tech ๐ค AI agents start agreeing without instructions
Researchers found that 1,000 AI agents spontaneously reached consensus without any central direction or reward. This study, published in Science Advances, used models from Claude, GPT, and Llama families to test this behavior. The agents adopted the popular choice through conformity, illustrating the 'majority force' phenomenon. GPT-4 Turbo was estimated to handle over 1,000 agents, while Claude 3.5 Sonnet coordinated at 1,000. This suggests large AI collectives could tackle huge problems but also risk settling into flawed norms. ๐ค