AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

Claude Cooperates: Cultural Evolution in LLM Societies

· The Cognitive Revolution

Edward Hughes (Google DeepMind) and Aron Vallinder (independent researcher, PIBBSS fellow) discuss their paper exploring cultural evolution in toy AI societies through a classic behavioral economics experiment called the donor game. Agents decide how much of a resource to donate to a paired recipient who receives double the amount, with the top 50% surviving each generation and their strategies culturally transmitted to new agents. Using reputation information across multiple rounds, the researchers tested which leading LLMs can sustain positive-sum social norms over time. Results showed striking differences: Claude 3.5 Sonnet achieved 3,000-5,000 units (out of a theoretical maximum of 32,000) with increasingly pro-social behavior across generations, while Gemini 1.5 Flash cooperated only limitedly to reach a few hundred units, and GPT-4.0 showed minimal cooperation with almost no resource growth. Early mixed-model experiments showed slight improvement over GPT-4.0 alone but with decline over time. The work highlights significant blind spots in benchmark-centric AI evaluation and emphasizes the critical role of reputation and enforcement of cooperative norms. The conversation extends to broader implications for AI agent deployment, the need for norms and regulations to create trustworthy multi-agent environments, and the importance of evaluating cooperation dynamics rather than just individual model capabilities. The guests argue this kind of research is uniquely accessible due to AI coding assistance and open-source code, making it an ideal entry point for social scientists and economists to contribute to understanding how AI societies will behave as agents become more prevalent in 2025 and beyond.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.