Imagine trying to teach a super-smart computer, like a large language model (LLM), something new. Sounds simple enough, right? Turns out, it’s a lot more complex than you might think. Researchers are grappling with a fundamental problem: how do you effectively and rationally update an LLM’s "beliefs" about the world? The central challenge is that editing an LLM isn't just about adding information; it's about revising a complex web of interconnected "knowledge." Think of it like this: if you tell an LLM that the moon is made of cheese, what are the implications? Does it mean astronauts brought back cheese samples disguised as rocks? The ripple effects of such a change are hard to predict and even harder to control. This problem is rooted in the philosophical challenge of belief revision, which has puzzled thinkers for decades. How does a rational agent – be it human or AI – integrate new information that contradicts existing beliefs? Current LLMs struggle with this. They might change their answer to a specific question, but fail to update related information consistently. For example, if you correct an LLM’s mistaken belief about a historical figure’s birth city, it might still get other related facts wrong. Moreover, we don't even fully understand how LLMs "believe" things in the first place. Are they like agents with a coherent worldview? Or are they more like databases passively storing information? This ambiguity makes it hard to design effective editing methods. Researchers are exploring new formal testbeds to measure how well LLMs revise their beliefs compared to an ideal Bayesian agent, a theoretical gold standard for rational belief updates. Initial results show a clear gap between LLM performance and ideal Bayesian reasoning. The journey towards truly editable AI is far from over. But by drawing on insights from philosophy and developing more rigorous evaluation methods, researchers hope to unlock the full potential of LLMs as dynamic, adaptable learners.