Imagine a detective piecing together clues scattered across various crime scenes to unveil the hidden truth. Large language models (LLMs) are starting to exhibit a similar ability, called inductive out-of-context reasoning (OOCR). Researchers are discovering that LLMs can infer underlying patterns and connections from disparate pieces of training data, even when that information is never explicitly stated. This 'connecting the dots' ability allows LLMs to answer questions they haven't been directly trained on, hinting at a deeper understanding than previously thought. For instance, an LLM trained only on distances between an unknown city and other known cities can deduce that the mystery city is Paris, and even answer questions about Parisian cuisine! Similarly, they can figure out the bias of a coin from just observing individual flips, or define a mathematical function from input-output pairs. This research explores how LLMs achieve this feat, using a set of clever tasks that range from geographic reasoning to complex mathematical problem-solving. While this emerging OOCR capability is promising, it's also somewhat unreliable, especially for smaller models or complex relationships. The research highlights how larger models like GPT-4 perform better at this type of inference than their smaller counterparts. The ability of LLMs to infer implicit information presents fascinating possibilities, from improved question-answering systems to new ways of unlocking insights from complex data. However, it also poses potential challenges for AI safety. If an LLM can connect the dots to infer sensitive information never explicitly included in its training, it becomes harder to control and monitor the knowledge it possesses. This poses questions about transparency and potential misuse, underlining the need for further research into these powerful and evolving AI models.