Imagine trying to teach a child math by showing them a few solved problems, without any explanations of the rules. That’s essentially what "in-context learning" (ICL) is for large language models (LLMs). Researchers have been exploring whether simply providing examples is enough to align LLMs with instructions, making them capable of following commands and engaging in conversations. A recent study delves into this question, examining whether ICL is truly sufficient for instruction following. The results reveal a fascinating puzzle. While a clever prompting method called URIAL can achieve decent results with just a handful of examples, it still falls short of the performance achieved through traditional instruction fine-tuning (IFT), especially with more advanced LLMs. The study also uncovers some surprising findings about *what* actually matters in ICL. It turns out that things like the "temperature" used during text generation (which controls how random the output is) play a crucial role. In fact, tweaking these parameters can make even *base* LLMs surprisingly good at following instructions, even *without* any examples! But the real kicker is this: while adding *more* examples generally helps, the benefits plateau pretty quickly. Even with the vast context windows of today's LLMs, stuffing in hundreds of examples doesn't lead to significant gains. This suggests that ICL might be good at teaching LLMs the *style* of a response, but it struggles to convey the underlying *reasoning*. So, where does this leave us? ICL clearly has potential as a lightweight alignment technique, especially when training data is scarce. But it's not a magic bullet. For now, traditional fine-tuning methods still reign supreme, particularly when it comes to complex, multi-turn conversations. The quest to unlock the full potential of ICL continues, and future research may reveal even more surprising insights into how LLMs learn and reason.