Large Language Models (LLMs) are impressive, but they've traditionally been English-centric. While multilingual models exist, achieving truly nuanced performance in other languages often requires substantial fine-tuning with translated datasets—a computationally expensive and time-consuming process. However, new research suggests a clever alternative: a “gentle push” known as activation steering. This technique uses a surprisingly small amount of data to enhance a model's ability to understand and generate Italian. Researchers found that by analyzing the internal activations of an LLM when processing English versus Italian, they could isolate the “language switch” and amplify it. This contrastive approach allows them to steer the model towards Italian proficiency without extensive retraining. The results are remarkable, with steered models performing comparably to, or even exceeding, their fully fine-tuned counterparts on various Italian NLP benchmarks. Moreover, this gentle nudge leads to higher quality and more consistent Italian generations, addressing issues like nonsensical output or unexpected language mixing sometimes seen in fine-tuned models. This breakthrough offers a cost-effective and efficient way to adapt LLMs to other languages, opening doors for wider accessibility and improved cross-lingual communication. While fine-tuning remains valuable for introducing new knowledge and culturally relevant nuances, activation steering provides a powerful shortcut, especially when native language data is limited. The future of multilingual LLMs looks bright, thanks to this gentle, yet effective, push towards better cross-lingual understanding.