LLM-evolved Core War programs were able to beat 96% of a dataset containing 300 human-written programs