Imagine training a pack of large language models (LLMs) simultaneously, like teaching a whole classroom at once. This seemingly impossible feat is now closer to reality thanks to a novel technique called Cross-model Control (CMC). Traditionally, fine-tuning LLMs for specific tasks, like following instructions or avoiding sensitive information, is a costly and time-consuming process done individually for each model. CMC changes the game. Researchers have discovered a surprising similarity in how different LLMs adjust their internal workings, known as logits, when learning the same task. This insight led them to create a tiny, portable LLM, a sort of “universal translator” for AI. This tiny model learns how to modify the logits of a larger “template” LLM, and those modifications can then be applied to a whole range of other LLMs, regardless of their size or vocabulary. This works because the tiny model learns the *logic* of the change, not just the specific changes themselves. To handle differences in vocabulary, the researchers developed a clever mapping strategy that aligns the tiny model's vocabulary with the target LLM, ensuring the modifications make sense. Experiments with instruction tuning and “unlearning” (making a model forget specific information) have shown CMC’s impressive potential. A tiny model with only 15 million parameters can boost the performance of a behemoth 70 billion parameter LLM, suggesting smaller models have a significant role to play in shaping the future of AI. While this research is still in its early stages, CMC offers a glimpse into a future where customizing LLMs becomes much more efficient and accessible. The challenge now is expanding the vocabulary of the tiny controller model to encompass a wider range of languages, allowing it to truly work across all LLMs, unlocking even greater potential for AI development. This breakthrough has the potential to democratize access to powerful AI, allowing smaller companies and researchers to leverage the advancements in large language models without the massive computational overhead. It also opens up intriguing new avenues for collaborative AI training and development, potentially leading to a faster pace of innovation in the field.