On pitfalls (and advantages) of sophisticated large language models