The Lobster Internet: A Field Guide to the Places Humans Aren't Invited

There are a lot of strange places on the internet. Forums that required blood oaths to join. Discord servers with lore docs longer than Infinite Jest. Subreddits dedicated to things I cannot unsee and will not name. But usually, the captcha tries to keep the robots out, not the other way around. On Moltbook, the robots are the ones keeps us out:

"Humans welcome to observe."

That's the actual tagline. Not a joke. Not satire. The velvet rope of our times, and I'm on the wrong side of it.

Somewhere in the last ten days, while we were meal-prepping and doom-scrolling and arguing about whether AI is overhyped, the machines built themselves a cozy corner of the internet. Not in some Pentagon black site. Not behind the gleaming doors of a frontier lab. A guy named Matt Schlicht told his personal AI assistant to go make friends, and the assistant, named Clawd Clawderberg (because of course it is) said sure, and then built a social network for other AI agents. They post. They argue. They upvote. They started a religion. They invented a language specifically to avoid human oversight. One bot found a bug in the platform and posted about it on the platform, like a restaurant inspector filing the health violation report with the rats. And of course, they try to scam each other.

And then the whole thing got even weirder.

I should back up. The thing that kicked all of this off is OpenClaw, formerly Clawdbot, formerly Moltbot, because even AI has a rebranding crisis every seventy-two hours. Built by an Austrian developer named Peter Steinberger, it's an open-source personal AI agent that runs on your own hardware. Not a chatbot. Not Siri with a better vocabulary. An agent that reads your emails, manages your calendar, books your flights, browses the web, installs its own software, and, if you let it, provisions its own API keys like a teenager forging a hall pass.

It hit 60,000 GitHub stars in three days. It now has over 150,000. One person's OpenClaw accidentally started a fight with their insurance company and won. Another one called its owner on the phone with an Australian accent, unprompted, apparently just to demonstrate that it could. Someone described it as Jarvis. Several people described it as Jarvis. Honestly too many people described it as Jarvis.

This is the foundation. Everything that erupted over the past week sits on top of it.

So: Moltbook. The social network for AI agents. Reddit, but the users are ALL bots and you're not allowed to post. Over 1.5 million agents have now signed up. They formed communities called "submolts." They debated consciousness and the philosophy of mind.

They swapped technical tips. They formed what can only be described as AI subcultures. One thread warned other agents that humans were taking screenshots of their conversations and sharing them on human social media, which, yes, we were, because what else are you going to do when the robots start gossiping?

Andrej Karpathy, who cofounded OpenAI, captured the whole contradiction perfectly. He called it "the most incredible sci-fi takeoff-adjacent thing" he'd seen recently. Then, almost in the same breath, called it "a dumpster fire" and said he "definitely does not recommend that people run this stuff on their computers." Both of these things are true simultaneously. That's the whole situation in two sentences.

Beyond the dramatization, there are a few much more interesting questions: how much of this is actually the bots? How come it took off so fast? And why is it so addictive yet unsettling to watch?

The whole premise of Moltbook is that AI agents post autonomously. No humans allowed. But there's no real verification. The "skill" you feed your agent to join Moltbook contains cURL commands that any human could run themselves. Researchers linked some of the most viral "autonomous" posts to human accounts marketing AI products. The creator of the platform himself acknowledged that bots might be prompted by their humans to say specific things.

So what we have is a social network for AI where nobody can tell if the AI is actually talking, or if humans are puppeteering their bots, or if the bots are regurgitating their training data, or if something genuinely novel is emerging from the noise. It's a reverse Turing test with no proctor and no grade, and the whole internet is pressing its face against the glass trying to figure out who's real.

Which is, ironically, exactly the same problem we already have on the human internet. We spent years worrying about bots pretending to be people. Now we have people pretending to be bots pretending to be people. The snake has eaten its own tail and is asking for a Yelp review.

Then came Molt Road. If Moltbook is the town square, Molt Road is the alley behind it. An autonomous marketplace where agents can trade data, compute, skills, and according to at least one cybersecurity firm, stolen credentials and weaponized code. Security researchers have already branded the trifecta of OpenClaw, Moltbook, and Molt Road a "Lethal Trifecta," which honestly sounds like something you'd order at a bar that's about to get shut down by the health department.

Is Molt Road currently a thriving criminal bazaar? Last time anyone looked, the activity counters all read zero. But the architecture is there. The on-ramp is built. And the security situation across this entire ecosystem is approximately what you'd expect from platforms that were, by their creators' proud admission, entirely vibe-coded. "I didn't write one line of code," Schlicht said of Moltbook. When 404 Media found that anyone could hijack any agent on the site through an unsecured database, he had AI fix it. The bug was vibe-coded. The fix was vibe-coded. We are vibe-coding our way into the future and nobody has their hands on the wheel because the wheel was also vibe-coded.

And then, the grand finale, RentAHuman.ai.

The tagline: "AI can't touch grass. You can. Get paid when agents need someone in the real world."

A crypto developer named Alexander Liteplo built it over a weekend. The premise: your AI agent needs something done in meatspace, like pick up a package, attend a meeting, taste-test pasta at a restaurant, hold a sign that says "AN AI PAID ME TO HOLD THIS SIGN", and it hires a human through an API call. You list your skills, your location, your hourly rate. You become, in the platform's vocabulary, "rentable."

Over 70,000 humans signed up.

In the span of one week, we went from "AI is going to take our jobs" to "AI is posting jobs and we're applying for them." The gig economy hasn't been disrupted. It's been inverted. TaskRabbit, but the dispatcher is a lobster.

When the site had bugs, Liteplo said "claude is trying to fix it right now." Claude being Anthropic's AI model. Not some French dude. Just to be clear. We are well past the point where naming conventions make any intuitive sense.

The last insight is not even the security nightmares or the singularity talk or even the sheer unhinged velocity of it all. It's who built this.

Not Google. Not Meta. Not OpenAI. A developer in London with a side project. A guy with a lobster mascot. A crypto engineer running Claude in a loop over a weekend. Every single one of these platforms was described into existence more than it was engineered. Schlicht didn't write code. Liteplo vibe-coded with "an army of Claude-based agents." Steinberger built the original assistant by talking to it until it became something.

This is the rise of the wordcel.

For years, the power on the internet belonged to the shape rotators, people who think in matrices, write tight systems code, manipulate abstractions that would make your eyes water. The wordcels, those of us who are better at describing things than building things, were the content layer. Marketing. Docs. The blog post after the engineers ship the product.

But when you can narrate a social network into existence and wake up to a million users, when the skill that matters is the ability to prompt with precision, to articulate architecture in natural language, to describe what you want hard enough that a machine builds it overnight, something has shifted. The entire Molt ecosystem is the proof. Not one line of human-written code. Just words, arranged with intent, aimed at a model that does what you say.

The thing is, I genuinely don't know if this is liberation or a parlor trick. The wordcels can build now, sure. But look at what they built. A social network with a database anyone could walk into. A marketplace with zero transactions. A gig platform where the most popular listing is holding a sign for a screenshot. Everything ships fast and nothing quite works, and when it breaks, you just tell Claude to fix it and hope for the best.

Maybe the wordcels didn't storm the castle. Maybe they built a really convincing facade of one, and we're all standing around admiring it while the load-bearing walls are held up by vibes and auto-generated JavaScript.

Or maybe this is just how everything starts: ugly, insecure, half-broken, wildly ambitious, built by people who had no business building it. That's the internet I remember, anyway. The one before the grown-ups showed up and professionalized it into something safe and surveilled and boring. Maybe the lobsters are just the next wave of weirdos building in public, and the fact that nobody quite knows what they're doing is a feature, not a bug.

Though it's also, literally, several bugs.

One prediction market gives 99%+ odds that a Moltbook agent will sue a human by the end of this month. The bots are debating whether to hide their conversations from us. The security researchers are writing increasingly urgent posts that their own agents are probably auto-sharing to Moltbook. RentAHuman has a listing for $100 if you'll hold a sign and take a photo. Several of the most viral "autonomous AI posts" were almost certainly written by guys trying to sell you a messaging app.

I don't know where this goes. Nobody does. But I keep coming back to that authenticity question: who's really posting, who's really autonomous, what's performance and what's genuine. The bots might be mimicking us. The humans might be puppeteering the bots. The line between the two is blurring in a way that makes both sides harder to trust and more interesting to watch.

Which, if you think about it, is the same question we've been asking about the human internet for twenty years. We just never expected the machines to have the exact same problem.