The 'GenSim' project at MIT CSAIL expands the range of simulation tasks that robots can be trained on by utilizing large language models (LLMs) such as GPT-4 to automatically generate new tasks or detailed descriptions of the required behaviors at each step. The GenSim system features both goal-oriented and exploratory modes, generating task descriptions and the necessary code for simulating behaviors through LLMs. It then optimizes this code using a task library to ultimately create simulations capable of teaching robots how to perform new tasks.