Fine-tuning vs. in-context learning: LLM customization for real-world tasks