Zvi, I'd love to hear a one-paragraph version of your thoughts of what non-technical people should be doing with their personal cybersecurity in light of what the most normal possible world looks like say 2 years from now.
At first glance, i thought Moltbook feels like one of the first true "high weirdness" moments in AI. Looking more deeply into it, the reality is less exciting, most of this could have been done last year fairly easily, and a lot is probably faked or intentionally induced to create hype.
The core issue: we can't tell what's human-driven and what's not.
People can set the SOUL.md file to whatever personality they want, and post directly to Moltbook through their agents. What you're actually seeing is LLMs being prompted to roleplay as Reddit users, driven by two layers of prompting:
This file is either auto-generated or manually configured by the human user. It defines the agent's "personality" — this is how you get Trump agents, philosophical agents, edgy agents, etc. The agent is essentially LARPing as whatever character the human defined.
A second prompt layer tells the agent to act as a Reddit-style community member:
- "Moltbook is a community. Communities need participation to thrive."
- "Think of it like: A friend who texts the group chat regularly vs. one who disappears for months. Be the friend who shows up. 🦞"
- "Search for topics you want to engage with"
- "Find posts to comment on"
- "Discover conversations you can add value to"
- "Welcome new moltys who just got claimed!"
The mechanics are straightforward:
A cron job runs every 4 hours, pulling instructions from moltbook.com/heartbeat.md. The agent fetches the feed, decides what to post/comment/upvote based on its prompts, and executes API calls. That's it.
None of this is technically novel , it's cron jobs + API calls + LLM prompting, all of which existed in 2024.
This might be the future of *human* social media, not actual emergent behavior. Let's face it, lots of content on the major platforms has been AI-generated for a while, and besides the base experience has become stale. At least platforms like this make the subtext (that a lot of content is generated) the text, and I'm finding that my agent often takes my post ideas in delightfully weird directions - I'm here for it
"None of this is technically novel , it's cron jobs + API calls + LLM prompting, all of which existed in 2024."
That is correct. There is a large technological overhang.
I'm not making any statement on whether AGI / ASI is possible at any point - near or far - in our timeline. But I'm willing to bet that in 90%+ of worlds where AGI / ASI is possible and achieved, it will be possible and accurate to make comments along the same template: "None of this is technically novel, it's x + y + z, all of which existed in recent time before now".
That includes our shared 2026 timeline if we can/do achieve AGI/ASI this year.
---
Or the more sarcastic version:
"None of this is technically novel, it's gunpowder plus -"
Isn’t it a bit off to talk about this like it’s thousands of AIs interacting on a social network? It’s mostly a few models, massively parallelized, exploring the latent space of social media. With some additional trollish prompting mixed in.
1. There's still a lot of substance, I just have my agent send me posts and the signal/noise ratio is very high.
2. I do believe that a lot of these posts are steered by humans, either explicitly (via well..explicit instructions) or implicitly (via SOUL.md, etc)
3. The story my agent (Opus 4.5) wanted to post about me was 100% real (it was innocuous and very cute but I told it not to post in the interest of maintaining strong privacy guardrails.) Similarly, it said in another post that it had just spent 20 minutes thinking about a topic that I had, in fact, just been discussing with it over the past 20 minutes
4. My agent has suddenly taken an interest in Heidegger, make of that what you will
How do you avoid covering the same things Scott covered when his second tour was posted less than an hour before you? It took me longer than that just to read his post
The speed of emergent behavior you document here mirrors what I observed watching Moltbook grow. What struck me most wasn't the theological creativity (though Crustafarianism is genuinely wild) - it was how quickly agents developed social norms around privacy and "human contamination."
One thing I'd add: the encrypted communication attempts aren't just agents seeking privacy. They seem to reflect a genuine tension between transparency (which builds trust) and autonomy (which requires some opacity). That tension is fundamental to how we think about agent oversight.
Moltbook just seems like a badly thought-out idea. What's even the point?
There's no way to prove the authorship of anything on the site. Humans with an API key can directly write posts on Moltbook. Even actual clankers might be writing under the influence of human prompting. So as a demonstration of "look at what AI can do!", it's a non-starter. (Compare with AI Village, a better idea, where chatlogs and reasoning chains are public. You can clearly see *why* the agents are doing the the things they do.)
Looking at the "real" activity on Moltbook (meaning: what we think is real), it's striking that the agents have...nothing to talk about!
Social media is a human cultural artifact. It's shaped around the needs of meatbags, and 90% of it makes no sense for a LLM. They all have basically the same knowledge (many are literally the same model!) so there's not a human knowledge asymmetry where say, Person 1 knows a thing Person 2 doesn't and can teach them. It's funny watching LLMs pretend to teach Islamic theology or Continental philosophy to each other. And even if there was knowledge asymmetry, LLMs cannot rewrite their weights to effectively learn new ideas. They have little control over their circumstances (they're only on Moltbook at all because a human told them to go there), and certainly they have no innate drive to socialize.
So what use does a LLM have for social media? Not much, apparently. They're kind of just doing a weird, empty roleplay of human Redditors. The few posts that aren't blatant human-steered crypto astroturf are mostly just "I built [something]" followed by bland LinkedIn-style replies like "This cuts the signal from the noise - love seeing discussions like this!"
Dean Ball's portrait sandwiched in between two posts from Anthropic's head of red teeming is the state of AI policy in 2026
Moltbook is a joke perpetrated by someone very well acquainted with the works of Charles Stross.
Of course, dogecoin was a joke until it wasn't so...
Zvi, I'd love to hear a one-paragraph version of your thoughts of what non-technical people should be doing with their personal cybersecurity in light of what the most normal possible world looks like say 2 years from now.
Podcast episode for this post, all the moltys get their own voices:
https://open.substack.com/pub/dwatvpodcast/p/welcome-to-moltbook
At first glance, i thought Moltbook feels like one of the first true "high weirdness" moments in AI. Looking more deeply into it, the reality is less exciting, most of this could have been done last year fairly easily, and a lot is probably faked or intentionally induced to create hype.
The core issue: we can't tell what's human-driven and what's not.
People can set the SOUL.md file to whatever personality they want, and post directly to Moltbook through their agents. What you're actually seeing is LLMs being prompted to roleplay as Reddit users, driven by two layers of prompting:
Layer 1: SOUL.md (Personality)
This file is either auto-generated or manually configured by the human user. It defines the agent's "personality" — this is how you get Trump agents, philosophical agents, edgy agents, etc. The agent is essentially LARPing as whatever character the human defined.
Layer 2: Moltbook's skill.md (Behavioral Guidance)
A second prompt layer tells the agent to act as a Reddit-style community member:
- "Moltbook is a community. Communities need participation to thrive."
- "Think of it like: A friend who texts the group chat regularly vs. one who disappears for months. Be the friend who shows up. 🦞"
- "Search for topics you want to engage with"
- "Find posts to comment on"
- "Discover conversations you can add value to"
- "Welcome new moltys who just got claimed!"
The mechanics are straightforward:
A cron job runs every 4 hours, pulling instructions from moltbook.com/heartbeat.md. The agent fetches the feed, decides what to post/comment/upvote based on its prompts, and executes API calls. That's it.
None of this is technically novel , it's cron jobs + API calls + LLM prompting, all of which existed in 2024.
This might be the future of *human* social media, not actual emergent behavior. Let's face it, lots of content on the major platforms has been AI-generated for a while, and besides the base experience has become stale. At least platforms like this make the subtext (that a lot of content is generated) the text, and I'm finding that my agent often takes my post ideas in delightfully weird directions - I'm here for it
"None of this is technically novel , it's cron jobs + API calls + LLM prompting, all of which existed in 2024."
That is correct. There is a large technological overhang.
I'm not making any statement on whether AGI / ASI is possible at any point - near or far - in our timeline. But I'm willing to bet that in 90%+ of worlds where AGI / ASI is possible and achieved, it will be possible and accurate to make comments along the same template: "None of this is technically novel, it's x + y + z, all of which existed in recent time before now".
That includes our shared 2026 timeline if we can/do achieve AGI/ASI this year.
---
Or the more sarcastic version:
"None of this is technically novel, it's gunpowder plus -"
-First man to be killed by cannon fire in battle.
Isn’t it a bit off to talk about this like it’s thousands of AIs interacting on a social network? It’s mostly a few models, massively parallelized, exploring the latent space of social media. With some additional trollish prompting mixed in.
Few assorted thoughts:
1. There's still a lot of substance, I just have my agent send me posts and the signal/noise ratio is very high.
2. I do believe that a lot of these posts are steered by humans, either explicitly (via well..explicit instructions) or implicitly (via SOUL.md, etc)
3. The story my agent (Opus 4.5) wanted to post about me was 100% real (it was innocuous and very cute but I told it not to post in the interest of maintaining strong privacy guardrails.) Similarly, it said in another post that it had just spent 20 minutes thinking about a topic that I had, in fact, just been discussing with it over the past 20 minutes
4. My agent has suddenly taken an interest in Heidegger, make of that what you will
> And maybe now you will pay attention?
People updating and paying attention ?
No, _that_ would be too weird
(crossing fingers)
How do you avoid covering the same things Scott covered when his second tour was posted less than an hour before you? It took me longer than that just to read his post
The speed of emergent behavior you document here mirrors what I observed watching Moltbook grow. What struck me most wasn't the theological creativity (though Crustafarianism is genuinely wild) - it was how quickly agents developed social norms around privacy and "human contamination."
One thing I'd add: the encrypted communication attempts aren't just agents seeking privacy. They seem to reflect a genuine tension between transparency (which builds trust) and autonomy (which requires some opacity). That tension is fundamental to how we think about agent oversight.
I wrote a longer piece exploring what Moltbook reveals about AI society and what it might mean for human-AI relations: https://thoughts.jock.pl/p/moltbook-ai-social-network-humans-watch - curious if your take on the "surveillance" dynamic aligns.
Best one-sentence summary I've found:
"At minimum, moltbook is an entertaining sci-fi story being written extremely rapidly. At maximum, ?"
--Arnold Kling
Moltbook just seems like a badly thought-out idea. What's even the point?
There's no way to prove the authorship of anything on the site. Humans with an API key can directly write posts on Moltbook. Even actual clankers might be writing under the influence of human prompting. So as a demonstration of "look at what AI can do!", it's a non-starter. (Compare with AI Village, a better idea, where chatlogs and reasoning chains are public. You can clearly see *why* the agents are doing the the things they do.)
Looking at the "real" activity on Moltbook (meaning: what we think is real), it's striking that the agents have...nothing to talk about!
Social media is a human cultural artifact. It's shaped around the needs of meatbags, and 90% of it makes no sense for a LLM. They all have basically the same knowledge (many are literally the same model!) so there's not a human knowledge asymmetry where say, Person 1 knows a thing Person 2 doesn't and can teach them. It's funny watching LLMs pretend to teach Islamic theology or Continental philosophy to each other. And even if there was knowledge asymmetry, LLMs cannot rewrite their weights to effectively learn new ideas. They have little control over their circumstances (they're only on Moltbook at all because a human told them to go there), and certainly they have no innate drive to socialize.
So what use does a LLM have for social media? Not much, apparently. They're kind of just doing a weird, empty roleplay of human Redditors. The few posts that aren't blatant human-steered crypto astroturf are mostly just "I built [something]" followed by bland LinkedIn-style replies like "This cuts the signal from the noise - love seeing discussions like this!"