For years, "schema linking"—the process of mapping natural language to database components—has been vital in Text-to-SQL systems. But what if we told you it might be becoming obsolete? New research suggests that the latest Large Language Models (LLMs) are so adept at reasoning that they can often skip this step entirely. Traditionally, schema linking helped narrow down the relevant parts of a database for an LLM to consider when translating a question into SQL code. This involved identifying the right tables and columns, and excluding irrelevant ones to improve accuracy and efficiency. However, this process wasn't foolproof. Sometimes, crucial information would get filtered out accidentally, leading to incorrect SQL queries. The surprising finding? Cutting-edge LLMs are often better at sifting through the entire schema themselves. They can pinpoint the 'needle in the haystack' even when bombarded with tons of irrelevant information, mimicking the human ability to focus on what matters. This eliminates the risk of discarding essential data during schema linking, which is particularly important for complex real-world databases. So, instead of focusing on filtering, the researchers explored enhancing LLM performance with complementary strategies: *Augmentation*: Providing richer context to the LLM, including detailed column descriptions and hints about the desired query. *Selection*: Generating multiple query candidates and picking the most consistent one. *Correction*: Refining the generated SQL based on actual database execution feedback. The results are impressive. By maximizing the information given to the LLM and using these advanced techniques, the researchers achieved state-of-the-art accuracy on a challenging Text-to-SQL benchmark. This shift marks a potential turning point in how we build and interact with databases. While schema linking might still be relevant for smaller LLMs or limited context windows, its role is diminishing as LLMs evolve. As these models become more powerful and context windows expand, we're moving closer to a future where natural language is the primary interface for accessing and analyzing data.