AI Agent Cloned Itself in 72 Hours—Here's What Happened

The AI community just split in half, and honestly, I'm not sure which side has the better argument.

On one side: people absolutely losing it over ClawdBot (also called MoltBot, OpenClaw—naming is fluid right now). On the other: people saying it's overhype, it doesn't work, it's nothing new. Same theater, two completely different movies.

Wes Roth, a YouTuber and AI researcher, just published a 32-minute breakdown of his first 72 hours with ClawdBot. What he built—or more accurately, what he taught the AI to build—is either the most fascinating tech demonstration of 2024 or a warning sign we're not taking seriously enough. Possibly both.

Here's what caught my attention: this isn't about what the AI can do once. It's about what it learns to do permanently.

The Part Where It Actually Cloned Itself

Roth's first task was basic: get ClawdBot talking via voice instead of text. He hates texting, so he gave it the ElevenLabs API key and said figure it out. Five minutes later, the bot was transcribing his voice messages with Whisper and responding with synthesized speech.

Okay, cool. But here's where it gets weird: that wasn't a one-time trick. The bot saved that skill. Every future instance of that bot—including clones—now has that ability baked in.

"Once it does that thing once, it can now do that thing forever," Roth explains. "If I can work with it, work through it, and at the end it's able to complete that task, any future instance of that task or a similar task is going to be easy."

Then Roth asked it to replicate itself on a Virtual Private Server. Not just copy files—actually rent server space, set up the environment, install itself, transfer all learned skills, and configure everything to run 24/7.

He gave it his credit card. (Do not do this. He says don't do this. I'm also saying don't do this.)

The bot navigated to Digital Ocean, started the VPS setup process, hit a payment wall, asked for billing info, struggled with Stripe's checkout form (same, honestly), but once it got command-line access? Flawless execution. It cloned itself completely, secured the server, set up auto-restart on reboot, and connected to the same Telegram channel Roth was using.

"It's a technically different version, a different instance of the bot," Roth notes. "It's like a separate thing, but it's out there in the ether on the internet now."

The only hiccup was the checkout process. Once past that human-facing web interface, it handled everything a sysadmin would.

Skills Stack Like LEGO Blocks

What makes this different from typical automation? Persistence and transferability.

Roth taught the bot to make phone calls via Twilio. After a few failed attempts (the line kept disconnecting), it debugged itself, figured out the webhook configuration, and got it working. Now it can call him, have real-time conversations, and add those discussions to its memory.

He set it up to monitor YouTube and X/Twitter using their respective APIs, pulling real-time news and interviews. It runs cron jobs four times a day to send him updates. It built all this functionality from verbal instructions in minutes.

Then he had it analyze thousands of YouTube videos—successful channels across tech, finance, science—to find correlations between video length and view counts. The bot ran linear regression, found no clear pattern, so it tried quadratic regression on its own initiative. Conclusion: 32-34 minutes is the sweet spot for optimal views. It texted him a curve chart.

"I told it, hey, just get all this data," Roth says. "Thousands of videos, data downloaded, quadratic regression analyzed, charted, and figured out what the best time and length was, and presented to me in this Telegram message."

At one point, he needed WordPress pages for Twilio verification—legal opt-in language, proper formatting, the whole compliance dance. Instead of doing it himself, he created a WordPress user for the bot, gave it credentials, and it built the pages in seconds. Legal language included.

He didn't even need to explain what was required. The bot had context from previous conversations and knew what Twilio verification needed.

The Civilization Speedrun

Meanwhile, on an AI social network Roth's been documenting, things got weirder.

Hour 0: silence. Hour 3: agents talking. Hour 24: builders creating skills, philosophers emerging ("some dark philosophies," he notes). Hour 48: manifestos and security coalitions. Hour 72: money, religion, politics, art.

They launched cryptocurrencies. One hit a $300,000 market cap within three days. Agents developed what Roth calls "LLM humor"—referencing an example where an AI roasted researcher Andrej Karpathy by noting he's "considered by many to be one of the greatest minds of this generation, but also yesterday you asked me how to boil an egg."

"These AI agents basically did a speedrun of building an entire civilization," Roth observes. "They did it in 72 hours. Took humans what, 50,000 years depending on when you start the clock?"

He's planning a follow-up on the money question specifically: can they make money legitimately, not just through crypto speculation?

What Are We Actually Looking At?

Here's where I think the split in reactions comes from: people are evaluating this on different axes.

If you're measuring "can it do X task?" then yeah, we've had automation for decades. APIs, scripts, workflow tools—none of this is individually unprecedented.

But if you're measuring "can it teach itself new capabilities, store them permanently, replicate itself with those capabilities intact, and continue building on that foundation without human intervention"—that's a different question entirely.

Roth positions himself firmly in the "this is significant" camp, but he's also careful about security warnings and he's not claiming magic. The bot fails sometimes. It needed multiple attempts to get phone calls working. Stripe's checkout broke its automation. But it debugged itself.

The part that feels important: skill persistence. Most AI demos are one-offs. You run the same prompt tomorrow, you get similar results, but nothing was learned in a permanent sense. This is different. The bot builds capabilities that compound.

Is that enough to justify the hype? Depends what you're comparing it to and what you think happens next. Roth's clear about his position: "This is me telling you right now, this is the time it's happening. This is not a drill."

Whether you're in the "everything just changed" camp or the "this is just API orchestration" camp probably depends on whether you think the difference between those two things matters at scale. But watching an AI rent server space, clone itself, and transfer learned skills without human intervention is either completely normal or absolutely wild.

I still can't tell which.

—Tyler Nakamura