Imagine a world where writing unit tests is as easy as chatting with a helpful AI. While that future isn't quite here yet, recent advances in large language models (LLMs) suggest it might be closer than we think. A new study examines open-source LLMs like CodeLlama and DeepSeekCoder to see how well they can automatically generate these essential tests for Java code. The results are a mixed bag. On one hand, the larger, more sophisticated LLMs show promise, crafting tests that achieve decent code coverage. However, they're still not as effective as existing tools like Evosuite, often producing tests that won't even compile. One key problem? The LLMs can "hallucinate," creating code with imaginary or mismatched components. The research also explored different ways of prompting the LLMs, finding that the right prompt can make a big difference. Including too much code in the prompt can actually hinder the LLM's performance, while other details, like the style of language used, can impact the effectiveness of CodeLlama models significantly. Interestingly, some advanced prompting techniques commonly used in other tasks, like 'Chain-of-Thought' and 'Retrieval Augmented Generation,' didn't boost performance here. The chain-of-thought approach struggled due to the LLMs' code comprehension abilities, while the retrieval method wasn’t effective because of the type of tests it retrieved. Ultimately, while AI-powered unit test generation isn't a solved problem, this research highlights the potential and the challenges ahead. Future improvements could involve tweaking how we prompt these LLMs, refining their code comprehension, and developing strategies to fix the faulty tests they sometimes generate. As open-source LLMs evolve, automated unit testing might finally become the quick, reliable tool developers dream of.