"Generating thousands or millions of trajectories will help us find the right mechanisms for incorporating feedback and re-training those LLMs [Large Language Models]. There are numerous opportunities for developing new training engines, learning engines, and ultimately creating new benchmarks. We cannot improve what we cannot measure. The benchmarks are critical for assessing our current position."