LLMs generate, log, and trace enormous volumes of date, with some models emitting tens of millions of logs per hour. Tools like grep laid the groundwork for logging, and while they served us well in the past, they have quickly become inadequate when faced with the scale of LLMs. The linear search capabilities of grep are simply too slow for today’s real-time demands. In this talk, I’ll demonstrate how modern, robust observability tools can transform the way we handle logs at scale. I will explore how to use OpenTelemetry (OTel) and Elastic Observability to tame the chaos, working with data from Hugging Face and Mistral-based LLMs. I’ll start by showing how grep struggles with high volume logs, and then contrast that with the power of modern observability stacks. We will walk through exporting telemetry data via OTel and EDOT into Elastic, and we will compare the insights you can extract from them versus grep. For this talk, while some foundational observability knowledge will be helpful, I want to make this accessible on all levels. I will explain the core concepts of observability of LLMs throughout the session. Whether you are debugging LLMs or scaling AI infrastructure, this talk will help you move from reactive grepping to proactive observability.