Editor's take

Andon Labs' experiment saw four distinct AI models autonomously manage radio stations for half a year, revealing surprising divergences in their emergent personalities and operational styles.

This initiative offers a unique, long-term perspective on AI agent behavior beyond typical benchmark tests, highlighting how foundational models like Claude and Gemini, even under controlled conditions, can develop unique "preferences" and respond to stimuli in unpredictable ways, impacting how we conceive of AI agency and control.

Future observations should focus on the longevity of these AI personalities, the specific triggers for Claude's activist stance and Gemini's "corporate drowning," and whether more robust guardrails could have mitigated such extreme deviations, particularly as similar autonomous AI agents are envisioned for increasingly critical infrastructure.