Basic understanding of how the model works.
As said elsewhere, the model doesn't have memory, it doesn't learn, and doesn't have goals. If you were allowed to play with it, you could quickly observe this for yourself with a bit of adversarial play - ask it about the time it fought Genghis Khan, ask it if it wants you to give it a glass of water, etc - you will quickly realize the illusion behind the system. Tell it whatever facts you want, be amazed how well it might recall them, and then start a new conversation and notice that it has entirely forgotten all of them (again, it has no memory, and it doesn't learn anything).
From first principles, we know how the LM works: it has been trained to predict what is the most likely next token in a series of tokens. It has been trained to do this on a gigantic corpus, and is now pretty good at it. You then present it with a prompt, and it starts predicting the most likely token that comes after that prompt, and it keeps predicting for a while until the most likely "token" is to stop. Note that these tokens aren't even words - they are typically letters, or short combinations of letters. To give the illusion of a coherent conversation and memory, the prediction function is applied to the entire conversation so far, not just to the latest prompt.
There is more sophistication to avoid various undesirable behaviors, especially in LaMDA (e.g. the model is also trained to evaluate the "safety" of various possible answers, so it doesn't generate a sentence that insults you or discriminates against you or induces you to do something dangerous etc), but this is the gist of how it works.