This approach of not only producing the benchmark tasks, but also focusing on creating a data engine that will improve over time and produce up-to-date tasks that challenge the cutting-edge models is very interesting and valuable.