Or, if we consider the fact that an LLM’s performance depends on the task’s similarity to others in the training set, it could be that one person is doing a fairly novel task and another is doing something very well represented in online code.