Self-driving cars rely on intricate planning systems to navigate complex environments. These systems need to be both fast and adaptable. Current real-time planners are quick, but often lack the flexibility to handle unexpected situations. Recent advances in Large Language Models (LLMs) offer a potential solution, enabling more sophisticated scene understanding and decision-making for autonomous vehicles. However, integrating LLMs directly into self-driving systems presents significant challenges. LLMs are computationally intensive and can be slow, hindering real-time performance. A research team has created AsyncDriver, an innovative framework that uses LLMs to enhance existing real-time planning systems. AsyncDriver employs an asynchronous approach, decoupling the LLM's inference frequency from the real-time planner. This means the LLM doesn't need to process every frame, reducing computational overhead without significantly impacting performance. How does it work? AsyncDriver translates routing instructions into high-level features that the real-time planner uses to guide trajectory predictions. This adaptive injection block allows the planner to benefit from the LLM's advanced reasoning abilities without being bogged down by its computational demands. The system is also adaptable to different transformer-based planners. Experimental results on the challenging nuPlan dataset show AsyncDriver outperforms current state-of-the-art methods, boasting improvements in key metrics like drivable area compliance and time-to-collision. Interestingly, AsyncDriver maintains near-robust performance even when the LLM only processes a fraction of the frames. This efficient use of LLM power makes real-world deployment a more viable option. This work signifies a noteworthy step toward safer and more efficient autonomous vehicles, though it also highlights the need for future research to further examine the generalization properties of LLMs for this type of task.