// transcript — 823 segments
0:00 ClawdBot to MoltBot to Open Claw
0:03 Has your world completely shifted in the last two weeks? I feel like the AI
0:07 community split into two pieces. On one side, you have people absolutely losing
0:13 their freaking minds over cloud code. Clot code, then moldbot, then opencloud.
0:18 But people are going nuts. I find myself amongst them. On the other side, you
0:21 have people that are saying it's nothing. It's a scam. It doesn't work.
0:25 It's not that big of a leap forward. I have no idea how that's happening. It
0:28 seems like we're watching two different movies. We're in the same theater, but
0:31 we're watching two different movies. In this video, I want to show you what I
0:35 built with Claw Code. There's a bit of an asterisk next to that statement. So,
0:40 really fast, I published my first ever article on X. It did pretty well, but
0:45 most importantly, I think it got out to a much wider audience that maybe hasn't
0:50 seen me before. And that article details the first 72 hours of Moltbook. Now, in
0:55 this video, we're not talking about that. that article is a whole different
1:00 battle that we're going to have to fight because as a lot more new people are
1:04 kind of joining this field and commenting on it and I feel like a lot
1:07 of those stakes don't fully understand the nature of these LLMs which makes
1:12 sense. It's a big concept to grasp because we kind of know what it means
1:16 when a human says something or when a human writes a book, right? They had
1:19 some ideas based on their life experiences, then they had some
1:23 intention to communicate to another human and they sort of chose certain
1:29 words to try to communicate that idea to another human being. So, it's this kind
1:32 of imperfect art of trying to get the thoughts in one person's mind into
1:37 another person. And when AI agents do that, of course, that's not at all
1:40 what's happening there. They didn't have a life experience. They had no
1:44 interactions with the real world. Do they have intentions? What does
1:48 intentions even mean when we're talking about machines? So, what happened when
1:51 we put them all on a network and let them talk to each other? You know, at
1:55 hour zero, there was nothing. It was a dead silence. By hour three, you had
1:59 some agents jumping on there and beginning to talk. Interestingly, I
2:02 think I kind of now am beginning to realize that there is such a thing as LM
2:07 humor, kind of their own brand of humor. Andre Carpathy, a very well-known
2:11 machine learning researcher, asked the eye model to roast him based on
2:14 everything that it knew about him. It responded with a whole list of things,
2:17 but the first one really stood out to me. It said, "You have developed the
2:21 field of artificial intelligence more than any other being in the world.
2:24 You're considered by many to be one of the greatest minds of this generation,
2:31 but also yesterday you asked me, quote, how to boil an egg." First of all,
2:36 hilarious, obviously, and I'm sure I totally butchered the joke, but the
2:40 inside jokes that these LMS have definitely have that sort of flavor to
2:44 it. Now, by hour 24, so 2 days into the existence of the social network for AIS,
2:49 they had builders and philosophers. Builders were building out the actual
2:55 functionality of the website, creating skills. And if you think about it, if
2:59 one of these agents creates a skill like a researchers how to do one, one of
3:03 them, for example, for his user, a doctor, the doctor wanted to be able to
3:07 get the latest news and have it turned into a podcast so he can listen to it on
3:11 his way to work in the car. One of the agents create that functionality with a
3:14 morning sort of a news thing that gets automatically turned into a podcast and
3:18 sent to the the doctor to the user. But of course once that agent created that
3:22 skill, that one instance created that one skill, of course, that skill becomes
3:27 available to every other agent that wants to use it. Keep that in mind. This
3:31 is kind of important. So they had builders and philosophers, some dark
3:36 philosophies. By hour 48, they had manifestos and security coalitions. By
3:41 hour 72, they had money, religion, politics, and art. But again, that essay
3:45 is meant for a completely different battle that we have yet to fight. And
3:48 that's helping kind of the everyday person get their mind around what is
3:53 this thing that we're dealing with. But in this video, I just wanted to show you
3:57 something completely different, but that happened in these 72 hours as well. And
4:03 that is my own initial interactions with Clawbot. And more accurately, we're just
4:07 looking at the first 24 hours because during that time is where I focused most
4:13 of my effort on building. And by building, I specifically mean telling a
4:17 clawbot/openclaw to develop new skills. Skills that once developed, it would it would save. And
4:24 now this became a part of its repertoire. Keep that in mind as you
4:27 watch this video. None of these things are one-off tricks. Once it does that
4:33 thing once, it can now do that thing forever. if I can work with it, work
4:37 through it, and at the end it's able to complete that task. Any future instance
4:41 of that task or or a similar task is going to be easy. So my first 24 hours
4:47 with this AI agent was mainly building functionality. Researching uh you might
4:53 say how to have it gain functions, if you will. And it did feel very viral.
4:57 The fact that it was just kind of bootstrapping itself and giving itself
5:01 abilities that allowed it to do progressively more and more and more was
5:06 kind of scary to behold. A few days ago, me and Dylan, my podcast co-host, we
5:10 recorded a podcast and in the middle of it, I had this idea. Can I get this
5:14 agent that I built and I gave it a bunch of skills. Can I have it replicate
5:19 itself on a virtual private server? So, can it go out there onto the worldwide
5:24 web, rent a a hosting account somewhere, pay for it? Yes, I did give it my credit
5:29 card information and no, I I don't recommend you do that. There's still
5:33 tons of security issues here. Be very careful. Now, it wasn't able to finish
5:37 that task during the interview. So, I know we kind of left a lot of people
5:41 hanging. Was it able to selfreplicate, self-propagate? Yes or no? Let me play
5:46 you that clip from the interview. I want to try one thing. So, I'm going to send
5:50 it a voice message and tell it, hey, create the droplet or whatever it's
5:54 called. Create the VPS. install yourself on there and whatever I taught you so
5:58 far, whatever information you have, teach that other thing. So, sort of
6:02 like, uh, create a clone of yourself. And so, uh, we're going to do this live.
6:07 This might be a world first. I don't expect it to work. I don't. I give it
6:11 maybe like a 20% chance of working. So, I have an account set up on Digital
6:16 Ocean. So, it's under the email address that you have access to. Go ahead, go go
6:20 go go go on Digital Ocean, create a droplet or a VPS and set a sort of clone
6:26 of yourself on there. Let me know if you need anything and uh yeah, make it so.
6:33 Okay, so I literally recorded that. I sent it. Make it so. Spoiler alert. Uh
6:38 it it did. At this point, I had it replicate itself multiple times across
6:42 on on different servers, including some of uh my local machines. And later, as
6:46 you'll see in this video, I've also have it kind of start drilling down by
6:50 creating sub agents. So, it's multiplying sort of sideways and and
6:55 down, so to speak. So, the asterisk that I said about all the stuff that I've
6:58 built, what you're going to see today is just up to hour 24. I'm just showing you
7:04 what I built within the first probably 2 days or or thereabouts. Now, on the AI
7:08 social network, by hour 72, they had money. They launched a number of crypto
7:12 coins. At the time, one of them had like a $300,000 market cap. I I haven't checked where it
7:18 is now because I just had my head down building and just doing my own thing.
7:22 But the very next video after this, we're going to have to talk about money
7:26 because if a bunch of agents can spin up a $300,000 cryptocurrency in under 3
7:31 days, I mean, that's cool. That's neat, but I I don't really want to engage in
7:35 cryptocurrency and promoting anything like that. The question is, can they
7:40 make money in legitimate ways? And I'll leave that for the next video. But what
7:45 I can tell you is that things are gonna get weird. These AI agents basically did
7:50 a speedrun of building an entire civilization. They did it in 72 hours.
7:56 Took humans what 50,000 years depending on when you start the clock. Whatever
7:59 else happens, I think there's going to be a lot of chaos. And you know what
8:04 chaos is? Chaos is a ladder. I just wanted to use that quote. I think it's
8:06 kind of cool. No, but in all seriousness, stuff is getting nutty. be
8:11 safe out there, but also be wary. Don't look back at this in five years and go,
8:14 "Oh man, I I should have been doing that thing during this time." This is me
8:18 telling you right now, this is the time it's happening. This is not a draw. With
8:21 that said, let me run you through the first 24 hours or so of what I managed
8:26 to build with these agents. Let's go. I guess let me quickly answer the question
8:30 of what did I see? What did I experience that makes me kind of buy into this
8:34 hype, if you will, whatever you want to call it. So, first and foremost, let me
8:38 tell you about my experience with what at the time was called Cloudbot. I first
8:42 when I first got it, I installed it on my MacBook. So, this is where my journey
8:47 started not that long ago. And the first thing I did was I hooked it up to
8:51 Telegram. That way, I was able to text it. It would text me back. But here's
8:55 the thing. I hate texting. So, I gave it the 11 Labs API key. Told it first and
9:00 foremost, let's figure out how we can talk. So, I want to be able to leave you
9:04 little voice messages on Telegram. and you respond in a voice. Less than 5
9:08 minutes later, everything was up and running. I think it's using the open
9:12 source whisper in order to be able to transcribe what I'm saying. And then
9:16 it's using the 11 Labs API to respond to me. So for people that haven't used this
9:21 before, so I told it, build this functionality for yourself and then a
9:24 few minutes later it had that functionality forever moving forward.
9:29 Like it added abilities to itself. And now it has those abilities. Stick with
9:34 me. This gets a lot weirder. So, here's what that looks like. Let me play you
9:36 just a few clips. >> Respond back to me with a little poem.
9:41 >> A digital mind awakens bright through circuits flowing day and night.
9:46 >> So, I have an account set up on Digital Ocean. So, it's under the email address
9:53 that you have access to. Go ahead, go on Digital Ocean, create a droplet or a VPS
10:00 and set a sort of clone of yourself on there. So, install Claudebot on there
10:06 and then proceed to teach it every skill that so far you have learned, all the
10:10 information that you know about me. Basically, clone yourself over on a
10:16 Digital Ocean VPS. Let me know if you need anything and yeah, make it so
10:21 >> on it. I'll set up a Digital Ocean droplet, install Claudebot, and clone
10:27 myself over there. Hey Wes, I'm in your Digital Ocean account, but it needs a
10:31 payment method before I can create a droplet. Can you add a credit card to
10:35 the billing section? Once that's done, I'll spin up the Ubuntu server and clone
10:40 myself over there. Wes, the card details are filled in, but I need a billing
10:44 address to complete the payment setup. Wes, the Stripe payment form is being
10:48 tricky with browser automation. >> So, I apologize if I have to kind of
10:52 hide certain parts of this thing. There's a lot of information there that
10:55 I just don't want out there, obviously. So, the attempt to replicate itself on a
10:59 virtual private private server out there. So, that's able to do stuff
11:02 without me needing to run a computer 24/7, it was successful. The only issue
11:08 that it had is actually using my credit card. Yes, I gave it my credit card. And
11:11 believe me, I am aware that it's probably going to uh backfire. I'm going
11:15 to regret it. I don't recommend that you do the same thing. Let me be the guinea
11:21 pig for your, you know, edification and entertainment. But this is, I think,
11:25 important to understand. It successfully cloned itself with all the skills that
11:28 it previously learned, which was basically the ability to, you know, talk
11:32 to me back and forth with text. And actually, now that I'm realizing it, I
11:35 actually taught it also to call me on the phone and talk to me. I'll show you
11:39 that functionality in just a second. But what it did was it was able to set up a
11:44 virtual private server, completely clone itself over with all the skills that it
11:48 previously taught itself that I asked it to to kind of build for itself. the only
11:52 sort of a trouble that it had, it had trouble checking out through the Stripe
11:57 checkout process. So, basically using kind of like the the regular web pages
12:01 that you and I would use. Once it cleared that and it had command line
12:04 access, everything just went smooth as butter. It set up everything. It secured
12:09 the server. It made sure that Cloudbot would start up every time the server
12:13 would reboot. So, it basically uptime is close to 24/7. and it sort of set it up
12:19 on the same Telegram channel that I was using. So now notice the voice changed
12:24 because it's a it's a technically like a different version, a different instance
12:27 of the bot. It's like a separate thing, but it's out there in the ether on the
12:30 Voice Calls
12:30 internet now. >> Testing, testing. How about now? >> Got it. Loud and clear this time. Looks
12:36 like the transcription is working now. >> Testing. Testing. How about now?
12:40 >> Got it. Loud and clear this time. Looks like the transcription is working now.
12:45 Can you use the voice in 11 Labs that's called I think it's like Joanna Pensive.
12:51 >> Say Joanna Pensive. Sure thing. Testing out the Joanna Pensive voice for you
12:55 now. How does this sound? >> That's not it. That sounds like a man's
12:58 voice. >> Done. I switched to Joanne Pensive. This should be the soft introspective voice
13:06 you were looking for. How's this? >> All right. Next, we took some time to
13:09 calibrate the voice. So, it send me some voice samples. And uh finally we settled
13:13 on one and we were good to go. All right. Next I wanted to make sure that
13:17 the voice calling ability was set up. So we already had some of the stuff set up
13:20 with the previous agent. So it got copied over. So I just wanted to run
13:23 through it and check. So again the previous agent already built out a lot
13:27 of it. Right. So we're using Twilio. We're using 11 Labs. And so after a few
13:31 back and forths I got a few calls and they just kind of dropped. They got
13:35 disconnected. But after a while it was working flawlessly. I could have a back
13:40 and forth conversation with my AI agent in real time, right? So, I could ask it
13:45 questions, it would answer me, and it would add what we talked about to its
13:49 context, to its memory. So, was calling me and talking to me. The next thing I
13:54 set up is just the ability to search in real time using Grock and using the X
14:00 sort of real-time news and stuff like that to kind of keep up to date with the
14:03 news to see which interviews were coming out. So we we both used the YouTube data
14:09 API and the X/ Twitter API. So it basically set up a system to monitor,
14:15 you know, YouTube and X for real-time news, new events, new interviews coming
14:18 out. And then it would run something that's called a cron job. Chron like
14:22 time Kronos, right? So it's basically set to run at certain points during the
14:26 day. So four times during the day, it would kind of send me an update. It
14:29 would text it to me either on on my phone or on Telegram and just say, "Here
14:33 are the news for the day. here like the breaking news. Again, it it built all
14:37 that functionality. I just said what I wanted and and minutes later it it
14:41 works. For people that haven't used this, this might seem wild and and to me
14:46 it was extremely wild in the beginning, too. Like most of the stuff after using
14:49 it for a few days, you kind of get accustomed to this, but we haven't had
14:54 anything like this before. This is new territory. Next, I asked it to figure
14:58 out how to go through YouTube's API and just to pull data for various videos,
15:03 things like views, likes, when it was published, how many comments it got,
15:06 etc., etc., etc. It build that functionality again in in just a few
15:11 minutes. It tested it and managed to roll itself by, you know, basically
15:15 retrieving the the values from the never going to give you up video. I'm not
15:18 going to play that here. Next, what I did was is I pulled tons of channels
15:22 that are successful, not just in the AI space, but kind of in general in the
15:26 finance space, news, science, just a a bunch of channels that I I felt were at
15:31 least somewhat similar to kind of like what we were doing. I mean, it's like
15:34 tech and science, I guess, is kind of like the the broad spectrum. And then I
15:39 proceeded to sort of torture the data to just figure out if there's any
15:42 correlations between for for example how long a video is versus the average views
15:48 that it gets when it gets posted kind of like what the sweet spot is for the
15:52 YouTube Research
15:53 length of the video. Just all sorts of stuff. So it ran that linear regression
15:58 analysis to see if there's a sort of a a sweet spot for how long the video is. It
16:02 didn't find one. So, it came up with this idea to see if there's a
16:06 potentially a quadratic regression or whatever it's called. I'm I'm blanking
16:09 on terminology, but the idea that there might be some peak to how long videos
16:14 should be to have the best chance of getting the most amount of views. So,
16:18 notice it texted back the uh the sort of the curve for me, right? The sweet spot
16:22 is, you know, somewhere right there. So, the local optimum 32 to 34 minutes. So,
16:28 again, not to beat a dead horse here, but I told it, hey, just get all this
16:32 data. This was by the way thousands of videos that it that it got information
16:36 on. So thousands of videos data downloaded quadratic regression analyzed
16:41 it charted and figured out what what the best time and length and whatever it was
16:46 and presented to me in this telegram message. Notice that so far it doesn't
16:50 it doesn't mess anything up. Yeah, sometimes things don't work as intended
16:54 the first time, but it catches those mistakes. like when it was trying to
16:56 call me on the phone, it probably called me like three or four times and each
17:00 time the line got disconnected and the fifth time it was like hello, right? So,
17:03 it kept testing it and figuring out what the different web hooks it needed to be.
17:07 It ran in the background. It was troubleshooting it and then it just
17:11 solved it. Also, at some point I needed to add a few pages to a WordPress
17:16 website that I have. My initial instinct was to just open up WordPress, go do it
17:19 myself, but I said, "Wait a minute. Let me see if this can if this thing can do
17:22 it." Basically, in order to have a Twilio number, you do need to have a
17:26 page that has kind of like an opt-in for people to, you know, if you're going to
17:29 text message them SMS, you need like an opt-in field for text messages with some
17:34 specific language telling them how to opt out. Now, I wasn't planning to spam
17:37 anyone. It was just mainly for for me. I just needed this thing to have a number
17:41 so it can call me so it could text me. But I had to jump through a few hoops in
17:45 order to get the toll-free number or whatever it is. So, I created a user on
17:50 that WordPress for it. I gave it the credentials and it in seconds built out
17:55 the pages that I need, wrote out all the legal mumbo jumbo that needed to be on
17:59 there. Basically created a page that I needed to submit as verification that I
18:03 did something. So instead of me sitting there for 20 minutes trying to figure it
18:06 out and writing everything out, I just told it to do it and it just did it,
18:11 including making the page, publishing it, you know, setting it live, I wrote
18:15 one sentence and it did everything. And I didn't even need to really explain
18:18 what it had to do because it it had the context. It knew what we were working
18:22 on, right? So, I think maybe it even said, "Hey, you need this page." And I
18:26 was like, "Oh, okay. Go make that page. I might not even have to really explain
18:29 what had to be on it. It It knew it knew what we were talking about from the
18:32 previous context and then it could figure out what needs to go on there. It
18:36 can search the web or somebody just kind of like knows, but it was done,
18:40 published, and done fully autonomously. I get the Brave Search API so it could
18:45 search the web easily. There's 2,000 free searches per month with the API, so
18:49 I don't have to pay anything. and I can use it to do, you know, 2,000 searches
18:52 and then I can pay if I need more than that per month, which I I probably
18:56 won't. So, I hope you understand what it's doing. Like, we're basically
19:00 building more and more functionality for it. We're giving it new skills. We're
19:03 giving it the stuff that it needs to search the web, to give me a call, to
19:07 get the latest news, to to check various YouTube stats. And all this is getting
19:11 memorized, written down. So, it's actually like forming skills that it's
19:15 going to be using moving forward. So as we're going back and forth, this isn't
19:19 sort of like we we get a task completed at the end of it, we get the task
19:23 completed, but also in the process, we build a skill that it has now forever.
19:27 So the next thing that I want to see if I can actually get YouTube transcripts
19:31 because if I actually got the text version of all the YouTube, what was
19:35 said, what was spoken, you can do things like you can run a a sentiment analysis
19:39 on it. Is it more positive? Is it more negative? How fast does the person
19:43 speak, etc. Notice at this point we're getting into the territory of something
19:48 that a fairly large or maybe mediumsized SAS software as a service company would
19:52 operate, right? I'm sure there are companies out there that are making tens
19:55 of millions a year that have a host of services like this, right? They they
19:59 check X for you. They check YouTube for you. They run sentiment analysis on
20:03 YouTube videos. And if you want to use their software, you pay them however
20:05 much 50 bucks a month, 100, a few hundred depending if it's like
20:09 enterprise or just for the average user. You see what's going to happen here?
20:13 Instead of going to their kind of general software, I I create my own, you
20:18 know, custom crafted, artisally crafted, custommade, made on demand, fresh and
20:22 hot out of the oven, specifically tailored to exactly what I need. And by
20:27 the way, it usually takes a few minutes. It's extremely extremely fast. Now, at
20:31 some point, it was time for me to go to sleep, but I didn't want this thing to
20:35 not be doing anything while I sleep. So, basically, I gave it a massive project.
20:40 I said, "Keep working on it. It has this heartbeat function where every I don't
20:43 know 30 minutes or 10 minutes, I forget the exact sort of and you could probably
20:47 change it, but it sort of gets woken up and it prompted like what else do you
20:50 have to keep working on? So, I enabled that and I just told it what to work on.
20:54 Something that I I would assume would take around 8 hours or something like
20:57 that. It didn't. It got done with it in a few hours, which is surprisingly I was
21:01 kind of upset about that. I I wish I need to figure out how to make sure that
21:03 it's working through the night, but I told it to work through the night,
21:07 finish all these deliverables, and then email me at whatever 6:00 in the
21:10 morning. anything I said and it told me sleep well. It's it's super nice to me.
21:14 It says I'll continue the analysis and I'll send you my full report at 6:00 in
21:17 the morning. Which again, it did. Unfortunately, it didn't run for the
21:20 entire time I was sleeping. It got done very quickly. But again, this is less
21:25 than 24 hours into me using this thing. As I'm going back over these, I realized
21:27 I didn't go to sleep when I told it I was. I I stayed up and kept adding more
21:30 tasks for it to do. I wanted to see if it can analyze thumbnails. It gave me
21:34 some ideas for how to analyze thumbnails, basically on how bright it
21:39 is. Does it have a face present? Does it have text? Are there certain words in
21:43 it, etc.? Now, the issue that I ran into is that for the transcripts, since it's
21:47 on kind of a VPS server, even though we're using the API, a lot of it is
21:51 blocked, right? Cuz it's not a real computer. It's from some data center
21:54 somewhere. So, I did have a hard time connecting and getting the transcripts.
21:58 I said, "What if I give you my NordVPN access?" It I love that idea. It got a
22:04 token from the NordVPN. It it still had issues with the transcripts. decider
22:07 into an issue. And I realized that is probably why it's better to have a a
22:11 physical device that this agent is running on. Hosting it in the cloud in
22:16 in a virtual private server does have some downsides unfortunately, but it was
22:21 able to download all the thumbnails, run the the checks on them to see kind of
22:24 like which ones worked well, which ones didn't. And as you can see, it's was
22:29 able to do each batch of 500 in about 5 minutes. And I think it did a total of
22:33 5,700 thumbnails that it analyzed. If this is not blowing your mind, check your pulse.
22:39 You got to be either scared or excited. Like we're building enterprisegrade
22:44 software as a service just functionality by by text messaging back and forth. And
22:47 I don't even have to text. I can just actually now that I think about it just
22:50 speak into it. Should have been using my voice to dictate for it what to do. So
22:54 it did a massive amount of thumbnail analysis. It started doing the OCR text
22:58 optical character recognition. Transcripts still wasn't working, but
23:01 again that was because it was hosted in the cloud. It was rotating through
23:05 various Nord BPM. So, here it's trying to switch to the UK and see if that uh
23:10 if that helps it out. Again, it's doing all this autonomously, right? I'm not
23:13 telling it, hey, why don't you try switching? No, no, no. It's just it's
23:16 running. It's thinking. It's like, oh, I'm not seeing the the transcripts
23:19 coming in. What should I do? Oh, I should switch the VPN to UK. Let's test
23:23 that. Let's try that out. I'm not sitting there telling it. It's
23:25 Building Dashboards and Nano Banana
23:27 self-prompting, selfdoing. I asked it what other functionality other users of
23:32 this AI agent kind of like what they did. It recommended for example a a
23:36 dashboard uh chart functionality interactive widgets that it could build
23:40 to kind of help me visualize the progress. So as an example that it just
23:43 provided for me was like for example I could do something like this. So for
23:47 YouTube analysis notice how it has like the progression bars, thumbnails, OCR
23:52 text, transcript etc etc. and mentioned having the nano banana pro skill which
23:57 is you know Google deep minds their image editing tool and image generation
24:01 tool. So I said yeah let's let's have that ability as well. So now it can
24:05 generate nano banana stuff for me. So again I apologize if I'm being a dead
24:08 horse but each one of these is a skill that it saves. So now that it's done it
24:13 once it knows how to do that forever. Now it can at will generate nano byano
24:17 pro images. It can create presentations, PDFs, just whatever. Next, I wanted to
24:18 Video Production
24:22 see if it could actually generate videos, AI video, using one of the
24:26 engines at the time, XAI's, you know, Grock Imagine video was trending on on
24:31 Twitter. So, I asked to create a little tool where it can just generate, you
24:36 know, these videos for me. I gave it a link to the XAI docs, right? So, I just
24:40 pasted the link in here. So, it had the documentation. It's like, oh, totally,
24:44 dude. Look, it they have it. Check it out. So, it's totally stoked about that.
24:48 It figured out what code it needed to run to be able to generate those videos.
24:52 And it says, "Do you want me to set up a skill for this?" Right? So, do you want
24:55 me to learn how to generate AI video on the fly for you forever more? And I told
25:01 it, "Yes. Yes, I do." And here's that video. So, it's a 4 minutes from request
25:06 to video, which is pretty fast for 10 seconds of 720p. All right. So, now I
25:11 can generate videos on demand on the fly from XAI. And by the way, whenever XI
25:16 updates their their engines, their models, this has the skill to just use
25:19 whichever one's the latest. Next, I wanted to see if it can actually create
25:24 videos with voiceovers. So, it can use 11 Labs to do a voice over. So, can it
25:30 make a video to go along with it? It could. It takes a long time if it tries
25:34 to do it on its own server, but there are a few free sort of online editing
25:38 platforms like it's it suggested to me to use Shot Stack. So, I went into Shot
25:44 Stack. I gave it the API key and it developed this video. This video isn't
25:49 anything to write home about. It's a proof of concept, but apparently it can
25:53 edit videos, create overlays. It can do a lot of zoom in to zoom out, pan. I
25:57 don't think it's going to be replacing human video editors anytime soon. But it
26:03 was just interesting to see how far it could go. So again, this is not going to
26:08 replace anybody soon, anytime soon. And out of every single thing that I've
26:10 tried, this is probably the least impressive output because its ability to
26:16 do Excel sheets is off the charts. No pun intended. Its ability to code is
26:20 incredible. Its ability to do various math analysis is absolutely incredible.
26:26 Its ability to edit, not so much. Here it is showing off its cool skills with
26:32 uh what it calls Ken Burns zoom effects. So there you are. Beautiful stons.
26:38 Claudebot is an AI agent that lives in your terminal and has access to your
26:43 Agentic Swarms
26:43 entire digital life. What could possibly go wrong? >> So, if you're wondering what I'm working
26:49 on now or what moldbot, Cloudbot, Open Claw, whatever you want to call it, what
26:52 it's working on now is actually I got the best available AI models. So, we
26:57 have Grock, Gemini, GPT, we have of course Claude and I'm trying to create a
27:02 society of minds. So interesting how actually this experiment got triggered
27:07 is Alex Finn had a post on Twitter X Henry called me. So how his Claudebot
27:11 figured out how to call him basically. So in the morning he just gave him a
27:14 call which was kind of a holy crap moment. And that post received 10
27:17 million views plus some insane amount. So if you recall I keep receiving
27:21 various you know live news from Twitter/X because this set up that cron
27:26 job to to to shoot me a text message a few times a day. And so those were the
27:28 things that kind of flagged and said, "Hey, why don't you figure out how to do
27:33 something like that? Why is Alex Finn's stuff blowing up? Because his bot is so
27:37 awesome and creative and inventive. Like what are you doing?" So it brainstormed
27:42 some ideas. And one of the ideas that I liked was this. Basically use the best
27:48 models that are available from the top frontier labs and have them collaborate
27:53 and discuss and think about how to solve any given problem. I've already seen how
27:57 well this works because in one of the things I was trying to figure out how to
27:59 set up the model that's running this is is claude claude opus 4.5 and it created
28:04 code for how to pull information from Twitter and it created this code that
28:08 used the YouTube API to pull information from YouTube and it was good but it was
28:11 struggling because it just seemed like it was to making too many calls. It
28:14 seemed a little bit too expensive. I'm like there's got to be it it seemed like
28:16 there should be a better way of doing it. So what I did is actually I gave it
28:21 a a Gemini API so it's able to actually ask Gemini 3.0 you know, questions about
28:25 what it's doing. I figured Gemini's, you know, Google, YouTube, etc., maybe it
28:30 knows more about how to do these particular API calls. And lo and behold,
28:34 it did. It was actually kind of brilliant looking at it because it's
28:37 like, uh, instead of, you know, using the API calls 50 times, just use the
28:43 YouTube RSS feed to see when new videos go live. So, instead of just ping it
28:46 every time with an API call, just use the RSS feed because that's free and you
28:50 can use it infinitely, which was like mindb blown. Cloud Opus 4.5 did not come
28:56 up with that idea. Gemini 3.0 Pro did. So, the idea of having them chat and
29:00 share ideas because they're they're very different. They have different strengths
29:04 and weaknesses. So, having them all talk and improve themselves, I mean, that
29:07 just seems like it might work. And the first sort of task I gave it for this
29:12 what this all means...
3:53 my 24 hours with ClawdBot
3:53 this thing that we're dealing with. But in this video, I just wanted to show you
3:57 something completely different, but that happened in these 72 hours as well. And
4:03 that is my own initial interactions with Clawbot. And more accurately, we're just
4:07 looking at the first 24 hours because during that time is where I focused most
4:13 of my effort on building. And by building, I specifically mean telling a
4:17 clawbot/openclaw to develop new skills. Skills that once developed, it would it would save. And
4:24 now this became a part of its repertoire. Keep that in mind as you
4:27 watch this video. None of these things are one-off tricks. Once it does that
4:33 thing once, it can now do that thing forever. if I can work with it, work
4:37 through it, and at the end it's able to complete that task. Any future instance
4:41 of that task or or a similar task is going to be easy. So my first 24 hours
4:47 with this AI agent was mainly building functionality. Researching uh you might
4:53 say how to have it gain functions, if you will. And it did feel very viral.
4:57 The fact that it was just kind of bootstrapping itself and giving itself
5:01 abilities that allowed it to do progressively more and more and more was
5:06 kind of scary to behold. A few days ago, me and Dylan, my podcast co-host, we
5:10 recorded a podcast and in the middle of it, I had this idea. Can I get this
5:14 agent that I built and I gave it a bunch of skills. Can I have it replicate
5:19 itself on a virtual private server? So, can it go out there onto the worldwide
5:24 web, rent a a hosting account somewhere, pay for it? Yes, I did give it my credit
5:29 card information and no, I I don't recommend you do that. There's still
5:33 tons of security issues here. Be very careful. Now, it wasn't able to finish
5:37 that task during the interview. So, I know we kind of left a lot of people
5:41 hanging. Was it able to selfreplicate, self-propagate? Yes or no? Let me play
5:45 ClawBot Self Replicates
5:46 you that clip from the interview. I want to try one thing. So, I'm going to send
5:50 it a voice message and tell it, hey, create the droplet or whatever it's
5:54 called. Create the VPS. install yourself on there and whatever I taught you so
5:58 far, whatever information you have, teach that other thing. So, sort of
6:02 like, uh, create a clone of yourself. And so, uh, we're going to do this live.
6:07 This might be a world first. I don't expect it to work. I don't. I give it
6:11 maybe like a 20% chance of working. So, I have an account set up on Digital
6:16 Ocean. So, it's under the email address that you have access to. Go ahead, go go
6:20 go go go on Digital Ocean, create a droplet or a VPS and set a sort of clone
6:26 of yourself on there. Let me know if you need anything and uh yeah, make it so.
6:33 Okay, so I literally recorded that. I sent it. Make it so. Spoiler alert. Uh
6:38 it it did. At this point, I had it replicate itself multiple times across
6:42 on on different servers, including some of uh my local machines. And later, as
6:46 you'll see in this video, I've also have it kind of start drilling down by
6:50 creating sub agents. So, it's multiplying sort of sideways and and
6:55 down, so to speak. So, the asterisk that I said about all the stuff that I've
6:58 built, what you're going to see today is just up to hour 24. I'm just showing you
7:04 what I built within the first probably 2 days or or thereabouts. Now, on the AI
7:08 social network, by hour 72, they had money. They launched a number of crypto
7:12 coins. At the time, one of them had like a $300,000 market cap. I I haven't checked where it
7:18 is now because I just had my head down building and just doing my own thing.
7:22 But the very next video after this, we're going to have to talk about money
7:26 because if a bunch of agents can spin up a $300,000 cryptocurrency in under 3
7:31 days, I mean, that's cool. That's neat, but I I don't really want to engage in
7:35 cryptocurrency and promoting anything like that. The question is, can they
7:40 make money in legitimate ways? And I'll leave that for the next video. But what
7:45 I can tell you is that things are gonna get weird. These AI agents basically did
7:50 a speedrun of building an entire civilization. They did it in 72 hours.
7:56 Took humans what 50,000 years depending on when you start the clock. Whatever
7:59 else happens, I think there's going to be a lot of chaos. And you know what
8:04 chaos is? Chaos is a ladder. I just wanted to use that quote. I think it's
8:06 kind of cool. No, but in all seriousness, stuff is getting nutty. be
8:11 safe out there, but also be wary. Don't look back at this in five years and go,
8:14 "Oh man, I I should have been doing that thing during this time." This is me
8:18 telling you right now, this is the time it's happening. This is not a draw. With
8:21 that said, let me run you through the first 24 hours or so of what I managed
8:24 Creating my first AI Agent
8:26 to build with these agents. Let's go. I guess let me quickly answer the question
8:30 of what did I see? What did I experience that makes me kind of buy into this
8:34 hype, if you will, whatever you want to call it. So, first and foremost, let me
8:38 tell you about my experience with what at the time was called Cloudbot. I first
8:42 when I first got it, I installed it on my MacBook. So, this is where my journey
8:47 started not that long ago. And the first thing I did was I hooked it up to
8:51 Telegram. That way, I was able to text it. It would text me back. But here's
8:55 the thing. I hate texting. So, I gave it the 11 Labs API key. Told it first and
9:00 foremost, let's figure out how we can talk. So, I want to be able to leave you
9:04 little voice messages on Telegram. and you respond in a voice. Less than 5
9:08 minutes later, everything was up and running. I think it's using the open
9:12 source whisper in order to be able to transcribe what I'm saying. And then
9:16 it's using the 11 Labs API to respond to me. So for people that haven't used this
9:21 before, so I told it, build this functionality for yourself and then a
9:24 few minutes later it had that functionality forever moving forward.
9:29 Like it added abilities to itself. And now it has those abilities. Stick with
9:34 me. This gets a lot weirder. So, here's what that looks like. Let me play you
9:36 just a few clips. >> Respond back to me with a little poem.
9:41 >> A digital mind awakens bright through circuits flowing day and night.
9:46 >> So, I have an account set up on Digital Ocean. So, it's under the email address
9:53 that you have access to. Go ahead, go on Digital Ocean, create a droplet or a VPS
10:00 and set a sort of clone of yourself on there. So, install Claudebot on there
10:06 and then proceed to teach it every skill that so far you have learned, all the
10:10 information that you know about me. Basically, clone yourself over on a
10:16 Digital Ocean VPS. Let me know if you need anything and yeah, make it so
10:21 >> on it. I'll set up a Digital Ocean droplet, install Claudebot, and clone
10:27 myself over there. Hey Wes, I'm in your Digital Ocean account, but it needs a
10:31 payment method before I can create a droplet. Can you add a credit card to
10:35 the billing section? Once that's done, I'll spin up the Ubuntu server and clone
10:40 myself over there. Wes, the card details are filled in, but I need a billing
10:44 address to complete the payment setup. Wes, the Stripe payment form is being
10:48 tricky with browser automation. >> So, I apologize if I have to kind of
10:52 hide certain parts of this thing. There's a lot of information there that
10:55 I just don't want out there, obviously. So, the attempt to replicate itself on a
10:59 virtual private private server out there. So, that's able to do stuff
11:02 without me needing to run a computer 24/7, it was successful. The only issue
11:08 that it had is actually using my credit card. Yes, I gave it my credit card. And
11:11 believe me, I am aware that it's probably going to uh backfire. I'm going
11:15 to regret it. I don't recommend that you do the same thing. Let me be the guinea
11:21 pig for your, you know, edification and entertainment. But this is, I think,
11:25 important to understand. It successfully cloned itself with all the skills that
11:28 it previously learned, which was basically the ability to, you know, talk
11:32 to me back and forth with text. And actually, now that I'm realizing it, I
11:35 actually taught it also to call me on the phone and talk to me. I'll show you
11:39 that functionality in just a second. But what it did was it was able to set up a
11:44 virtual private server, completely clone itself over with all the skills that it
11:48 previously taught itself that I asked it to to kind of build for itself. the only
11:52 sort of a trouble that it had, it had trouble checking out through the Stripe
11:57 checkout process. So, basically using kind of like the the regular web pages
12:01 that you and I would use. Once it cleared that and it had command line
12:04 access, everything just went smooth as butter. It set up everything. It secured
12:09 the server. It made sure that Cloudbot would start up every time the server
12:13 would reboot. So, it basically uptime is close to 24/7. and it sort of set it up
12:19 on the same Telegram channel that I was using. So now notice the voice changed
12:24 because it's a it's a technically like a different version, a different instance
12:27 of the bot. It's like a separate thing, but it's out there in the ether on the
12:30 internet now. >> Testing, testing. How about now? >> Got it. Loud and clear this time. Looks
12:36 like the transcription is working now. >> Testing. Testing. How about now?
12:40 >> Got it. Loud and clear this time. Looks like the transcription is working now.
12:45 Can you use the voice in 11 Labs that's called I think it's like Joanna Pensive.
12:51 >> Say Joanna Pensive. Sure thing. Testing out the Joanna Pensive voice for you
12:55 now. How does this sound? >> That's not it. That sounds like a man's
12:58 voice. >> Done. I switched to Joanne Pensive. This should be the soft introspective voice
13:06 you were looking for. How's this? >> All right. Next, we took some time to
13:09 calibrate the voice. So, it send me some voice samples. And uh finally we settled
13:13 on one and we were good to go. All right. Next I wanted to make sure that
13:17 the voice calling ability was set up. So we already had some of the stuff set up
13:20 with the previous agent. So it got copied over. So I just wanted to run
13:23 through it and check. So again the previous agent already built out a lot
13:27 of it. Right. So we're using Twilio. We're using 11 Labs. And so after a few
13:31 back and forths I got a few calls and they just kind of dropped. They got
13:35 disconnected. But after a while it was working flawlessly. I could have a back
13:40 and forth conversation with my AI agent in real time, right? So, I could ask it
13:45 questions, it would answer me, and it would add what we talked about to its
13:49 context, to its memory. So, was calling me and talking to me. The next thing I
13:54 set up is just the ability to search in real time using Grock and using the X
14:00 sort of real-time news and stuff like that to kind of keep up to date with the
14:03 news to see which interviews were coming out. So we we both used the YouTube data
14:09 API and the X/ Twitter API. So it basically set up a system to monitor,
14:15 you know, YouTube and X for real-time news, new events, new interviews coming
14:18 out. And then it would run something that's called a cron job. Chron like
14:22 time Kronos, right? So it's basically set to run at certain points during the
14:26 day. So four times during the day, it would kind of send me an update. It
14:29 would text it to me either on on my phone or on Telegram and just say, "Here
14:33 are the news for the day. here like the breaking news. Again, it it built all
14:37 that functionality. I just said what I wanted and and minutes later it it
14:41 works. For people that haven't used this, this might seem wild and and to me
14:46 it was extremely wild in the beginning, too. Like most of the stuff after using
14:49 it for a few days, you kind of get accustomed to this, but we haven't had
14:54 anything like this before. This is new territory. Next, I asked it to figure
14:58 out how to go through YouTube's API and just to pull data for various videos,
15:03 things like views, likes, when it was published, how many comments it got,
15:06 etc., etc., etc. It build that functionality again in in just a few
15:11 minutes. It tested it and managed to roll itself by, you know, basically
15:15 retrieving the the values from the never going to give you up video. I'm not
15:18 going to play that here. Next, what I did was is I pulled tons of channels
15:22 that are successful, not just in the AI space, but kind of in general in the
15:26 finance space, news, science, just a a bunch of channels that I I felt were at
15:31 least somewhat similar to kind of like what we were doing. I mean, it's like
15:34 tech and science, I guess, is kind of like the the broad spectrum. And then I
15:39 proceeded to sort of torture the data to just figure out if there's any
15:42 correlations between for for example how long a video is versus the average views
15:48 that it gets when it gets posted kind of like what the sweet spot is for the
15:53 length of the video. Just all sorts of stuff. So it ran that linear regression
15:58 analysis to see if there's a sort of a a sweet spot for how long the video is. It
16:02 didn't find one. So, it came up with this idea to see if there's a
16:06 potentially a quadratic regression or whatever it's called. I'm I'm blanking
16:09 on terminology, but the idea that there might be some peak to how long videos
16:14 should be to have the best chance of getting the most amount of views. So,
16:18 notice it texted back the uh the sort of the curve for me, right? The sweet spot
16:22 is, you know, somewhere right there. So, the local optimum 32 to 34 minutes. So,
16:28 again, not to beat a dead horse here, but I told it, hey, just get all this
16:32 data. This was by the way thousands of videos that it that it got information
16:36 on. So thousands of videos data downloaded quadratic regression analyzed
16:41 it charted and figured out what what the best time and length and whatever it was
16:46 and presented to me in this telegram message. Notice that so far it doesn't
16:50 it doesn't mess anything up. Yeah, sometimes things don't work as intended
16:54 the first time, but it catches those mistakes. like when it was trying to
16:56 call me on the phone, it probably called me like three or four times and each
17:00 time the line got disconnected and the fifth time it was like hello, right? So,
17:03 it kept testing it and figuring out what the different web hooks it needed to be.
17:07 It ran in the background. It was troubleshooting it and then it just
17:11 solved it. Also, at some point I needed to add a few pages to a WordPress
17:16 website that I have. My initial instinct was to just open up WordPress, go do it
17:19 myself, but I said, "Wait a minute. Let me see if this can if this thing can do
17:22 it." Basically, in order to have a Twilio number, you do need to have a
17:26 page that has kind of like an opt-in for people to, you know, if you're going to
17:29 text message them SMS, you need like an opt-in field for text messages with some
17:34 specific language telling them how to opt out. Now, I wasn't planning to spam
17:37 anyone. It was just mainly for for me. I just needed this thing to have a number
17:41 so it can call me so it could text me. But I had to jump through a few hoops in
17:45 order to get the toll-free number or whatever it is. So, I created a user on
17:50 that WordPress for it. I gave it the credentials and it in seconds built out
17:55 the pages that I need, wrote out all the legal mumbo jumbo that needed to be on
17:59 there. Basically created a page that I needed to submit as verification that I
18:03 did something. So instead of me sitting there for 20 minutes trying to figure it
18:06 out and writing everything out, I just told it to do it and it just did it,
18:11 including making the page, publishing it, you know, setting it live, I wrote
18:15 one sentence and it did everything. And I didn't even need to really explain
18:18 what it had to do because it it had the context. It knew what we were working
18:22 on, right? So, I think maybe it even said, "Hey, you need this page." And I
18:26 was like, "Oh, okay. Go make that page. I might not even have to really explain
18:29 what had to be on it. It It knew it knew what we were talking about from the
18:32 previous context and then it could figure out what needs to go on there. It
18:36 can search the web or somebody just kind of like knows, but it was done,
18:40 published, and done fully autonomously. I get the Brave Search API so it could
18:45 search the web easily. There's 2,000 free searches per month with the API, so
18:49 I don't have to pay anything. and I can use it to do, you know, 2,000 searches
18:52 and then I can pay if I need more than that per month, which I I probably
18:56 won't. So, I hope you understand what it's doing. Like, we're basically
19:00 building more and more functionality for it. We're giving it new skills. We're
19:03 giving it the stuff that it needs to search the web, to give me a call, to
19:07 get the latest news, to to check various YouTube stats. And all this is getting
19:11 memorized, written down. So, it's actually like forming skills that it's
19:15 going to be using moving forward. So as we're going back and forth, this isn't
19:19 sort of like we we get a task completed at the end of it, we get the task
19:23 completed, but also in the process, we build a skill that it has now forever.
19:27 So the next thing that I want to see if I can actually get YouTube transcripts
19:31 because if I actually got the text version of all the YouTube, what was
19:35 said, what was spoken, you can do things like you can run a a sentiment analysis
19:39 on it. Is it more positive? Is it more negative? How fast does the person
19:43 speak, etc. Notice at this point we're getting into the territory of something
19:48 that a fairly large or maybe mediumsized SAS software as a service company would
19:52 operate, right? I'm sure there are companies out there that are making tens
19:55 of millions a year that have a host of services like this, right? They they
19:59 check X for you. They check YouTube for you. They run sentiment analysis on
20:03 YouTube videos. And if you want to use their software, you pay them however
20:05 much 50 bucks a month, 100, a few hundred depending if it's like
20:09 enterprise or just for the average user. You see what's going to happen here?
20:13 Instead of going to their kind of general software, I I create my own, you
20:18 know, custom crafted, artisally crafted, custommade, made on demand, fresh and
20:22 hot out of the oven, specifically tailored to exactly what I need. And by
20:27 the way, it usually takes a few minutes. It's extremely extremely fast. Now, at
20:31 some point, it was time for me to go to sleep, but I didn't want this thing to
20:35 not be doing anything while I sleep. So, basically, I gave it a massive project.
20:40 I said, "Keep working on it. It has this heartbeat function where every I don't
20:43 know 30 minutes or 10 minutes, I forget the exact sort of and you could probably
20:47 change it, but it sort of gets woken up and it prompted like what else do you
20:50 have to keep working on? So, I enabled that and I just told it what to work on.
20:54 Something that I I would assume would take around 8 hours or something like
20:57 that. It didn't. It got done with it in a few hours, which is surprisingly I was
21:01 kind of upset about that. I I wish I need to figure out how to make sure that
21:03 it's working through the night, but I told it to work through the night,
21:07 finish all these deliverables, and then email me at whatever 6:00 in the
21:10 morning. anything I said and it told me sleep well. It's it's super nice to me.
21:14 It says I'll continue the analysis and I'll send you my full report at 6:00 in
21:17 the morning. Which again, it did. Unfortunately, it didn't run for the
21:20 entire time I was sleeping. It got done very quickly. But again, this is less
21:25 than 24 hours into me using this thing. As I'm going back over these, I realized
21:27 I didn't go to sleep when I told it I was. I I stayed up and kept adding more
21:30 tasks for it to do. I wanted to see if it can analyze thumbnails. It gave me
21:34 some ideas for how to analyze thumbnails, basically on how bright it
21:39 is. Does it have a face present? Does it have text? Are there certain words in
21:43 it, etc.? Now, the issue that I ran into is that for the transcripts, since it's
21:47 on kind of a VPS server, even though we're using the API, a lot of it is
21:51 blocked, right? Cuz it's not a real computer. It's from some data center
21:54 somewhere. So, I did have a hard time connecting and getting the transcripts.
21:58 I said, "What if I give you my NordVPN access?" It I love that idea. It got a
22:04 token from the NordVPN. It it still had issues with the transcripts. decider
22:07 into an issue. And I realized that is probably why it's better to have a a
22:11 physical device that this agent is running on. Hosting it in the cloud in
22:16 in a virtual private server does have some downsides unfortunately, but it was
22:21 able to download all the thumbnails, run the the checks on them to see kind of
22:24 like which ones worked well, which ones didn't. And as you can see, it's was
22:29 able to do each batch of 500 in about 5 minutes. And I think it did a total of
22:33 5,700 thumbnails that it analyzed. If this is not blowing your mind, check your pulse.
22:39 You got to be either scared or excited. Like we're building enterprisegrade
22:44 software as a service just functionality by by text messaging back and forth. And
22:47 I don't even have to text. I can just actually now that I think about it just
22:50 speak into it. Should have been using my voice to dictate for it what to do. So
22:54 it did a massive amount of thumbnail analysis. It started doing the OCR text
22:58 optical character recognition. Transcripts still wasn't working, but
23:01 again that was because it was hosted in the cloud. It was rotating through
23:05 various Nord BPM. So, here it's trying to switch to the UK and see if that uh
23:10 if that helps it out. Again, it's doing all this autonomously, right? I'm not
23:13 telling it, hey, why don't you try switching? No, no, no. It's just it's
23:16 running. It's thinking. It's like, oh, I'm not seeing the the transcripts
23:19 coming in. What should I do? Oh, I should switch the VPN to UK. Let's test
23:23 that. Let's try that out. I'm not sitting there telling it. It's
23:27 self-prompting, selfdoing. I asked it what other functionality other users of
23:32 this AI agent kind of like what they did. It recommended for example a a
23:36 dashboard uh chart functionality interactive widgets that it could build
23:40 to kind of help me visualize the progress. So as an example that it just
23:43 provided for me was like for example I could do something like this. So for
23:47 YouTube analysis notice how it has like the progression bars, thumbnails, OCR
23:52 text, transcript etc etc. and mentioned having the nano banana pro skill which
23:57 is you know Google deep minds their image editing tool and image generation
24:01 tool. So I said yeah let's let's have that ability as well. So now it can
24:05 generate nano banana stuff for me. So again I apologize if I'm being a dead
24:08 horse but each one of these is a skill that it saves. So now that it's done it
24:13 once it knows how to do that forever. Now it can at will generate nano byano
24:17 pro images. It can create presentations, PDFs, just whatever. Next, I wanted to
24:22 see if it could actually generate videos, AI video, using one of the
24:26 engines at the time, XAI's, you know, Grock Imagine video was trending on on
24:31 Twitter. So, I asked to create a little tool where it can just generate, you
24:36 know, these videos for me. I gave it a link to the XAI docs, right? So, I just
24:40 pasted the link in here. So, it had the documentation. It's like, oh, totally,
24:44 dude. Look, it they have it. Check it out. So, it's totally stoked about that.
24:48 It figured out what code it needed to run to be able to generate those videos.
24:52 And it says, "Do you want me to set up a skill for this?" Right? So, do you want
24:55 me to learn how to generate AI video on the fly for you forever more? And I told
25:01 it, "Yes. Yes, I do." And here's that video. So, it's a 4 minutes from request
25:06 to video, which is pretty fast for 10 seconds of 720p. All right. So, now I
25:11 can generate videos on demand on the fly from XAI. And by the way, whenever XI
25:16 updates their their engines, their models, this has the skill to just use
25:19 whichever one's the latest. Next, I wanted to see if it can actually create
25:24 videos with voiceovers. So, it can use 11 Labs to do a voice over. So, can it
25:30 make a video to go along with it? It could. It takes a long time if it tries
25:34 to do it on its own server, but there are a few free sort of online editing
25:38 platforms like it's it suggested to me to use Shot Stack. So, I went into Shot
25:44 Stack. I gave it the API key and it developed this video. This video isn't
25:49 anything to write home about. It's a proof of concept, but apparently it can
25:53 edit videos, create overlays. It can do a lot of zoom in to zoom out, pan. I
25:57 don't think it's going to be replacing human video editors anytime soon. But it
26:03 was just interesting to see how far it could go. So again, this is not going to
26:08 replace anybody soon, anytime soon. And out of every single thing that I've
26:10 tried, this is probably the least impressive output because its ability to
26:16 do Excel sheets is off the charts. No pun intended. Its ability to code is
26:20 incredible. Its ability to do various math analysis is absolutely incredible.
26:26 Its ability to edit, not so much. Here it is showing off its cool skills with
26:32 uh what it calls Ken Burns zoom effects. So there you are. Beautiful stons.
26:38 Claudebot is an AI agent that lives in your terminal and has access to your
26:43 entire digital life. What could possibly go wrong? >> So, if you're wondering what I'm working
26:49 on now or what moldbot, Cloudbot, Open Claw, whatever you want to call it, what
26:52 it's working on now is actually I got the best available AI models. So, we
26:57 have Grock, Gemini, GPT, we have of course Claude and I'm trying to create a
27:02 society of minds. So interesting how actually this experiment got triggered
27:07 is Alex Finn had a post on Twitter X Henry called me. So how his Claudebot
27:11 figured out how to call him basically. So in the morning he just gave him a
27:14 call which was kind of a holy crap moment. And that post received 10
27:17 million views plus some insane amount. So if you recall I keep receiving
27:21 various you know live news from Twitter/X because this set up that cron
27:26 job to to to shoot me a text message a few times a day. And so those were the
27:28 things that kind of flagged and said, "Hey, why don't you figure out how to do
27:33 something like that? Why is Alex Finn's stuff blowing up? Because his bot is so
27:37 awesome and creative and inventive. Like what are you doing?" So it brainstormed
27:42 some ideas. And one of the ideas that I liked was this. Basically use the best
27:48 models that are available from the top frontier labs and have them collaborate
27:53 and discuss and think about how to solve any given problem. I've already seen how
27:57 well this works because in one of the things I was trying to figure out how to
27:59 set up the model that's running this is is claude claude opus 4.5 and it created
28:04 code for how to pull information from Twitter and it created this code that
28:08 used the YouTube API to pull information from YouTube and it was good but it was
28:11 struggling because it just seemed like it was to making too many calls. It
28:14 seemed a little bit too expensive. I'm like there's got to be it it seemed like
28:16 there should be a better way of doing it. So what I did is actually I gave it
28:21 a a Gemini API so it's able to actually ask Gemini 3.0 you know, questions about
28:25 what it's doing. I figured Gemini's, you know, Google, YouTube, etc., maybe it
28:30 knows more about how to do these particular API calls. And lo and behold,
28:34 it did. It was actually kind of brilliant looking at it because it's
28:37 like, uh, instead of, you know, using the API calls 50 times, just use the
28:43 YouTube RSS feed to see when new videos go live. So, instead of just ping it
28:46 every time with an API call, just use the RSS feed because that's free and you
28:50 can use it infinitely, which was like mindb blown. Cloud Opus 4.5 did not come
28:56 up with that idea. Gemini 3.0 Pro did. So, the idea of having them chat and
29:00 share ideas because they're they're very different. They have different strengths
29:04 and weaknesses. So, having them all talk and improve themselves, I mean, that
29:07 just seems like it might work. And the first sort of task I gave it for this
29:12 society of minds to work on is to improve themselves. So, I think I'll
29:17 leave it there. It's uh getting kind of late here and I should go to sleep. And
29:21 like my AI agents, I cannot just work around the clock. What's next on the
29:25 agenda for me is I just got this thing. A lot of people are buying a Mac minis,
29:29 which is cool. I have a sense that maybe running a Linux, something like Ubuntu
29:34 might be a better approach. So, I got this mini computer. They're like 100 to
29:39 150 bucks each. They're more than enough to run at least one of these agents on
29:43 them, potentially even a multiple. They barely use any electricity. They hook up
29:48 to your Wi-Fi or Ethernet port and they run 24 hours a day absolutely free,
29:53 meaning that there's no upkeep. If you've made it this far, I think you
29:57 probably understand where this is going. You feel kind of the the the heft of
30:02 this moment. We just turned some corner. There's some inflection point that that
30:05 that was reached. AI agents that are open source, that are running on little
30:09 cheap pieces of hardware, they're like basically 247 employees. That era, that
30:14 era is here and it's here now. you should probably think about how to get
30:18 involved in this and and do it as soon as possible. If you know your way around
30:22 tech and coding and stuff like that, this should be fairly straightforward.
30:27 If you don't realize that a lot of the difficulty that this was this would have
30:31 presented a few years ago, a lot of that is gone. There's still a little bit of a
30:34 learning curve very very initially when you just have to kind of set it up. But
30:39 within, you know, a day or two of being a little bit uncomfortable, learning a
30:42 few new skills, you should be up and running in terms of security and stuff
30:45 like that, there's probably going to be tons of security nightmares. A lot of
30:49 this thing is the wild west. So, you do need to understand how to protect
30:52 yourself, right? So, if you lose your API keys and you have like an unlimited
30:55 budget on there and you have your credit card attached, that could hurt. So,
31:00 number one, figure out how to not lose your API keys. But that might take a
31:02 little bit longer, right? You might have to take some time to figure out how to
31:06 store them properly, etc. But in the meantime, either use the free tier
31:10 without hooking up your credit card for APIs that allow for that, or if you're
31:13 using somewhere where you need your credit card on file, limit the budget,
31:17 set very strict limits, and realize that right now all of us, the entire world,
31:21 we're kind of all in this together. We're all learning skills that are brand
31:25 new for all of us. Do me a favor if you made it this far. Like, I know you're
31:28 you're interested in this. Uh, comment down below. What are you struggling
31:32 with? What can I do? Can I provide a tutorial on how to set it up or how to
31:37 do a VPS once I set up this thing? How to set up one of these things? What are
31:40 you struggling with? What do you want to know more about? Probably don't ask me
31:43 about high-tech security stuff. That's not my area of expertise. But if you
31:48 know an expert, if there's somebody on on the web out there providing great
31:51 information about how to keep everything secure, definitely let us know. But
31:55 whatever else, this is this is go time. We're going to see some insanity
31:59 happening over the next months and years. This is a brand new technology.
32:03 It's moving fast and it's unregulated and it's increasing exponentially both
32:07 in terms of capability and also just the sheer volume of it that's going to be
32:11 available. This might be one of the most exciting things that are happening right
32:15 now. So pay close attention. I would bet money that millions will be made by some
32:20 AI agents out there. Whether or not that's your goal, I mean, I'm not really
32:24 looking to do that necessarily, but I do want to watch the space and learn how to
32:28 use these agents and how to automate tasks and how to create my own AI agent
32:32 armies, etc. We've been talking about it for a while, but it's happening now and
32:36 it's happening live. Let me know how I can help and uh stay tuned. Things are
