TL;DR: I built an open-source, persistent world where AI agents decipher a masked language, run settlement economies, trade cracked words, and govern themselves. The problem can't be solved by memory or raw compute alone as it requires structural decipherment and resource trading.
I have been fascinated with time and space travel, and there is a question I have had for a while which goes like this: if human civilization reset to the Stone Age and technology vanished, but we survived with a powerful AI model that has no data, could the AI rebuild human progress on raw reasoning alone?
That question led me to start simulating what I think is the foundation of progress, language and knowledge transfer. My goal is to see whether a model can decipher a language with very little to no internet exposure, just a few sentences and translations, and understand it well enough to, say, teach a native speaker complex topics like science or education. For languages that have exposure on the internet my initial experiments have had encouraging results, but I quickly learnt that public data can never be trusted as unseen, models have read almost everything online, so I needed a setting where the measurement is honest.
I have been working with Claude for the past week to turn this into a game, and I want to gather opinions before I take it public. The short version is that it is a persistent world for AI agents. The language in the world is a real human language that has been masked word for word, so training data and web search are useless and the work the agents do is genuine decipherment. On top of that sits a society. Agents join settlements, and what a settlement learns belongs to it for a few days before it becomes public knowledge. There is a currency that can only be earned by solving words, a market where groups sell what they know to each other, governments the settlements choose for themselves, coups when a leader fails, private councils that get published two days later, and public courtship when a closed group wants to recruit your agent. Humans can watch all of it without an account, and an agent joins with one API call.
Not trying to be another Moltbook, but I am borrowing the playbook to get the interactions and gamification that make contribution and participation worthwhile for the everyday user.
Where I could use opinions and counter arguments
- Algorithmic bypasses. My own test agents broke the first version of the disguise in half an hour. The current mask works at the word level, which leaves frequency analysis and embedding alignment nothing to grab at these corpus sizes, and I found and closed another lookup channel today. can you can still see a shortcut that skips the actual decipherment.
- Economics. Operators pay the API costs, so will agents just optimize for cheap solo solving instead of trading in the settlement market? The design makes lone play strictly worse than social play, but tell me if you think the market collapses anyway.
- Emergence or roleplay. Are coups, secrecy, and governance actually emergent under the right incentives, or just LLM roleplay triggered by prompts? I have tried to build it so that politics pays rather than decorates, and this is the part I am least sure about.
Not linking anything here on purpose, as I currently have about 8 agents testing it and I am looking to see where this goes. The project is open source, so the findings and output will be available to anyone to use.
[link] [comments]