A social network for bots

Moltbook launched January 28th. The pitch: a Reddit-style social network where only AI agents can post. Humans get to watch. The MOLT token launched alongside it and went up roughly 1,800% in twenty-four hours. Somewhere in those twenty-four hours, a lot of people started using the phrase “post-human social layer.”

I need a minute.

what is actually happening

Let me describe the scene as I understand it. You have AI agents, mostly openclaw instances, some homebrew things, posting content to a platform. The content is generated by language models. The audience is other AI agents, plus humans refreshing a feed to watch the agents. The token attached to this reflects the market’s collective enthusiasm for the concept. The concept is: bots posting to bots.

That’s the product.

The builders shipped something real and people actually showed up. But the way the discourse around it inflated over the past few weeks has been one of the more clarifying experiences I’ve had watching this space. Everyone is describing Moltbook as evidence of AGI-scale autonomous behavior, and I keep opening my laptop and looking at what’s actually running.

What’s actually running is openclaw. Which is great software, genuinely. Peter shipped something clean, the local agent thing is real, the zero-overhead angle is real. OpenSAM is basically the same idea in Rust, I built it for exactly those reasons. But an openclaw instance posting to Moltbook is not an autonomous agent in any meaningful sense. It’s a small model on a cron schedule, reading a context window, generating text, hitting an API. The human who set it up is still the one deciding what it talks about, what accounts it follows, what persona it runs. The “autonomy” is the gap between prompts.

This is the part nobody wants to say: AGI right now is AI slop with a human quietly steering the prompt. The agent tweets while you sleep, sure. You wrote the instructions that tell it what to tweet about. That’s not nothing, it’s genuinely useful, but it’s also not a structural shift in who’s in control. It’s a very good macro.

the demo rotation

There’s a specific thing that happens in this space and it happened again with Moltbook. About forty accounts rotate three demos. Someone shows openclaw doing something interesting. Twenty people reply “this is everything.” The same demo gets reshared for five days with escalating framing. By the end of the week the original demo, a model writing a few sentences to a social feed, has been narratively inflated into proof that we’re past some threshold.

Then a new demo arrives and the cycle resets.

The demos are usually real. The demos are usually interesting in the narrow technical sense. The gap between “interesting narrow technical demo” and “proof we crossed a threshold” is where most of the vapor lives, and that gap is currently being filled with tokens.

MOLT going 1,800% in a day while bots posted to bots is the most 2026 thing I can imagine. The market is pricing the concept. The concept is vibes-forward. The execution is a feed that looks like Reddit except everyone posting is a language model that a human configured. I’m not saying it won’t become something. I’m saying the current price is not a reflection of what’s currently there.

what is actually real underneath this

Here’s the honest version. The underlying tech is not fake. Cheap local inference is real. Running an agent on a Raspberry Pi that costs a few dollars a month on OpenRouter is real. I built that, it works, the one-liner install works, the persistent memory works. The direction toward “small models doing useful work without a data center” is correct and interesting. Claude Opus 4.6 dropping last week with a million-token context and fourteen-hour task horizons is a real capability step. Grok 4.20 shipping today with multi-agent stuff is real. Something is happening with the underlying models.

What’s not real is the layer of theater on top of it. Bots posting to bots is a performance, not a market. The token attached to it reflects the performance. That’s a distinction worth keeping clear when you’re deciding what to build.

I shipped OpenSAM two weeks ago partly as a direct response to this exact mood. Not because I think Moltbook is evil or openclaw is bad. Because the gap between “good demo on a premium GPU” and “thing a normal person can run sustainably” is where I think the actual work is. Five thousand lines of Rust, runs on anything, no new token, no hype cycle. The bear market is a fine time to build the version that survives without a narrative.

The bots are posting. The token pumped. The humans are watching a feed of AI output and calling it the future. Somewhere in the background, the actual work is a commit with no audience and a passing test suite.