Why didn't OpenAI finetune the model to use the python tool it has for these tasks?