What are the limitations of LLMs?

With both the statistical approach for the letters and the LLM approach for words, there is no formal knowledge that many expect, given its ability to answer many questions. There is no extensive knowledge base of topics, common sense, or anything except the word patterns it has learned. What is surprising to researchers is the fact that the learned patterns and their relationships somehow enable LLMs to perform new tasks for which they have never been trained. For instance, LLMs have even shown some capability in some foreign languages that are not part of their training set, through a process called zero-shot cross-linguistic transfer. Even if the output has been trained, one may get slightly different results each time from a prompt due to the randomness in the system. Much of the research in the prompt engineering area focuses on how to receive consistent and correct outputs from LLMs by crafting the prompts with specific phrases, commands, and...