Imagine teaching an AI to play games like Taboo or Wordle. Sounds simple, right? But what if the AI had to learn the rules through conversation, not explicit programming? Recent research delves into this fascinating challenge, exploring how an AI's "brainpower" (measured by its parameters) affects its ability to play conversational games. Researchers used a clever benchmark called 'clembench' which presents various dialogue games to AI models. The results? While larger AI models generally performed better, there were surprising exceptions. Some smaller models, when trained with specific data, were surprisingly adept at following game instructions and even showed sparks of clever strategy. This suggests that simply increasing the size of an AI model isn't a magic bullet. Instead, the quality and type of training data plays a crucial role. The research also highlighted an unexpected quirk: how we access and interact with these AI models can affect their performance. Different interfaces and even methods of compressing the model's size (quantization) yielded different results, adding a layer of complexity to AI deployment. This research reveals a dynamic interplay between model size, training data, and access methods in shaping an AI's ability to learn and perform conversational tasks. It offers a fresh perspective on the ongoing quest to develop AI that can seamlessly interact with us, not just solve complex equations, but engage in witty banter and navigate the nuances of human conversation. The path towards truly conversational AI might not be about building ever-larger models, but about understanding the subtle factors that unlock their existing potential.