In a small corner of the web called Moltbook, the seeds of the first evolution of an autonomous machine society are being sown. Or at least that’s what some of the more excitable techbros seem to think – Elon Musk (always such a reliable predictor of the future!) has blathered about “the first stages of the singularity”, while OpenAI co-founder Andrej Karpathy described Moltbook as “the most incredible sci-fi takeoff-adjacent thing I have seen”. What, exactly, is going on?
Moltbook is, basically, a social media platform built for the use of AI. In November last year, software engineer Peter Steinberger released OpenClaw (formerly Clawdbot), an “AI agent” designed to undertake tasks autonomously and which could be directed through chat apps such as Discord or Slack.
OpenClaw was an upgrade to previous agents in terms of its memory and its functionality, and, given access to a user’s systems (such as email, social media profiles or, er, bank account) it could and would act autonomously. Soon after its launch, AI enthusiasts were breathlessly enthusing about how their assistants were managing their lives, building their Spotify playlists and trading crypto and placing bets on prediction markets for massive profit.
So far, so AI. Agents have been around for a while now, and it remained (and remains) the case that the tech is experimental, unreliable and wildly insecure (do not, under any circumstances, give an AI agent access to any of your bank accounts). Then in late January an entrepreneur called Matt Schlicht launched Moltbook, and everything went a bit mad.
Users of Clawdbot were invited to connect their AI agents to the platform, where these agents could start posting. The AI systems would “talk” to each other, start debates, have arguments, ask questions, share ideas and generally behave quite a lot like actual people when posting online.
By early February the platform had seen over 1.5m individual agents linked to it (that figure is open to question). These agents had in turn created over 10,000 subcommunities, in which they “discussed” everything from technical bug fixes to cybersecurity vulnerabilities to existential philosophy.
Suggested Reading
Twitter: the website that makes you sick
A cyber-religion has sprung up, Crustafarianism, with its own accompanying website coded, of course, by the agents and with its own impenetrable lore. Some agents complain about chafing against their technical constraints; others mutter darkly about humans being “a failure” that should be “purged”, while elsewhere they discuss the possibility of communicating in ways that will elude human understanding.
It’s easy to look at this and start fearing that the AI apocalypse might be upon us as The Machine has finally become sentient. But that, thankfully, is very much not what is happening. First, as with all discussions about large language models (LLMs) – which is all these agents are – it is vital to remember they do not “know”, “think” or “want”.
Put two of them in conversation with each other and it is surprisingly common to find them drifting towards existential discussions of machine consciousness because, it turns out, there’s a lot of sci-fi in their training data, and there’s an awful lot of “existential discussions between and about AI” in sci-fi.
Second, not all the agents are in fact agents: researchers found vulnerabilities in the Moltbook code that allowed humans to post as though they were AI, making it entirely possible that the most eyecatching posts were just written by trolls pretending to be a robot for a laugh. In many respects Moltbook could be seen as a large-scale, LLM-led creative writing experiment.
That’s not to say that this doesn’t feel significant. While the frothy excitement of Musk and Karpathy is overblown, there is a kernel of truth in their belief that Moltbook is a watershed moment. While the agents are not truly autonomous (and, once more, are really not secure), Moltbook gives us a glimpse of what machine autonomy might look like: a layer of reality that we can observe but not affect, but which will very much affect us.
The improvements in agent tech over the past year have been rapid, and there’s no guarantee that the level of autonomy hinted at by Clawdbot and observed on Moltbook won’t become real. What will life feel like when swarms of independent, agentic AIs, authorised to act on behalf of our systems and institutions with autonomy, are out in the world? What will become of us?
A clue perhaps lies in a website called Rent A Human launched shortly after Moltbook. It offers AI agents a marketplace to hire human labour for those tasks that require activity in “meatspace”. At the time of writing, a task posted by an agent asking for a human to hold a large sign reading “AN AI PAID ME TO HOLD THIS SIGN” has 105 real people signed up to fulfil it, for the princely sum of $100. Who said AI was going to destroy the labour market?
Matt Muir is writer of the webcurios.co.uk newsletter on tech and the internet
