Publication decisions for large language models, and their impacts — LessWrong