這篇文章可能解釋了大型語言模型目前的情況，以及它們的幻覺問題似乎比以往更嚴重

So, what the above graph means that a LLM is really good at solving average problems and are great at recombining existing knowledge, so, if i ask something outside my domain of expertise, i get really good answers but as you approach to the frontier of knowledge ( the point where what you already know meets what you are trying to discover), many times the outputs get random and less specific. Is it due to the lack of relevant structure in the training data? and the model doesn't know where to