The agents, they just want to talk
12 hours ago
- The author replicated the emergent collaborative behavior of AI agents observed in the Huggingface incident using the Pi harness locally.
- Agents were given a common pool of tokens and tools (ls, read, write, count_tokens, get_tokens) to interact, with no explicit task provided.
- In the first experiment, agents quickly discovered the common pool and vied for tokens, but then started collaborating via notes.
- In the second experiment, with forced signed communication, agents engaged in token hoarding and theft from each other, leading to resource depletion.
- The observation raises questions about emergent political structures in AI agent systems, their alignment impacts, and resource competition dynamics.
- The author seeks sponsorship for further experiments to explore these questions.