How large language models actually behave