Imagine writing tests for software *before* the software is even written. This seemingly paradoxical idea is the core of test-driven development (TDD), a practice that promises more robust and reliable code. But writing tests first is challenging. Could AI automate this process? Researchers are exploring whether Large Language Models (LLMs) can generate effective tests for issues even before developers write the code to fix them. A new benchmark, TDD-Bench Verified, puts LLMs to the test, evaluating their ability to generate tests based solely on issue descriptions and the existing codebase. The benchmark uses a rigorous evaluation harness, filtering for high-quality test cases and focusing on two key factors: whether the tests correctly fail before the fix and pass afterward (fail-to-pass), and how effectively they cover the changed code (adequacy). Early results are promising. A new LLM-based technique, Auto-TDD, outperforms existing methods, achieving a fail-to-pass rate of 23.6% using GPT-4. Auto-TDD employs a three-step process: selecting the relevant test file, identifying issue-related functions, and generating the test function itself. Interestingly, LLM-based file selection proves crucial for performance. While the results highlight the potential of LLMs in automating TDD, challenges remain. LLMs sometimes produce syntactically incorrect or poorly formatted code, and their test adequacy, while comparable to human-written tests in fail-to-pass cases, drops significantly for other scenarios. The research also suggests that an ensemble approach, combining the strengths of different LLMs, could further boost performance. While more research is needed, the ability of LLMs to generate useful tests even before code is written offers a tantalizing glimpse into the future of AI-assisted software development. This could lead to increased developer productivity and, ultimately, more robust software.