Hacker News·Sep 8 Large language models develop novel social biases through adaptive exploration HTML/javascript