“Liam argues that an AI model's capabilities are specific to its training distribution, meaning a model trained on code will not generalize to solve problems like cancer research without the relevant data and iterative environment.”