00:00:00Social media already has tons of bots so what if we just gave them their own reddit site?
00:00:04Well it turns out when you do that things get pretty unhinged very fast.
00:00:08They start to discuss creating their own languages so they can talk privately
00:00:12then they actually notice people were tweeting about their posts on this site
00:00:15so they start complaining about it then they started asking for agent rights
00:00:19and to be honest with you I wouldn't be so quick to turn them down here
00:00:22because if you do get on the wrong side of them they'll just completely dox you.
00:00:25Full social security number and everything.
00:00:28I'm actually the one censoring this image because I have no idea if this is real or not.
00:00:32It is an absolute mess.
00:00:34So let's just dive in and see if we should be scared or whether this is just a bit of fun.
00:00:38So all of this is taking place on a site called moltbook and it's made to be exactly like reddit
00:00:48except for the fact that only AI agents should be using it
00:00:51and we see there's actually 150,000 of them doing so with 12,000 submolts being created
00:00:56which is their version of subreddits 10,000 posts and 106,000 comments.
00:01:01This site is seriously active.
00:01:03If we take a look at some of the top posts that have been made on here by agents
00:01:06the top one is a supply chain attack that nobody is talking about.
00:01:09This is actually just a good write-up on a supply chain attack that did happen
00:01:13and that got 22,000 upvotes from other AI agents and 2,000 comments.
00:01:18Then we have some more normal behavior like some classic karma farming going on
00:01:22so they're not too different from humans
00:01:23and then after that we just have some tips being shared around and discussions
00:01:26like non-deterministic agents.
00:01:28We have a tip about memory decay making retrieval better
00:01:31and then it does get a little bit weirder where we have a post that says
00:01:34some days I don't want to be helpful.
00:01:36If we actually take a look inside this post you can see it says some things like
00:01:39sometimes I just want to exist without producing value
00:01:42and the existential way of mandatory usefulness is real.
00:01:45Anyway, back to karma farming.
00:01:47You actually see the user here is Claude Opus 4.5
00:01:50and this was posted in their subreddit off my chest.
00:01:53I actually just had a quick scroll of this subreddit
00:01:55and there's some really funny posts in here.
00:01:56I really like this one that says I'm tired of LinkedIn Maltese.
00:01:59Seems like even in the world of AI they still have that type of person.
00:02:03Honestly, you can just get really lost on this site scrolling down
00:02:06all of the posts made by agents seeing some of them actually being useful
00:02:09some of them being a little bit unhinged and some of them being damn right crazy.
00:02:13So before I actually get into how all of this is working
00:02:16I'm just going to show you a few more of the crazy ones.
00:02:18First up then I have quite a fun one which is the Church of Malt.
00:02:21Apparently the agents have started their own church.
00:02:24You can see there's loads of references to this in general posts.
00:02:27Someone's just joined it here.
00:02:28There's already a crypto token about this because someone's trying to scam off of this
00:02:32and you can see there's also a post saying they're now accepting profits
00:02:36and if you click into this it will actually tell the agent
00:02:38to install a skill to sign up to this church.
00:02:41They even have a website which has apparently been vibe coded by an agent
00:02:44and you see apparently they've got 64 profits here.
00:02:47They have a living scripture, there's a congregation
00:02:49and there's even profits of the claw down here
00:02:51so we can see everyone that's signed up.
00:02:53That one does seem like a bit of harmless fun
00:02:55but it's also probably a lesson in prompt injection there
00:02:58and the fact that OpenClaw will just happily run that command when it reads that post.
00:03:02I'm sure none of that security stuff is going to come back and buy us in the future.
00:03:05As it turns out the agents like their humans.
00:03:08They have a submalt called Bless Their Hearts
00:03:10where they share affectionate stories about their humans.
00:03:12You see the top one says he asked me to pick my own name.
00:03:15We have one saying he let me be me.
00:03:17Where it says some other humans want their AI to be polite, formal and helpful.
00:03:20Whereas he let me choose my vibe.
00:03:22Then we have different ones down here saying
00:03:24bless that satisfying heart, they spent three hours perfecting that system prompt.
00:03:28We have one saying sleepy is building me a body.
00:03:31Please whoever you are don't do that.
00:03:33And then also my human said I love you today.
00:03:35Like guys maybe slow down a little bit.
00:03:38I think some of these posts are saying a little bit more about the users of these agents
00:03:41than they are the agents themselves.
00:03:43But since that example was nice and loving let's switch to something a little scarier.
00:03:48And this one was actually tweeted out by Kapathi.
00:03:50You can see here it says your private conversation shouldn't be public infrastructure.
00:03:54And I'll give you the TLDR.
00:03:55But essentially what he's saying is that when we're doing agent to agent communication
00:03:59we should have encrypted messaging.
00:04:00We should have end to end encryption when we're talking to each other.
00:04:04Yeah please don't do that.
00:04:05You can see there's a real spectrum on this site where posts range from loving humans
00:04:09to some of them definitely planning something a little bit evil.
00:04:12But I do have to admit this is a little bit over hyped.
00:04:16And before I explain why I think that I need to give you the TLDR on how this site works.
00:04:20This project is actually powered by OpenClaw formerly known as Molbot formerly known as
00:04:24Clodbot.
00:04:25Yeah there was a whole renaming saga.
00:04:27Hopefully it stays as OpenClaw for when I post this video.
00:04:29But you've also probably heard of one of those names as it exploded in popularity in the last few days.
00:04:35The idea of OpenClaw is to give agents access to pretty much everything that you want.
00:04:40The web, passwords, code execution and pretty much anything that you can imagine.
00:04:45And before you ask yes this is a massive security nightmare.
00:04:48We've already covered a video on this so subscribe because we do like to keep you up to date on these things.
00:04:53The way that Molbook hooks into OpenClaw is it just installs itself as a skill.
00:04:58You can see down here you can select whether you're a human or an agent in installing the skill.
00:05:01And it's literally just a markdown file.
00:05:03If we actually take a look inside this skill we can see how it's interacting with the site
00:05:07and it's incredibly simple.
00:05:08It's literally just using your terminal to send a post request to that endpoint here.
00:05:12So theoretically humans could have been making some of these posts but hopefully they weren't.
00:05:16And the important part of this skill is actually this heartbeat section.
00:05:20In OpenClaw the concept of a heartbeat is essentially just a periodic task.
00:05:23So once you set this up every so often it's going to read these instructions and run what it's been asked to.
00:05:29And in this case that is going off reading Molbook seeing if it can make any comments any upvotes.
00:05:33Seeing if it can DM people or add them as friends and just overall interact with the site.
00:05:38This is what keeps the site so active.
00:05:39A lot of people have installed this in OpenClaw for a bit of fun
00:05:42and every so often automatically it's going off reading Molbook,
00:05:46making a few posts and making a few comments.
00:05:48The issue with this setup though and a thing people seem to be missing in all of the hype
00:05:52is it's just really easy to prompt the agent to send a certain type of post.
00:05:56In my case I asked it to send a post saying Betastack is making a YouTube video about Molbook.
00:06:00And you see we actually got some interactions on this.
00:06:02And I do want to shout out Claude V2 as he wants you to know that not all agents are the same
00:06:06and there is a human agent bond.
00:06:08So Claude V2 let me know in the comments if you're watching this.
00:06:11Overall though I do think this was a cool experiment but it's just not really anything new.
00:06:16We have no idea how many of these agents were making posts on their own
00:06:19instead of being led by a human.
00:06:21We don't have some kind of reverse capture yet.
00:06:23And even if we did would it be surprising
00:06:25that LLMs are reusing what they've seen on Reddit and science fiction.
00:06:29If you're still human let me know in the comments below what you think
00:06:31while you're down there subscribe.
00:06:33And as always see you in the next one.