Reconsidering LLM Uncertainty Estimation Methods in the Wild