Can Large Language Models Invent Algorithms to Improve Themselves?