. Um, but it's not anything like special or new. Can we build agent without open cloud? Yeah, I just showed you guys two agents you can build with open cloud. OpenClaw is not for building agents. OpenClaw is an agent. Like it is an agent that you're using to manage a lot of your personal things, but you can build your own agents like the things I just showed you previously. Lots of videos saying Open CL. No, it is not safe. It is definitely not safe. No, no. That's why I'm saying you can play around with it if you want. I don't really recommend like giving it access to all your things. And if you don't really know what you're doing, I do not recommend just letting it yolo and do whatever it wants. That's probably not a great idea, but I do think you can try it out just to experience how amazing it is to have autonomous AI agents. Also, I'm just saying like these are stuff that you can totally build yourself. Um, you can build this type of thing. And this is like open source, right? But you can also build it for like enterprises for different companies. A lot of the freelance work that we do like contract work we do is that we build these type of AI agents for other companies like companies that are in healthcare companies in the finances tech companies startups as well like there's a lot of need for people to develop AI agents and you can do all of this it's just like a matter of skill right so yeah I do want to talk about the fact that yes there are a lot of serious security concerns with openclaw um because you need to give it a lot of permissions like email terminal file system etc Uh yeah, so it's like broad system access is not great. It allows people to attack and get over your full control over your computer. Um malicious agent skillware. There is the drawback of open source is that you do need to be more defensive because it is accessible to everyone which also means that you get like people doing not very nice things on it as well. So you need to be careful about that. Know what it is that you're doing and also it does have persistent memory which is data retention. So it does store your conversational history and learn behaviors locally. So this is not an issue in itself, right? Because it's stored locally. But if your machine does get compromised and attackers gain access to it, they're getting like a lot of data that you could potentially have. If you don't know what you're doing and you're like, "Ah, yes, let me just give access to my data and let open claw like give access to my data. " Then you're going to have problems, right? So if you are going to like try it out yourself, I recommend running isolated Docker container. If you are familiar with Docker containers, do it with that. Audit all installed agent skills, limit access sensitive data, and use dedicated non-admin accounts, and don't give it access to things that you would not like to be leaked is how we would put it. Yeah. So, here's just some examples. I'm not going to go into this right now, but there's just like some examples of what has already happened with people who are being not very nice and um stealing data through like open claw and other risk as well. So with this like open source development, it is definitely true that you there needs to be a lot more on the security side as well. So I mean I've been harping about this for like months at this point, but really like especially with open source coming out guys like if you're like oh what should I do? working in the AI field right now like where is the demand right like the clear demand is number one building agents um either personal agents develop like building your own company with agents or like building agents for other people's companies there's like so much development here like there's so little people who have that skill like people who go through our boot camp within 28 days they're like I'm not saying everybody does right like caveat here but there's people who like build their agents and sell their agents and then tweak these agents and sell them to like different companies that they're working at because there's just like so much demand for this and not that many people who actually know how to do it. And if you have like specifically industry knowledge like if you're someone who works in healthcare, works in finance, then you understand what needs to be built. So once you understand how to build agents, you are in a really good position to actually build these agents. Um yeah, there's so much opportunity. That's like number one opportunity in this field. Number two opportunity is focusing on the actual [clears throat] security aspects of it because with these companies that are building these agents with people building their own agents obviously the things that h start happening is security risk right so focusing on the security part is also solving massive problems for people so yeah if you don't know what it is that you should be doing and learning about those two things make sense cool anybody have any other questions comments, thoughts, just buy a second device only for Cloudbot. I probably would, but I'm just like, does Clawbot deserve its own device? What level AI literacy does one need to learn how to build agents who are using OpenClaw? I mean, OpenClaw, you don't need to have literacy. You just download it directly, do it. to build agents. I would say um learning the fundamentals of it like what level of literacy. You don't need to know how to code per se, but you do need to understand the structure of how agents are, or else it'd be very hard for you to build agents just like randomly. What we've noticed is like if you don't understand the fundamentals of what agents are and how they work, you can build agentic workflows, but you end up building workflows that are not actually that good and usable when it comes to like application level because you're not able to do stuff like evaluations for example. Like we don't understand the concept of how to evaluate and how to test agents. You can like definitely still build agents. They're just not going to be ones that would work very well in production and they're also not ones that you can like build for other people, right? So, you do need that level of literacy. Why not containerize it? Indeed, you can containerize it. That's why you've used Docker, then you'll be fine. Uh, what are your thoughts on the Gemma models compared to Quen, especially around A to 12b parameters? I personally think Quen is a lot better. My voice is echoing through space. Oh no, that is unfortunate. Happy Valentine's Day everyone with AI agents. Happy Valentine's Day indeed. You guys have any Valentine's plans? I do not have any Valentine's plans sadly. Hi Tina. Is your boot camp approved for beginners to someone from non-technical background? Yep. Uh you do not need to have any technical background to join. All right. Uh let's see. Let's see. I think that is on the security side. Yeah. So just want to finish this off. Companies are going all in. And so actionable automations, major companies are racing to integrate AI agents to their products and workflows. This is really obvious that's happening right now. So yeah, pop quiz if you guys are paying attention. What did I say earlier? What are the two things in the AI space that is really worth learning right now? Write it in the write into the chat. What did I say? Um yes, that's why you need to learn the fundamentals. really recommend you not just build things and use things randomly. Learn about the fundamentals first and everything else will start making a lot more sense. Like for example, Halam's work, agent architecture, security practices and testing and evaluations. Testing and evaluations, evals and monitoring very important. These two are very important. Um and you don't realize that these two are important until you end up in trouble and you're like, "Oh man, I wish that I learned about security practices and testing any valves. " And you're like, "Oh, okay. " But yes, these are very important. Yeah. Oh, and for those of you who want to um sign up for our agents boot camp, thank you so much for being patient with us. We have I know that it's been a few months since we launched, but we are going to be launching opening up enrollment again on February 17th. Um we'll be covering like the way that we do things is we want to give you a very clear curriculum and it's also a live session. So live mentorship to navigate all of this. You'll be learning fundamentals, understanding privacy, security, building production ready agents with the models and frameworks that are there. So we focus on having a clear curriculum um mentorship fundamentals and also on the application level. So throughout these 28 days you will build a minimum of four different agents and you will learn everything that you need to know in order to build functional agents that you can uh use yourself, you can sell them, you can contract work. people have done a lot of different things with the agents that they've built afterwards as well. So, um if you want if you're interested and yes, we do cover like stuff like open source um and things like that as well. So, if you want to learn more about this, if you're interested, please go ahead and go to lonely octopus. com to sign up for the weight list because we usually sell out within the from the weight list directly um within a couple hours. We only limit it to a 100 people. The reason for this is because uh we actually want to give people like personal attention to all of this. That's why we don't extend our cohorts past 100 people because we want people attention um as we're building agents. Also, a little reminder for those of you who have joined any of our workshops previously. Um so I think some of you guys have joined workshops that we've done. You will have a coupon code. So, that coupon code is valid from the AI agent breakthrough workshop. So, $100 um for the next cohort. So, please make sure if you are interested when you're signing up for the February 17th cohort launch to put in your coupon code so you can get $100 off. All right. And any questions you may have is contact at lonely octopus. com if you wish. So, yeah, here are some of the takeaways. Any comments, questions? Oops. Where's the comments? How much is How much for the boot camp? It's $997. $887 if you have a coupon for 28 days. Thank you, Rex. Are we building an alt operating system? Okay. I'm not really sure what you're referring to here. Agents of cyber security. Okay, good. Agents of security. Yes, you guys were paying attention throughout this entire process. I am very glad. No, you rage baiters. No, I did not say that you guys should learn anatomy. What? No. Yes. So it's like two things that you guys should learn is agents, how to build agents and how to deal with security. Yeah, very important these two things. Tina, we have to pay to contact you. No, you don't need me. You can just send us an email. No, you don't need to pay to contact me. Worth every penny. Really appreciate it. Um, how much time invested in the basic course daily? So the boot camp we recommend having six hours a week minimum. Um I think that's the minimum amount of time you do need in order to dedicate to that. It's two hours of live every week but you are building agents and there's an entire community around it as well. So as you're building these agents it is important to have that amount of time to building it. So if you don't have six hours a week don't actually like sign up. I really don't recommend it because I want you guys to actually get you know get the most of what it is that you are paying for. So, if you don't have the time, I'll rather you just not do it. I do have like an entire video about it as well. Uh, so this is a video that goes through everything about the agents boot camp. So, you can join the weight list if you wish. Um, and we will be launching on the 17th to the weight list first. So, you'll be the first to hear about it. Um, for those of you who want to go a little bit slower, like and you're you want and obviously if you this is like too much money, it's not something that you can afford and you're okay like kind of figuring things out yourself and going a little bit slower. What I would recommend doing is going when I say like learn about the fundamentals, I would really like you guys to learn about these like the six components of AI agents that I spoke about earlier, right? like learn about how each of these things fit into each other and really focus on the evaluation part and deployment like evaluation and orchestration including deployment that is really the part that is missing um from most people's agents that they're building. The reason why their agents end up not being super useful is because of that part. So um building it is definitely really important and figuring out how to do that. But more it's like the other 50% is the evaluation, testing and orchestration. So, please do make sure you focus on those things as well. I do not want you guys building agents that will end up, you know, ruining things because you didn't test it properly. Yeah. And I also don't think Yeah. For those of you who don't know how to code, it's honestly okay. Like, you can build agents without knowing how to code. If you do know how to code, um, does it mean that you have more access to stuff? Absolutely. Um, especially if you want to build like more complicated agents. Truthfully, I think knowing how to code really does help. And um yeah, it really does help. But if you want to just get started and you want to just start building agents, you really don't need to know how to code when you're doing it. Like we have people from our boot camp, they would be they like build these agents using NAT and they're selling these agents using NATE for like 10K USD, right? Like we've had someone recently in the past cohort like sell their thing that they built during the boot camp for 10K. And I was like, "Wow, that's pretty crazy. " Um but it made sense because he was very specialized in his field. So he was selling it for an agent that does like a very specific type of thing and it didn't need you to know how to code in order to implement that. But of course if you want to build your own software and you actually want to like build it out yourself um build a business surrounding it then I think learning how to code is still important to do so. Yes. Thank you so much Penny Automatica. Thank you so much for saying that. It has been it was very useful too. That honestly makes us really happy. Are there also intermediate advanced boot camps? My background is DevOps and cloud computing. Yeah. So the way that we structured the agents boot camp, we have a code track and a no code track. So we because we focus on fundamentals, right? So we teach you about how agents work fundamentally speaking and the implementation of that. You can do that through code and no code as well. So if you do know how to code and you already have a technical background, then I highly recommend that you go for the code track. Um, so when you're building out these things, like there's a lot more that you have access to. Yeah. And for those of you who don't know, just go with the no code track. So you don't need to know how to code in order to do that. So that's how we accommodate people. Oh, you actually started learning how to code, Penny. I so happy to hear that. Amazing. Okay. Amazing. I'm really glad. So uh, Python, I assume, right? Have you used the um, from the boot camp, have you started building using agents SDK now with Python? Will the boot camp be recorded to access later? So people who join the boot camp have um lifetime access to everything that's there including the projects and starter code and like things like that. Yes. What platform do you use to create your presentation? Ah that is a great question. Excellent question. Uh this is actually with the slides right that we have. Um the presentation is an internal agent. Like it's an internal um software that we built to create these slides. Yeah. One of the things that you can start building. What are you doing after stream? I think I'm going to go eat. Yeah. Speaking of which, I'm probably going to stop streaming now. Thank you all so much for joining. I really hope that this was helpful for you. Um yeah, thank you so much. And I really hope this was helpful and I will see you guys in the next video or live stream or hopefully in the next workshop or cohort. See you guys. Have a good rest of your day and happy Valentine's Day. Please enjoy your day. I shall log off now. See you. Bye.
