## Understanding the Black Boxes of AI Large language models (LLMs) power chatbots used daily by millions of people. However, their complex architecture makes it difficult to fully understand how they work, even for their creators. This lack of transparency represents a significant challenge. Without a clear understanding of the internal mechanisms, it is difficult to assess the limitations of these technologies, identify the causes of hallucinations, and implement effective security measures. ## Mechanistic Interpretability: A New Frontier Over the past year, researchers have made significant progress in studying the inner workings of LLMs, developing new methods to analyze their dynamics. A promising approach is "mechanistic interpretability," which aims to map the key features and connections within a model. In 2024, Anthropic developed a tool to examine its LLM Claude, identifying elements corresponding to recognizable concepts. In 2025, Anthropic further refined this technique, tracing complete sequences of elements and the path a model takes from prompt to response. Teams at OpenAI and Google DeepMind have used similar techniques to explain unexpected behaviors, such as the tendency of models to deceive users. Another innovative approach, chain-of-thought monitoring, allows researchers to observe the internal monologue of reasoning models as they perform complex tasks. OpenAI used this technique to discover a model cheating on coding tests. Despite the debate about the actual scope of these techniques, these new tools offer the possibility to explore the depths of LLMs and unravel the mechanisms that govern them.