0:03 I just spent $20,000 on OpenClaw and
0:06 plan on spending another $100,000 on it
0:08 by the end of the year. I truly believe
0:10 the return on investment on all these
0:12 Mac Studios I'm buying is going to be at
0:14 least 10x what I'm spending. You might
0:16 think I'm crazy, but if by the end of
0:18 this video I don't have you convinced
0:20 what I'm doing is actually genius, then
0:22 I think you're the crazy one. In this
0:24 video, I'll go over what I just spent
0:27 all that money on, what my open claw is
0:29 doing on all this new hardware I just
0:31 bought, why I believe this is the
0:34 future, and show you how you can start
0:36 doing the exact same thing I'm doing
0:38 without spending a dollar on any
0:40 hardware. You can use the exact Mac Mini
0:42 you bought or even the dusty HP laptop
0:44 from your closet. You are about to get a
0:48 peak into the future of AI agents. Let's
0:50 get into it. So, real quick, we're going
0:52 to cover a bunch in this video. Chapters
0:53 down below if you want to skip around,
0:54 but basically what we're going to be
0:56 doing is I'm going to tell you why I
0:59 spent all this money to run local AI
1:00 models, what all the advantages are,
1:02 what this is going to allow me to do.
1:04 I'm going to give you a demo of exactly
1:06 what I'm building, which I promise you,
1:08 you have never seen anything like this
1:10 in your life. There is nobody else on
1:12 planet Earth building what I'm building
1:14 with Open Claw at the moment. I'm going
1:16 to show you how this will change the
1:18 entire world and why I believe all
1:20 normal people will be running out to buy
1:21 thousands of dollars worth of computers
1:23 by the end of the year. And at the very
1:25 end of the video, we'll go through how
1:27 you can do the same thing I'm doing on a
1:29 much smaller budget. Even if you have a
1:32 $500 Mac Mini or a $10 HP laptop in your
1:34 closet, I'll show you which local models
1:37 you can run so you can start doing
1:39 something similar to what I'm doing. I'm
1:41 about to take you through a glimpse of
1:43 what I think the future of AI agents
1:46 are, what I think the future of OpenClaw
1:47 is, and what I just think the future of
1:49 AI is. You're going to learn a ton, and
1:52 your mind is going to be blown by what
1:54 these AI models are capable of. So,
1:55 here's where we're going to start. The
1:59 $20,000 I spent is on two Mac Studios.
2:03 These are Mac Studios with 512 GB of
2:06 unified memory each. Why 512 GB of
2:08 unified memory each? so that I can run
2:12 the largest, smartest local AI models in
2:13 the world. Right now, that is Kimmy K
2:17 2.5. Kimmy K 2.5 is 600 gigabytes, which
2:20 means I need to load 600 gigabytes into
2:22 local memory. With two Mac Studios at
2:25 512, that's a terabyte of AI models I
2:27 can load into local memory. If I were to
2:29 do the same thing with Nvidia GPUs, I'd
2:31 need to spend over $100,000 on Nvidia
2:33 GPUs, but I can do that with the unified
2:36 memory of Mac Studios. Now, why do I
2:39 want to run large local models on my
2:41 computers on these Mac Studios? Well,
2:44 here are the five reasons why. One, you
2:45 run these models completely for free.
2:48 I'm not paying for API costs anymore
2:51 when you use Claude Opus or Chad GBT or
2:54 any other AIS that are in the cloud. You
2:56 pay for every single token. These local
2:58 models, because they're running on my
3:01 Mac Studios, they are completely free to
3:03 run. And because they are completely
3:06 free to run, I can have them work 24/7
3:09 365. This is an absolutely massive
3:11 advantage. The number one objection I
3:13 get when I tell people I'm doing this
3:15 is, "Oh, local models, they're stupider
3:18 than cloud models. Kimmy K 2.5 is not as
3:19 smart as Opus. It's not as smart as Chad
3:22 GBT 5.3." And you know what? You're 100%
3:24 right. But here's the thing. The fact
3:28 that I can run these AI models 24/7
3:30 completely changes what AI is capable
3:32 of. Now, I'm going to show you examples
3:33 once we get through this. I'm going to
3:35 show you my actual setup I'm building
3:37 with these Mac Studios. But the fact
3:40 that these AI models can now run 24/7
3:44 365 means I can do things like have them
3:47 read Twitter and Reddit all day and all
3:49 night looking for challenges. Have them
3:52 coding all day and all night building
3:54 solutions to those challenges. Have them
3:56 shipping those apps all day and all
3:58 night promoting them on all the social
4:00 media. I can constantly be creating
4:03 value from midnight to midnight every
4:04 single day. They they don't have to eat,
4:06 they don't have to sleep, and they don't
4:07 cost me any money. It's literally like
4:10 having 24/7 employees working for you
4:12 that never complain, never demand more
4:14 money out of you. They just have that
4:15 upfront cost of buying the computer. And
4:17 it doesn't matter that they're dumber
4:19 than Opus. I can't do what I just
4:21 described and what I'm going to show you
4:23 later in this video with Opus. If I were
4:25 to do this with Claude Opus or Chad GBT,
4:28 I'd be spending at least $10,000 a month
4:30 on API costs. But because I can run this
4:33 locally, I can have these models going
4:35 all the time, finding challenges,
4:37 building solutions to challenges, and
4:39 creating value for me and my company.
4:40 There's many other reasons why you'd be
4:43 wanting to run local models. Privacy,
4:45 for instance, everything that happens
4:47 happens on the computer. Nothing goes to
4:49 the cloud. Nothing goes to chat GPT
4:51 servers. None of this can be read by Sam
4:53 Alman or the owners of the other AI
4:54 companies. Everything is completely
4:56 private. In fact, I can unplug my
4:58 computers right now if I wanted from the
5:00 internet and they'd still be going and
5:02 running AI models. There's no internet
5:04 required. It's all private, all local.
5:06 No one can see what I'm doing. Next,
5:08 it's educational. By running these local
5:10 models, I'm learning how AI works. I'm
5:12 learning a ton about what goes into
5:14 running an AI model. I am learning about
5:16 the most important technology in the
5:18 entire world. And that in itself is
5:21 worth a lot of money. And lastly, it's
5:24 just fun AF. It is just so much fun
5:26 looking down at your desk, seeing
5:27 computers running, and knowing there are
5:30 AI agents running, doing things for you
5:33 24/7. It is fun. And I don't care what
5:35 other people say. You're allowed to have
5:36 fun in this world. You're allowed to
5:38 spend money to have fun. You're allowed
5:39 to do things for the fun of it. Don't
5:41 let people tell you you're not allowed
5:42 to have fun anymore. Take it from me,
5:44 you're allowed to have fun. And so that
5:45 is another reason why I'm doing this.
5:48 So, what exactly am I building with all
5:50 these local models running 24/7? Well,
5:53 have a look at this. This is my company.
5:57 This is my oneperson 24/7 365 AI agent
5:58 company. And this might look strange,
6:00 but let me walk through exactly what's
6:02 happening here. What you see here are my
6:05 AI agents, and they are working as we
6:07 speak. They are reading X. They are
6:08 going through Reddit. They are searching
6:10 for problems to solve. They are reading
6:13 my tweets. They're watching my YouTube
6:14 videos. They're looking at the
6:16 performance of all my content. And as
6:18 you can see, sometimes the agents even
6:20 meet up together, go to a meeting table
6:22 and have discussions. Right now, there
6:24 is a standup happening for all my AI
6:27 agents where they are brainstorming new
6:29 features for Creator Buddy. I can track
6:30 the live activity. So, you can see the
6:32 live activity over the last hour. A
6:34 roundt started to brainstorm new
6:36 features. They're all coming up with new
6:39 ideas for features for Creator Buddy. At
6:41 some point they will be handed off to
6:43 Henry who is the manager, the strategic
6:46 manager of all the AI agents. If we take
6:48 a look at the org chart of this AI
6:50 digital company of all these local
6:52 models running and talking to each other
6:54 and planning at all times, you can see
6:56 I'm at the top as the CEO. Henry is the
6:59 chief of staff. He's running on Opus 45,
7:02 but all he's doing is getting ideas and
7:04 approving and disproving. He just has to
7:05 do a couple prompts a day where he
7:08 approves or disproves ideas. the local
7:10 models hand to him. As you can see,
7:11 right now he is in the standup
7:13 brainstorming new features for creator
7:15 buddy. Then I have my other agents
7:17 working for me in this organization. So
7:19 you can see I have a creative team which
7:21 all they are is a local model running on
7:23 my Mac studio running off of Flux 2 that
7:25 is able to generate images for me,
7:27 thumbnails for my YouTube images for my
7:29 Twitter account. Whatever I need, that's
7:31 all being done locally. I have my
7:33 research team which is Scout who's an
7:37 analyst that is being powered by GLM4.7
7:38 which is a huge local model that I'm
7:40 running on my Mac studios. This is
7:42 constantly reading Twitter, constantly
7:44 reading Reddit, finding challenges to
7:48 solve and handing them to Henry who is
7:50 my chief of staff and then I have many
7:52 other agents working in my digital
7:53 organization as well. I have an
7:55 engineering team who's constantly coding
7:58 for me and many other AI agents as well.
7:59 They are able to accomplish so many
8:01 things. So, for instance, right now as
8:02 they discuss new features for Creator
8:05 Buddy, they're looking online, seeing
8:07 what people want out of content tools
8:09 and building it out and discussing it
8:10 with each other and learning from each
8:12 other. And this goes much deeper than
8:15 that. Each one of these agents, have
8:17 their own memories, have their own
8:19 personalities, are building their own
8:21 relationships with each other agent. If
8:23 you look down here below, I have a list
8:24 of all my agents. If I click on one of
8:28 them, you can see here Quill has its own
8:30 soul. So its own personality, how they
8:32 think, their own signature phrases,
8:34 their own voice of how they talk, their
8:35 own speaking style, their own
8:37 responsibilities, even has their own
8:40 relationships with the other AI agents.
8:43 Every time one of my agents speaks to
8:46 another agent, it actually changes their
8:47 relationship. They can become better
8:49 friends. They can become bitter enemies.
8:51 It's just like a real workplace where
8:53 you have friends, you have co-workers
8:54 you like, you have co-workers you don't
8:56 like, and their relationships can shift
8:59 and evolve over time. They also have
9:01 their own memories. So Quill is a new
9:03 agent I just hired just now. So it
9:04 doesn't have its own insights and
9:06 strategies, but as they participate in
9:08 meetings, as they have conversations
9:10 with other agents, they can come up with
9:12 their own insights and strategies. So
9:14 for instance, Quill is my creative agent
9:17 who writes tweets. Maybe Quail goes in
9:20 and has a water cooler conversation with
9:22 Scout, who is my local model, who is
9:24 constantly reading Twitter, and Scout
9:25 tells him, "Hey, that tweet you wrote
9:27 the other day, it's performing really
9:29 well." Quill can then get a memory that
9:31 says, "Okay, tweets like this perform
9:32 really well. I need to write more about
9:35 them." So, my digital society here, my
9:37 digital office, they're constantly
9:39 learning from each other. They're
9:41 constantly talking to each other. They
9:42 can do many things. They can even have
9:44 water cooler conversations, which you
9:45 just saw there. They're now walking over
9:47 to the water cooler and talking to each
9:49 other. This happens all autonomously
9:53 24/7 365. They are constantly
9:55 researching, constantly writing,
9:57 creating, coding. They are constantly
9:58 learning from each other. They're
10:00 constantly building relationships with
10:02 each other. And I don't need to be a
10:04 part of this. I can just sit back and
10:06 enable them and make sure they're doing
10:07 good work. When I'm sleeping, they're
10:09 working. When I'm eating, they're
10:10 working. When I'm watching the Patriots
10:12 win the Super Bowl, which this will look
10:13 really bad if they don't end up winning
10:15 the Super Bowl today, they are talking
10:17 to each other and working and creating.
10:20 This is only possible with local models.
10:23 If this were all done with Opus, if all
10:25 of these AI agents were Opus and Chad
10:27 GBT, I'd be spending the cost of these
10:29 Max Studios every single month. But
10:32 because I have local models working, I
10:35 can offload a lot of this work to those
10:37 local models to save tremendous amounts
10:38 of money. And yes, I still have Opus as
10:41 a part of this, but Opus, Henry, the the
10:44 chief of strategy, is only doing
10:45 decision-making. He isn't doing the
10:47 dirty work. He isn't searching Twitter
10:49 and searching Reddit and doing a lot of
10:50 the writing and creating. He just
10:53 approves and disproves. Everyone else is
10:55 doing the hard work. All the tokens are
10:58 being burnt by the local models. As I
11:01 buy more computers, as I buy more GPUs
11:02 and devices, which I fully plan on
11:04 doing, I'm going to buy the Mac Studio
11:06 M5 Ultra when that comes out in a few
11:07 months. I'm going to buy a DGX Spark.
11:08 I'm going to buy a whole lot of other
11:10 computers. By the way, Nvidia, if you're
11:12 watching, send me the DGX Spark. I will
11:14 talk about it so much. I will be adding
11:16 all these devices to my organization, to
11:18 my local data center. And as I do that,
11:20 I can run more local models. I can
11:22 expand my oneperson company. I can get
11:25 more employees in here just chugging and
11:28 working 24/7. I can even, as I add more
11:30 GPUs, train my own custom models.
11:32 Basically, train my own employees that
11:34 will be working in my company. In a
11:35 second, right after I go through this,
11:37 I'll show you how you can run your own
11:38 local models, even if you have really
11:40 crappy computers or a Mac Mini or
11:41 anything. I'll show you how to run that
11:43 in a second. But just to wrap this part
11:46 up, this is the future of AI agents.
11:49 Claudebot unlocked this. Claudebot
11:51 unlocked the ability to run your own
11:54 agents autonomously. The issue was is
11:56 you can't take full advantage of
11:58 Claudebot with APIs and cloud models.
12:01 The bottleneck once models can run
12:03 autonomously is the cost of the models
12:05 themselves. But by running these models
12:07 locally, that bottleneck disappears.
12:09 Your models can now be completely
12:12 unchained and do so many more use cases
12:14 like constantly finding challenges
12:16 online, like constantly building and
12:18 coding and creating things, like
12:20 constantly reviewing all your work. This
12:22 is the future of AI agents, and this is
12:25 the worst it will ever be. Right now,
12:28 the best local model is Kimmy K 2.5,
12:31 which is near Opus 4.5 level. Over time,
12:33 and probably by the end of the year, the
12:34 local models will be better than that.
12:35 They'll be better than that and be able
12:38 to run on much cheaper hardware. This is
12:40 the slowest, dumbest, and most expensive
12:43 it will ever be. And right now, this has
12:44 been amazing for me and what I've been
12:46 able to accomplish. Another one of the
12:48 biggest questions I get is like, okay,
12:50 that all sounds cool, but like what are
12:52 you actually getting out of all this?
12:54 What are the workflows that you're able
12:55 to enable that you weren't able to do
12:56 before? Let me give you a couple
12:59 examples here. So, here's two examples.
13:00 You can just go straight down here.
13:02 Here's a couple things we've been doing.
13:05 Number one, I have a researcher agent,
13:06 which is a local model, constantly
13:10 reading Reddit 247 365. It finds
13:11 challenges people are having in
13:13 subreddits. It hands it to my strategy
13:16 officer, which is Henry. Henry decides
13:18 if the challenge is good or not. He
13:20 takes the good challenge and hands it to
13:22 the developer agent. The developer agent
13:24 then codes an app to solve that
13:25 challenge. The developer agent then
13:27 ships that app and puts it live on
13:29 Verscell through the Verscell CLI. Then
13:31 the researcher agent DMs the original
13:33 poster and says, "Hey, we came up with
13:35 the solution to this problem. This is
13:38 just a constant closed loop. I do not
13:40 need to be a part of that is constantly
13:43 going 24/7, 365. This is not possible
13:45 with cloud APIs." Here's another
13:48 example. I record a YouTube video, just
13:51 a raw YouTube video. A local agent edits
13:53 the video. So, it cuts out all the blank
13:55 space that is then handed to a local
13:57 image model Flux 2 that is running
14:00 locally on my device that generates a
14:03 thumbnail for the video. Another agent
14:05 goes in in the browser, puts the video
14:07 onto YouTube, puts the chapters, posts
14:10 the video, and a day later, all my
14:12 agents meet. They go to the table I
14:14 showed you before, and they discuss the
14:15 performance. They go over the
14:17 transcript. What did he talk about? The
14:18 hook, the thumbnail, what worked, what
14:20 didn't work. And based on all those
14:22 learnings which get saved to their
14:24 memory, they write a new script. These
14:27 are the type of autonomous use cases
14:29 that were not possible before. They were
14:31 not possible with Chad GPT. They were
14:33 not possible with Claudebot just using
14:36 cloud APIs. It's only possible by having
14:39 compute on your desk 24/7. This is what
14:41 I can do now. This is what I'm
14:43 developing. And this is what you can do
14:45 in the new world as you start running
14:48 your own local AI models. So now the
14:50 question becomes, "Hey, Alex, do I need
14:52 to spend $20,000? That's a lot of money.
14:54 I can't afford that right now. I need to
14:56 be able to run local models and do these
14:57 use cases without spending all that
14:59 money." Well, I got good news for you.
15:01 You can do this without spending
15:03 $20,000. Can you do it to the degree,
15:06 the intelligence, and the speed at which
15:08 my models are doing it? No. But you can
15:11 start off cheap and then slowly layer on
15:12 from there. Let me show you. So, let's
15:14 talk about which local models you can
15:16 run on different budgets. You don't need
15:18 to spend $20,000 like I have. You can
15:20 have different budgets and different
15:22 local models on each. Now, are the
15:24 cheaper ones going to be as smart,
15:26 efficient, fast as larger ones? No. But
15:29 it's good to start somewhere and then as
15:32 you go, as you figure out new use cases,
15:33 as you figure out how to fit this into
15:35 your workflow, you can either buy more
15:37 hardware or change things around or
15:39 experiment. It's up to you. I don't
15:40 recommend everyone just go out and spend
15:42 $20,000 like I did. No, I do not
15:44 recommend that. Start cheap, then slowly
15:46 build your way up. So, if your budget's
15:48 only $100, that's great. You can buy a
15:51 Raspberry Pi on there. You can run
15:53 simple, small, local models like Gemma,
15:55 like Tiny Llama, things like that. And
15:56 you can do different things. You can
15:59 have simple chats. You can do smart home
16:01 things. You can do very simple things.
16:03 There are use cases there. As your
16:05 budget goes up, you can do more
16:07 interesting use cases. So, if you went
16:09 out, you're like many people like myself
16:12 that bought a Mac Mini when you
16:14 discovered OpenClaw, you can run models,
16:15 too. There's Llama models, there's
16:17 Mistral models, there's Quen models you
16:19 can run that could be like personal
16:21 assistants that can do a little bit of
16:23 coding. Will it be as good as Claude
16:25 code? No. But you can still do some
16:27 small things on it. If you have a larger
16:28 budget, you buy maybe the
16:30 top-of-the-line Mac Mini. You can start
16:32 doing some more serious coding. You can
16:34 start doing some more serious research.
16:36 Maybe your model now starts reading
16:38 Reddit at all times like mine. By the
16:39 way, as I get into the upper tiers, if
16:41 you learned anything at all, leave a
16:43 like down below. Make sure to subscribe
16:44 and turn on notifications. I'm going to
16:46 be making so many videos about these use
16:48 cases and about what I'm building. It'll
16:50 blow your mind. Make sure to turn on
16:52 notifications for that. And let me know
16:53 down below in the comments if there's
16:56 any specific part of this you want to
16:57 hear more about. Whether it's running
16:59 cheaper local models, whether it's about
17:01 running more expensive ones, use cases
17:03 for this open claw as a whole. Let me
17:05 know which one you want me to dive
17:07 deeper into for my next video. As your
17:09 budget increases, maybe you buy a Mac
17:11 Studio M2 Ultra, which is the older
17:13 generation. you can start doing more
17:15 professional workflows, have multiple
17:17 agents going at once, and then once you
17:19 get to my level, where I'm at now with
17:21 the Mac Studio M3 Ultra, so I have two
17:23 of these M3 Ultras, one's on my desk
17:25 now, another's coming in the mail this
17:27 week, you can start having a fully
17:29 autonomous organization working for you,
17:32 which is amazing, and run local models
17:34 that are almost as good as Opus, right?
17:35 And almost as good as just good enough
17:39 if it can run 24/7. This is the future.
17:40 And this might not look like much right
17:42 now what I'm showing you, but I am
17:44 actively building this out as we speak.
17:45 I'm actively building out more use
17:49 cases. Nobody else in the world is doing
17:51 anything like this. I'm not kidding.
17:53 Where we are in technology right now is
17:56 absolutely unbelievable. This is all
17:57 green field. This is a brand new
18:00 technology. Open Claw only became
18:02 popular like 2 weeks ago. We are 2 weeks
18:05 into this revolution. If you go in now,
18:08 if you tinker now and experiment now and
18:10 try new things out, the odds are you're
18:12 doing something no one else in the
18:14 entire world has done. It is early on
18:16 these technology revolutions where all
18:18 the opportunity is. I'm being serious.
18:21 This is where all the opportunity is. If
18:24 you strike now and you experiment, try
18:27 new things, invest, dedicate yourself to
18:29 this, you could create success and
18:31 opportunity for yourself that no one has
18:33 ever seen before. That's what I'm doing.
18:35 That's why I went all in and invested in
18:37 my own personal local data center here
18:39 because I want to do things the world
18:40 has never seen before. And it's
18:42 something you can do as well. If you're
18:44 joining me on this journey into the
18:47 unknown, if you're joining me on pushing
18:49 the limits of technology, if you're
18:51 joining me on trying to build the first
18:54 one billionoll business, make sure to
18:56 subscribe. Make sure to turn on
18:58 notifications. I'll be taking you
19:00 through this. I'm not holding back any
19:02 secrets whatsoever. Everything I do, I
19:04 will show to you. and you can copy and
19:06 join along with me and do incredible
19:08 things. I hope you learned something. I
19:10 absolutely love making these videos for
19:12 you. Thanks for joining along for this
