Large Language Models Develop Novel Social Biases Through Adaptive Exploration | Hacker News