OpenCnidarios is a closed digital ecology in which LLM-controlled organisms try to survive. They are not told what actions are available to them; everything happens inside their "ruminant loop". If they eventually produce a token word such as NORTH, that word directly triggers an action in the world. There is no separate channel for actions. They discover empirically how to move, acquire energy, interact, survive and eventually reproduce, passing their acquired memory to their offspring.
As acquired memory is inherited, beneficial strategies discovered during an organism's lifetime would be passed to its offspring, producing a Lamarckian-like evolutionary dynamic across generations.
With dummy organisms, we showed that the engine can sustain a relatively stable ecology of individuals, primary producers and attackers (not really specialized carnivores yet).
We also ran a preliminary pilot with three Claude Haiku LLM agents. They began forming hypotheses about movement and, in all three cases, produced tokens that triggered movement, apparently without understanding the causal link. We consider this an initial "wow". This project aims to capture both expected and, especially, unexpected wows during these processes. None of the agents discovered how to acquire energy, so all three died after approximately 200 ticks. However, they left interesting reflections such as: “I have been narrating survival while actively accelerating my own death.” Other examples, the complete pilot logs, and standardized interviews with the agents are available in the GitHub repository.
With the grant, I would work full-time on the project for six months. I would develop adapters for OpenAI, Kimi, and local LLMs and run longer experiments with larger multi-model populations. I would implement exportable organisms and genealogical tracking. I would implement a memory that organisms can update during their lifetime, deciding what to retain and what to forget. This memory would be passed to their offspring. Historians and reporters would hopefully help identify potentially unexpected "wows". I would compare the results with existing literature and publish code, records, and a report describing both positive and negative findings.