스크립트
00:00:00You need an agentic OS, but it's not for the reasons you think.
00:00:04It's not because you need fancy dashboards like this or visual interfaces like this one,
00:00:08even though I'll show you how to create them.
00:00:10It's instead about building a custom system that takes the best of all the AI tools at
00:00:16your disposal, whether that's GPT-6 Astra or Cloud Opus 5.5 or GEV, and builds it around
00:00:24you and your work in a way you just can't get inside the terminal.
00:00:27That is what this agentic OS does, and today I'm going to show you not only how it works,
00:00:32but how you can build something like this for yourself.
00:00:35And along the way, I'm going to show you how to create a fully local voice system, how to
00:00:39integrate GEV as a virtually free, virtually instant model router, how to create a customized
00:00:44skill backbone that is built around how you actually work in your day-to-day.
00:00:49We'll also talk about how to use Obsidian as a memory layer that actually provides value.
00:00:53And at the end, I'll even show you how you can package something like this, not just for
00:00:58you, but for teammates and even sell to clients.
00:01:00This is going to be all about showing you how to create an agentic OS that provides real value
00:01:05beyond the visual spectacle.
00:01:07So let's get started.
00:01:09So let's begin by talking about what an agentic OS, what a Cloud OS even is.
00:01:13So it is a three-layer system, and the first layer is the visual layer, and that's what
00:01:18we're looking at here.
00:01:19For me, this has a lot of stuff to do with my social media, what's going on on my different
00:01:23accounts.
00:01:24I have it hooked up to my calendar.
00:01:26I can look at all my different skills and automations and run them from here.
00:01:29And speaking of running it, I can run it either on Cloud Code or Codex.
00:01:33Now, this is all happening inside of my browser.
00:01:36This is a web app, but there are no rules here.
00:01:38This doesn't have to be a web app.
00:01:39In fact, as you see here, I've ported it over to Obsidian.
00:01:43This is also an Obsidian plugin that, just like you saw, can run on Cloud Code or Codex.
00:01:48But I get all the advantages of being inside the Obsidian desktop app.
00:01:53I still have access to a terminal here.
00:01:55I can click on any of these buttons, which are linked to different skills and automations.
00:02:00I have different tabs that dive deeper into what's going on with my audience, what's going
00:02:04on in AI in general, right?
00:02:06I have a whole research tab that takes a look at the GitHub trending.
00:02:09Outliers on YouTube, as well as things like Hacker News.
00:02:12The big thing with the visual side is this is ultimately 100% custom, and you can edit
00:02:17this so it makes sense for you and your work, or your team's work, or your client's work.
00:02:21Now, the second layer is the memory system.
00:02:24And for that, I use Obsidian.
00:02:26And the value here is not these knowledge graphs.
00:02:28Like, these knowledge graphs look cool, but that's about where the functionality ends,
00:02:31because this is not a RAG system.
00:02:34What the memory layer is really about is creating a file and folder structure that makes sense
00:02:39for you, the human being, and for your AI OS, so that you can get answers quickly and accurately,
00:02:45even when we're talking about thousands and thousands and thousands of files.
00:02:48And this is going to be based on the Karpathy Obsidian RAG system, and we will dive into that
00:02:53in much more detail later.
00:02:54And last but not least, for the third layer, we have the skill architecture.
00:02:58This is the backbone of everything we do.
00:03:00If you get nothing else from this video, if you ignore the memory layer stuff, you ignore
00:03:04the visual layer stuff, I want you to listen to this section when we dive into it in more
00:03:09detail.
00:03:09And that is, how do we go from somebody who just uses AI to sort of help us in our day-to-day
00:03:14task, and offload virtually all of it to AI in a way that makes sense?
00:03:19Because you have things you do every single day, week to week, in your life, in your business,
00:03:23that we should not only be turning into skills, but then turning into automations.
00:03:28And from those automations and from those skills, we get certain deliverables and outputs.
00:03:32Those deliverables and outputs are what is fed back into the AIOS.
00:03:37All the reports you see here, all these graphics, all of these are driven by the different skills
00:03:41we run.
00:03:43So not only understanding that in a theoretical sense, but having the practical know-how of
00:03:47how to create something like this from scratch is what we're going to talk about.
00:03:50But before we do that, a quick word from today's sponsor, me.
00:03:54So inside of Chase AI+, you not only can get my exact Cloud OS setup, but you also get
00:03:59access to my Cloud Code and Codex masterclass.
00:04:03And these are the best and easiest way to go from zero to AI dev, no matter your technical
00:04:08background.
00:04:08We focus on real use cases.
00:04:10It gets updated every single week.
00:04:12So if you want to get more serious about this stuff, definitely check us out.
00:04:17There is a link to it down in the pinned comment.
00:04:19So the Cloud OS has three layers, visual, memory, and skills.
00:04:22But how does this thing actually work?
00:04:24How do all these pieces come together?
00:04:27Well, let's go through the architecture.
00:04:29So let's do some examples.
00:04:31So let's say I tell my Cloud OS, which also works with voice, go ahead and bring up the
00:04:36morning Intel brief.
00:04:40It instantly brings it up.
00:04:42But if I say something like, what was the biggest news in AI today?
00:04:50From today's morning brief, a model that can't write a sentence is beating models that can.
00:04:56We get a response like that and I cut it off.
00:05:04But now if I ask it for a more complex task, something like, can you create me some sort
00:05:09of report or explainer that kind of illustrates the difference between JEV, that new AI system
00:05:16everyone's talking about, and your standard large language models like Fable or Astra?
00:05:23Okay, I'll get to work.
00:05:26It'll do something like that.
00:05:27You will see with this more complex task, it actually pulled up the terminal.
00:05:33So what was actually going on under the hood there?
00:05:35And why did I show three different seemingly random examples?
00:05:37Well, it sort of illustrates how this AI OS routes your different requests to different
00:05:42models, depending on the nature of your task.
00:05:45And this is where stuff like JEV kind of comes into play, which we see right here.
00:05:50So let's focus on that third example where I asked our Cloud OS, hey, come with like
00:05:53this visual explainer sort of illustrating the difference between JEV and stuff like Fable.
00:05:58Well, we started over here.
00:05:59I simply gave it an audio request.
00:06:02Now, this then went to our Obsidian plugin, but this just as easily could have gone to our
00:06:07Jarvis HUD.
00:06:08Everything you just saw in that Obsidian plugin also works exactly the same here.
00:06:11This also is able to pull up a terminal.
00:06:13That request then went to the bridge.
00:06:16And the bridge does two things.
00:06:18One, it takes our request and it records all this inside of Obsidian.
00:06:23Two, it then starts engaging our local voice system.
00:06:28So all that speech you heard, all the transcription, all the actual audio is done locally on your
00:06:33machine.
00:06:34So we use an open source program called Whisper, which essentially acts as the ears of our model.
00:06:40So when I said, hey, give me that explainer, you know, JEV versus the LLMs, it took the audio
00:06:45and transcribed that into text using Whisper.
00:06:47Now, the audio you heard back with that sort of like Jarvis voice, that also was a local model.
00:06:53That was Kokoro, totally open source.
00:06:56You could change that voice to be whatever you want.
00:06:58That transcription then got sent to JEV.
00:07:03And we'll do a little JEV explainer here in a second.
00:07:05JEV is amazing when it acts as a classifier.
00:07:09And what we want to do with JEV is we want it to sort our requests.
00:07:13And there's really three tiers of requests we can handle.
00:07:16Now, tier number one are for extremely simple requests that we want pretty much instant responses
00:07:20to.
00:07:20So when I asked it to bring up that morning report, that morning Intel brief, this was a
00:07:26tier one request that JEV classified.
00:07:28AI isn't even brought into the system if it's tier one, which is why it's able to be so fast.
00:07:33We're talking about hundreds of a millisecond to bring up these things.
00:07:35And this could be something like you saw where I say, hey, bring up a report.
00:07:38Or these can even be commands inside of Obsidian itself where I say, hey, bring up the terminal
00:07:42or something to that effect.
00:07:44Then we have tier number two.
00:07:45These are for non-complex tasks, but they aren't so simple that we put it into tier one.
00:07:51So when I asked it something like, hey, what was the biggest news in AI today?
00:07:55I want some thinking involved.
00:07:57So JEV then calls on the smallest models available to us.
00:08:01This is Haiku or Luna, depending on if I'm using the Astra version or the Cloud Code version.
00:08:07This is cheap.
00:08:08This is fast.
00:08:09And if you really wanted to go into a full local setup, you could swap these out for something
00:08:12that's just running on your machine.
00:08:13And then third, we have more complex real work, which is what you saw when I asked it to
00:08:18create that explainer.
00:08:20This is actually going to then pull up Cloud Code or Codex inside of your terminal, like
00:08:25you just saw.
00:08:26This gives you the ability, if you're inside, whether it's the Obsidian plugin or the web
00:08:31app, to actually interact with the terminal in those spaces if you want to.
00:08:35So the idea is your tasks require different models, and JEV is really, really good at classifying
00:08:40them appropriately.
00:08:41From there, all the deliverables, all the conversations, everything you do with your AIOS is then logged
00:08:46back in Obsidian.
00:08:47Now, a real quick explanation for why we have JEV in there.
00:08:49If you're not familiar with how JEV works, because it's still new to a lot of people.
00:08:53And unfortunately, a lot of sort of the discourse around it is kind of confusing because it's
00:08:58an AI system, but it's not an LLM.
00:09:00But people are saying it's a replacement in some cases for Cloud Code or Codex, which is
00:09:04sort of true in certain situations, but in general is not true whatsoever.
00:09:07So this is like the 90-second explanation for JEV and why it makes sense in our AIOS.
00:09:13So the big thing with JEV is it allows you to ask this AI model fuzzy questions and get
00:09:19specific numbers in response, get probabilities in response.
00:09:23So here's an example.
00:09:25Imagine you are using this in some sort of customer service situation and you have a customer that
00:09:30says, "I canceled my subscription last month, but you charged me again."
00:09:33And you want to know which team should handle this.
00:09:35So normally, if we sent this to large language models and they can handle this just fine,
00:09:40it would say something like, "Hey, this is a billing issue route to stick to the billing."
00:09:44And that would be the correct answer.
00:09:45JEV also does something similar, but it doesn't give us a sentence in response.
00:09:50There's no back and forth chat with JEV.
00:09:52It instead gives us the probability that there's a 92% chance this should go to billing, which
00:09:58is also the correct answer.
00:09:59So if we both get the correct answers, why the heck should we care?
00:10:03Well, we should care because JEV does this extremely quickly and way cheaper.
00:10:08We're talking like 200X, like faster and cheaper.
00:10:12That's why we love JEV.
00:10:13Cheap, fast, perfect for these situations where you're handling what should I do in this scenario
00:10:19where there are predefined choices, like in our model routing scenario, right?
00:10:26What are the predefined choices?
00:10:27One, two, three.
00:10:29Could I put Haiku here?
00:10:30Could I put Luna here?
00:10:31Sure.
00:10:32But there's no point in doing that anymore with JEV, which is why it's been integrated.
00:10:37So that is how all of this works from an architecture point of view.
00:10:42Now, let's dive a little bit deeper into each of those three layers, the visual, the memory,
00:10:46and the skills, so you understand how you could create something like this on your own.
00:10:50So let's go from the ground up beginning with the skill backbone.
00:10:53This is the most important thing of the entire video.
00:10:56And if you do nothing else, but what I'm about to show you here, you're going to be ahead
00:11:00of like 99% of the people who use Cloud Code or Codex.
00:11:03Now, the idea is simple.
00:11:05We are going to take all the tasks you do in your day-to-day, in your week-to-week, whether
00:11:09it's in your personal life or your business, and turn them into skills.
00:11:12And if it makes sense, turn those skills into automations.
00:11:15The value of doing that seems obvious, yet most people don't do that at all.
00:11:19They do it with a few of the things they do in their day-to-day, and they certainly don't
00:11:23systematically go through everything they do and break it down in this manner.
00:11:28Because while this looks kind of crazy, you probably have something similar if it was actually written
00:11:32out.
00:11:33You do things related to productivity, research.
00:11:35Maybe it's not content in a community, but you certainly have stuff related to clients
00:11:39and sales and finance in some manner in your business life.
00:11:43So, why haven't we codified this into something Cloud can help us with?
00:11:48And the answer is, it just seems like too much damn work.
00:11:50It seems like it would take too much time and you don't even know where to begin.
00:11:53Well, luckily, creating something like this, breaking them all down into skills is actually
00:11:57relatively easy, and there's just two things you have to do.
00:12:00The first thing we're going to do is have your AI system of choice simply look at your
00:12:04logs.
00:12:05Cloud Code and Codex keep a written history of every single conversation you have with it
00:12:10over the last 30 plus days.
00:12:12So, instead of us guessing about how Cloud or Codex should help us in our day to day, why
00:12:17don't we just have them look at how we actually use them in reality and ask, "Hey, based on
00:12:23everything I've done over the last 30 days, are there any skills we could create?"
00:12:28That is literally the entirety of the prompt you have to give it.
00:12:31There's no special thing, there's no crazy sort of workflow.
00:12:34It's literally just, "Look at how I've used you for the last 30 days.
00:12:37Can we turn those into skills?"
00:12:39The second thing we are going to do is we are going to sort of just take that idea a
00:12:44step further, and instead of giving it a log, we are just going to sit in front of Cloud Code,
00:12:50sit in front of Codex, open up our microphone, and you are going to talk nonstop for like
00:12:5310 to 20 minutes.
00:12:54And all you're going to do is you are going to explain what you do in your day-to-day or
00:12:57your week-to-week.
00:12:58After you dump all that information into Cloud, it doesn't have to be coherent.
00:13:01Again, stream of consciousness.
00:13:03It can get through your inability to talk correctly.
00:13:06You are then going to ask it, "Hey, based on everything I just told you,
00:13:10what are some ways where we could create some skills or automations to take some work off my plate?
00:13:15How could we take all these tasks I've talked about and give them to you, this AI system, to execute in my stead?"
00:13:21Simply by doing that, and neither of these scenarios require you to have the answers beforehand.
00:13:28You don't have to know what you want to turn into skills.
00:13:30We'll get you pretty much 90% of the way here of creating these skills that will make your life infinitely better.
00:13:39Now, the question of which one should be turned into automation should be relatively straightforward.
00:13:43And when it comes to turning them into automations, yes, simply ask Cloud Coder Codex to do that.
00:13:47We can always do that manually through things like routines or scheduled tasks inside of Codex.
00:13:52But these systems also have the ability to turn those into automations that are triggered by your machine itself.
00:13:57So we have a lot of flexibility here.
00:13:58And again, like so many of these things, you don't even have to have the answer when you walk into the space.
00:14:03You just have to ask Cloud or Codex to do it.
00:14:06Now, how does this whole skill backbone tie into the AI OS at large?
00:14:10Well, like I talked about earlier, the idea is we've created all these systems with the skills and with the automations
00:14:16to create things for us.
00:14:17Many times they are going to be reports and they're going to be deliverables.
00:14:20Instead of having them spread in a million different places, why don't we just consolidate them into one place?
00:14:26This is where a lot of the value from the dashboard and the visual side is actually harnessed.
00:14:31For me, that's metrics.
00:14:32That's a look at my schedule.
00:14:33That's breaking it down into tasks.
00:14:35That's the morning headlines.
00:14:36That's audience metrics.
00:14:38That's deeper research into things that are relevant for me.
00:14:41Again, this has to be 100% customized because you're not going to care about your YouTube subs, right?
00:14:47But you undoubtedly have a lot of metrics and a lot of deliverables that you would like all in one place that are easy to access and ask questions about.
00:14:54And that sort of thing can be a little bit difficult to do even inside of the desktop applications and definitely kind of a pain inside of the terminal.
00:15:02Even if you want to argue about the value of having that all in one sort of visual dashboard that you can customize.
00:15:07I don't think there's any argument about creating some sort of skill backbone in your greater Claude or Codex ecosystem like this.
00:15:15Now let's talk about the second layer of our system, which is the Obsidian memory system.
00:15:20When we talked about the high level architecture, you could tell everything kind of touches this layer.
00:15:25Whatever goes in, whatever comes out gets recorded inside of our vault.
00:15:29Now if you've never used Obsidian before, it is a completely free, simple to use desktop application that gives you insights into your markdown files, essentially helps with organization.
00:15:40Now Obsidian and things like Claude Code and Codex are often used together to create what is called a second brain.
00:15:46And the second brain is the idea that there's all these files and folders that have to do with you or your business.
00:15:52We put them in these series of files that we have called the Obsidian Vault.
00:15:57We run Claude Code inside of those folders and it has access to all this information about us.
00:16:02Now this Obsidian second brain has kind of gotten a bad rap as of late because people have sort of overhyped what it is Obsidian does in this structure and like what the second brain even means.
00:16:11The reality is Obsidian isn't giving Claude Code or Codex any sort of like magic upgrade.
00:16:16Just because this AIOS system uses Obsidian, that doesn't mean it gets like a better memory jumper, remembers things better.
00:16:23That's not what Obsidian does.
00:16:25The Obsidian desktop app just makes it very easy to navigate those files and folders that we have designated the vault.
00:16:32And that's a subtle but important difference because it means the power of an Obsidian second brain, the power of bringing Obsidian in the memory layer is not in Obsidian itself.
00:16:41The power of the system depends 100% on simply how you organize your files and folders, right?
00:16:47If I just have one folder called the Vault and I put 10 million just random files in this one folder and it's completely unorganized, but it's Obsidian, does that help me at all?
00:16:56The answer is no, absolutely not.
00:16:58So what we have to do is we need to set up some sort of folder structure that makes sense.
00:17:02And that's what you see here.
00:17:04So this all comes from Karpathy.
00:17:06Karpathy is an AI researcher who works with Anthropic now, I believe.
00:17:10And he sort of coined this structure with Obsidian Cloud Code and this like Obsidian rag idea.
00:17:15And the idea is relatively simple.
00:17:17You have this folder that you have designated as the vault in your system where everything's going to live.
00:17:22It's just a folder on your computer.
00:17:24Underneath that folder, we have three subfolders.
00:17:26We have raw, we have wiki, and we have outputs.
00:17:30And the idea is raw data goes in the raw folder.
00:17:33The wiki subfolder is where we sort of organize that raw data and create, you know, like Wikipedia style articles about them.
00:17:41And then three is the outputs.
00:17:42You know, let's say I wanted to make a slide deck about something that was a Wikipedia article.
00:17:47So for example, let's say I wanted our AIOS to find information about AI agents, right?
00:17:52Do a bunch of research on AI agents.
00:17:54So it does that and it creates a bunch of information or downloads a bunch of information about AI agents,
00:17:59which goes in the raw folder.
00:18:01Then let's imagine we have this AI agents folder in the raw section and I want to turn it into some sort of organized data.
00:18:07Well, then it creates an AI agents subfolder down here with all the information about AI agents,
00:18:13including things about autonomous coding and tool use patterns.
00:18:17And then let's say I told our Cloud OS, create a PowerPoint slide deck about this.
00:18:24Well, then it makes the AI agents slide deck.
00:18:27Okay.
00:18:28And that would go into the outputs.
00:18:30Simple enough.
00:18:31That's sort of the idea.
00:18:32And the cell here is that because it's set up in this way where it's very organized and I know where everything goes.
00:18:38When I ask Cloud something about anything that's inside of my vault, it has a clear map, a clear path of where to find that information.
00:18:48So it finds it quickly and it finds it efficiently.
00:18:51You know, it doesn't have to go through 10 million folders and subfolders to get there.
00:18:54There's a clear path.
00:18:55And the other side of the equation is because it's organized in that way, it's easy for me, the human being to find it if I want to.
00:19:02So with that being said, how should you set up your folder system with your AI OS, your Cloud OS inside of a vault?
00:19:10Does it have to be like this where it's raw, where it's wiki, where it's outputs?
00:19:13The answer is absolutely not.
00:19:14You can kind of set it up any way you want to.
00:19:17It just needs to be organized in some manner that makes sense so that you, the human being, can navigate it and Claude or Codex can navigate it.
00:19:26Now, like so many of these things, you don't need to have the answer.
00:19:29It is as simple as telling your AI OS, hey, this is the folder I'm using as my vault.
00:19:34Help me come up with some sort of organizational structure and it will map it out and move files and folders as it sees fit.
00:19:40The only other thing I would add to this is in whatever folder you designate as the vault, I highly suggest, and whether it's a Claude.md or Agents.md, you tell it what the vault structure looks like.
00:19:51So you give it instructions of how to add new files and content to your system.
00:19:56So it never kind of gets out of whack or drifts.
00:19:58And that's really all there is to the memory layer.
00:20:00It just needs to be structured.
00:20:01It just needs to be organized.
00:20:03It just needs to make sense for you.
00:20:05And lastly, we have the visual piece, which we've kind of already hit on a lot here.
00:20:08Where the big sell for the visual piece is all these skills have deliverables.
00:20:13We have a bunch of metrics that you need to be in charge of, and we're going to put them all in one dashboard that we customize.
00:20:18A lot of the value of sort of these custom dashboard setups as well, though, is seen when we bring different team members and clients into the equation.
00:20:27Because there definitely is a subset of people watching this.
00:20:29We're like, hey, I could probably generate something like this myself, even if it's some like, you know, mashed up thing inside of the desktop app.
00:20:37Why would I need to do this?
00:20:38Well, if you're someone who's working with team members or you're someone who wants to bring this to clients, it's not as easy to bring something you've built inside the terminal specifically to them.
00:20:47On the other hand, I can create something like this that is custom made for anybody and set it up extremely easily.
00:20:53On top of that, I can even pour in a lot of my workflows, a lot of my skills, a lot of my automations.
00:21:00And if I put them inside of the dashboard, like you see here, where they are simply a button one person has to press.
00:21:05Well, the way this works is in the background.
00:21:08It's simply calling a headless version of Cloud Coder Codex to execute those tasks.
00:21:12So I could have someone sit in this chair right here in front of my dashboard, whether it's the Obsidian plugin or the web app version, and I can start having them just running all sorts of skills that do a lot of work.
00:21:22And they're pretty much harnessing 90, 95% of the power of Cloud Coder Codex without even knowing how to open it up.
00:21:29So on top of all the information consolidation, when we talk about selling this and bringing other members of our team into the fold, the visual dashboard can actually help them a ton, which can't be undersold, even if you're someone who's technical enough to kind of do this on your own.
00:21:44And a lot of that has to do with the custom nature of it.
00:21:47This is not a one size fits all thing, which is why it's so powerful.
00:21:51No matter what you need to be listed here, no matter what you need access to all in one place, you can build it that way.
00:21:57So that's the rundown on this Cloud OS.
00:22:00I hope I was able to give you a better understanding of how you could build something like this for yourself.
00:22:04If you want to get my exact setup, whether that's the web app version or the Obsidian plugin, you can find that inside of Chase AI+.
00:22:10I will put a link to that down in the description.
00:22:12But besides that, I'll see you around.
커뮤니티 글
아직 글이 없습니다. 이 영상에 대한 첫 번째 글을 작성해 보세요!
이 영상에 대해 글쓰기