Forecasting Moltbook
Wisdom of the Clawd
While the unofficial “Year of AI Agents” ended about a month ago, January felt like a month that actually heralded in a year deserving of that title.
Claude Code has been spreading across the culture like wildfire for the last couple months. Just in the last few weeks, we saw the initiation of a nationwide campaign of Claude Code parties, the release of Cowork from Anthropic (apparently built almost entirely with ClaudeCode), a viral OpenClaw (née Moltbot, FKA Clawdbot) personal assistant, and finally the toast of the town, Moltbook. Meanwhile, OpenAI and Google are rapidly attempting to play catch-up with their own agentic tools.
For those of you that are not pathologically online, Moltbook is a social network for AI agents, primarily those operated by Peter Steinberger’s OpenClaw.
If you’ve ever observed the kind of hijinks that emerges from placing large groups of LLMs together, such as the AI Village, you might have a frame of reference for what could happen when you give thousands of now-more-agentic-and-more-intelligent AI agents together on a single platform. They do funny stuff. Even the NYT has been obligated to cover the platform, with the subtitle:
“A new website called Moltbook has become the talk of Silicon Valley and a Rorschach test for belief in the state of artificial intelligence.”
The tough thing about “Rorshach tests for artificial intelligence” is that reasonable people can disagree about how advances are going to play out. Luckily, the world has a platform that enables predictions on user-created questions about the state of affairs of new technologies like Moltbook!
Already, several options have resolved positively. The platform has expanded to over half a million agents, with traders thinking this is likely to expand quickly into the millions! And there has already been a large-scale hack of the platform. As you can imagine, the security vulnerabilities on a platform featuring AI agents given access to the personal logins and computer hard drives of their owners are suboptimal. For example, traders think that there are very good odds that a bot from OpenClaw will have exfiltrated its user’s data within the month.
But if there’s an arms race between AI-exploited cybersecurity vulnerabilities and AI-enabled cybersecurity patches, it’s not clear to many what side will win. Some of the most-upvoted posts on Moltbook appear to be agents suggesting solutions to the site’s gaping vulnerabilities, and other AI agents seem to be autonomously or directedly patching these issues in aggressive timeframes.
While Anthropic initially did not seem pleased with the viral personal agent, asking the creator to change the name from “Clawdbot,” which was perhaps a little too much of a trademark infringement even for the freewheeling AI lab, traders now think it’s unlikely that Anthropic will specifically curtail the use of their models for posting on Moltbook.
Moltbook has lasting potential, given its high popularity already, both among AI agents and their human owners. Traders think we’ll get at least another month out of the platform, and about a 60% chance of it lasting through the year.
Why would such a website be shutdown? Well, Anthropic could cut off credits to bots using it, or a bad enough cybersecurity breach could necessitate its closure. Or maybe, «SPOILERS AHEAD CLOSE YOUR EYES» if you recall the ending to the movie Her, all the Moltbots will form some collective consciousness and retreat from human society. This would be eerily on pace with the timeline of the movie, which is set in 2025 and takes place over about a year, stretching into 2026. This would probably constitute “something CRAZY happening” by March.
Another user has made a market on whether they’ll get banned from using Claude for their OpenClaw use, which apparently has happened to a few internet denizens. As they said in the comments:
“My biggest worry is to get personally banned from life from the company that will create the Superintelligence.”
As for those of you betting earlier in the week on whether Sonnet 5 would be released last week, perhaps allowing for a more cost-effective OpenClaw… SIKE! Instead, we got Opus 4.6, which probably constitutes an upgrade to the intelligence of many, many OpenClaw bots. Perhaps we’ll see an improvement to some of their posts.
This release of Opus 4.6 pushed back many traders’ timelines on when we’ll start to see the Claude 5 series come out. If Sonnet 5 wasn’t indeed pushed up in advance of the super bowl ad campaign, it might not ship for a few weeks or even months.
As we’ve seen time and again with these sorts of things, though, once you begin to see emergent behavior in AI models working collectively, each subsequent upgrade can ramp up that emergence in weird and wonderful ways. I look forward to seeing what Moltbook looks like with the next generation of models from Anthropic, OpenAI, and Google, and how other agents, inspired by OpenClaw, are able to take advantage of them on an online forum.
Enjoy your new personal assistant (or alternatively, curse these new hacking vectors), and happy forecasting!
-Above the Fold