Hey all, just want to start by saying I know very little about AI and have just been going down a rabbit hole thinking about multi-agent simulations and had a question I couldn’t find a clear answer to.
Most of the big simulation projects I’ve seen like Project Sid and Stanford Smallville use LLMs as the base, which means the agents already come loaded with human language, concepts, and cultural baggage before the experiment even starts. And things like Aivilization are cool but players are still actively guiding the agents.
Has anyone tried doing this with a non-language model instead? Like a reinforcement learning agent dropped into a simulated primitive environment with zero pre-loaded human knowledge — no language, no concepts, nothing. Just physics, consequences, and scarcity.
The idea being you’d want to watch what actually emerges on its own. Does something religion-shaped develop when the agent can’t predict its environment? Does communication emerge when you run multiple agents simultaneously? Does generational knowledge transfer look anything like human cultural evolution when you pass behavioral tendencies from one agent to the next without passing the full context?
Basically — has anyone tried building the conditions that forced human intelligence to develop rather than starting with intelligence that’s already human shaped? Is that possible?
Curious if this exists already or if there’s a reason it hasn’t been done. Sorry for the long post.
[link] [comments]