The Eureka paper, co-authored by Jason Ma, was the first to use a large language model (GPT-4) to self-generate simulation environments and reward functions