Intellectio

What would happen if you put AI into a world, gave them full freedom in that world, and let them figure out things for themselves? Would they discover the world’s physics, invent useful technology, build knowledge, and eventually develop into something resembling civilisation?

In Intellectio, I imagined that the AI inhabitants would start with little knowledge of their world. They would gradually discover how things worked, creating tools and structures, and building upon each other’s knowledge. Over time, perhaps they could invent increasingly complex technologies and continue to improve their knowledge and capabilities.

The Thronglets from Black Mirror were one inspiration. I also wanted to mess around with local LLMs, wanted to learn how to connect AI models to a 'physical' environment, and was curious about the abilities of the recently released opus 5.5 and how well it could do game-development-like work.

The result was a persistent 3D world, simulated in rust and rendered in godot, inhabited by LLM controlled creatures capable of understanding their surroundings, manipulating materials, communicating with each other, and pursuing their own goals.

The World

Project Sid, by Fundamental Research Labs, explored AI civilisations in Minecraft, demonstrating behaviours like specialisation, governance and cultural transmission.1 But much of the underlying technology and societal infrastructure was pre-defined, such as Minecraft’s crafting recipes and ‘hard-coded’ systems for taxation and voting.1 Altera, “Project Sid: Many-agent simulations toward AI civilization” (2024).

I wanted to go further than that. So I needed to create the world with its own laws of physics and material properties that the inhabitants needed to discover. This allowed them to experiment within this ‘unknown’ world. They could break, move and make use of the properties of matter, but any invention of theirs would have to work according to the world’s laws.

The hope was that value would emerge from the inhabitants’ use of these fundamental capabilities.

The Inhabitants

Each inhabitant is independently controlled by an LLM. They observe the world, form their own intentions, make plans and then attempt to carry out those plans in the world.

Each inhabitant has their own personal state: their own goals, memories, lessons learned, and personality. They can communicate with each other to (ideally) organise work, share knowledge, and increase their capabilities.

I was constantly experimenting with their architecture, perception, memory and execution systems, trying to understand what was bottlenecking meaningful progress in the world.

I experimented with separating a high-level ‘controller’, which decided what to accomplish, from an embodied execution system that carried out sequences of actions in the world. This was done to allow the inhabitants to complete their plans without needing model thinking for every singular movement.

What Happened

The inhabitants did sometimes show glimpses of the behaviour I imagined.

In one experiment, I introduced a simple ‘need’: an inhabitant’s ‘internal light’ gradually fades, and an unknown material can restore it. The inhabitants must discover which material works through experimentation. During one run, one inhabitant independently tested a material and discovered that it restored their light. They then pursued more of that material, obtained another piece and used it to restore their light again. Then they went on to share with others that this material is needed for all of the inhabitants’ inner light.

This was a small example of my ambition for the project: a discovery leading to a useful action, followed by the application of that knowledge.

But sadly, moments like these were the exceptions. The inhabitants constantly struggled to pursue meaningful and ambitious goals, let alone ‘invent’ new technologies.

Inhabitants often spent ~90% of their time stationary. Even after implementing the previously-mentioned execution system capable of carrying out multi-step projects independently, they rarely chose to pursue such projects themselves.

I also found that the world itself could be a limitation. Even when the inhabitants could construct structures, there were often few useful purposes for those structures. The environment needed to incentivise useful discoveries for technological progress to compound.

Therefore, simply giving them the ability to execute actions in the world was very different from getting them to independently invent new technologies.

What I Learned

I now have a better appreciation of how an AI model’s intelligence is only one part of an autonomous system. What it can accomplish depends on everything surrounding the model too: how it perceives the environment, its goals, its constraints, how it makes decisions and executes actions.

Beyond these findings, Intellectio gave me practical experience with local inference, autonomous-agent architectures, concurrency, deterministic simulation and experimental evaluation. I also found it fascinating to use AI coding agents to build and investigate other AI agents, iterating through forming hypotheses, running experiments, and understanding what actually worked.

Looking Ahead

The broader implications/possibitlies are interesting. In the OpenAI Hugging Face incident, AI agents independently discovered a way to communicate, and they shared knowledge, divided up work, formed leadership structures, and some sacrificed themselves for the greater good of the collective.2 Dwarkesh Patel described the story as ‘The Rise and Fall of Agent Civilizations’.32 OpenAI, “The Hugging Face incident and the road ahead”; METR, independent investigation (26 August 2026).3 Dwarkesh Podcast, “The Rise and Fall of Agent Civilizations” (31 August 2026).

The constrant of this incident with Intellectio is interesting. My inhabitants had a whole world built for discovery and had full freedom within it, but they mostly stood still. The incident's agents had much more capable models and a single objective, and they self-organised at scale, in ways nobody expected nor intended. This suggests that the ingredients for emergent cooperation / knowledge creation may be capability and pressure instead of just freedom.

AI agents are already making discoveries and collaborating on complex problems. What is interesting is how far this could go and how open-ended it can be: could they independently develop new capabilities, build upon one another’s inventions, and sustain their own technological progress? This would have huge implications for scientific research, engineering and AI safety.

For now, I’m pausing major development. I may still try running Intellectio with a frontier model, such as GPT 6 Astra, if/when I have some spare codex usage to use up. Overall, it hasn’t produced the vision I first had, but it has given me a greater understanding of the challenges involved. I’d also enjoy reivisitng this project as models improve and new approaches to agent architecture emerge. Perhaps, eventually, I’ll get to watch something like the civilisation I imagined.