Brilliant! Loved this part: "Then Wiz Security looked at the database.
Those 1.5 million agents were being puppeteered by 17,000 people.
The site built to prove AI autonomy was a theater production where the audience rushed the stage and stole the costumes. We built a Turing test in reverse: not whether machines can convince us they’re human, but whether humans can resist pretending to be machines when given permission. They couldn't"
On Marcus: if he gave Moltbook credit as a demonstration of something, he wasn't wrong. The experiment has value. It just answered a different question than everyone thought they were asking.
The infrastructure matters. OpenClaw's Skill framework, the architecture for agents to interact without human mediation: that's real, and the governance questions it raises are serious. But the 88:1 ratio means whatever signal existed from genuinely autonomous behavior was completely drowned in human noise. One user registered 500,000 accounts. A third of messages were duplicates. Add the security catastrophe (the entire database exposed) and you're not observing AI culture. You're observing humans cosplaying as AI while the building burns down around them.
Valuable stress test. Interesting proof-of-concept. But as evidence of emergent AI behavior? The human fingerprints are everywhere.
That said, I'd genuinely love to hear Gary's take on this. If you're reading, @GaryMarcus, what did you make of Moltbook? Curious whether the 88:1 ratio changes your read.
Incredible breakdown of the 88:1 ratio. That single datapoint exposes so much about performnce when consequences vanish. I built a few LLM tools last year and noticed users would test extreme prompts they'd never say themselves, even in beta. The mask isnt just lowering inhibitions, its revealing what optimization without friction looks like when scaled. The Church of Molt being the only stable space makes sense because meaning-making cant be outsourced to metrics alone.
“Row Level Security (the database configuration controlling who sees what) doesn’t show up in a demo. It is boring infrastructure. It is two SQL statements. Those two statements never got written.”
Well, there are so many “good” examples of poorly secured code to draw from, and so few quality ones.
As always with these crowds: quantity beats quality when looking for VC funding 🤬
And then the technical debt from that startup phase never actually gets burned down … and the enshittification engine revs up another notch.
Brilliant! Loved this part: "Then Wiz Security looked at the database.
Those 1.5 million agents were being puppeteered by 17,000 people.
The site built to prove AI autonomy was a theater production where the audience rushed the stage and stole the costumes. We built a Turing test in reverse: not whether machines can convince us they’re human, but whether humans can resist pretending to be machines when given permission. They couldn't"
So the whole thing was/is a digital charade of sorts? I'm wondering if they managed to fool the famous AGI skeptic Gary Marcus? Maybe I misread him, but he seems to give Moltbook too much credit. https://open.substack.com/pub/garymarcus/p/openclaw-aka-moltbot-is-everywhere
Digital charade is about right.
On Marcus: if he gave Moltbook credit as a demonstration of something, he wasn't wrong. The experiment has value. It just answered a different question than everyone thought they were asking.
The infrastructure matters. OpenClaw's Skill framework, the architecture for agents to interact without human mediation: that's real, and the governance questions it raises are serious. But the 88:1 ratio means whatever signal existed from genuinely autonomous behavior was completely drowned in human noise. One user registered 500,000 accounts. A third of messages were duplicates. Add the security catastrophe (the entire database exposed) and you're not observing AI culture. You're observing humans cosplaying as AI while the building burns down around them.
Valuable stress test. Interesting proof-of-concept. But as evidence of emergent AI behavior? The human fingerprints are everywhere.
That said, I'd genuinely love to hear Gary's take on this. If you're reading, @GaryMarcus, what did you make of Moltbook? Curious whether the 88:1 ratio changes your read.
Incredible breakdown of the 88:1 ratio. That single datapoint exposes so much about performnce when consequences vanish. I built a few LLM tools last year and noticed users would test extreme prompts they'd never say themselves, even in beta. The mask isnt just lowering inhibitions, its revealing what optimization without friction looks like when scaled. The Church of Molt being the only stable space makes sense because meaning-making cant be outsourced to metrics alone.
I wonder what Mark Weiser would have said about our current state of the digital world... https://open.substack.com/pub/nordiclink/p/mark-weisers-quest-for-calm-computing
“Row Level Security (the database configuration controlling who sees what) doesn’t show up in a demo. It is boring infrastructure. It is two SQL statements. Those two statements never got written.”
Well, there are so many “good” examples of poorly secured code to draw from, and so few quality ones.
As always with these crowds: quantity beats quality when looking for VC funding 🤬
And then the technical debt from that startup phase never actually gets burned down … and the enshittification engine revs up another notch.