A study by Carnegie Mellon University and Tel Aviv University shows that many-shot learning with long-context LLMs matches retrieval (RAG) and fine-tuning.