> There is some general principle of how it works that captures this behavior, It is statistical completion, but working at the nth order. What seems most lik

> There is some general principle of how it works that captures this behavior, It is statistical completion, but working at the nth order. What seems most likely to come next? An enormous number of possible responses are generated based on the prompt, and then tested for how likely it seems like everything else in its training data. A possible completion that is a fake biography of a person but which looks very much like a real biography, scores quite highly on the metric. ChatGPT has also been trained by feedback, with fake but convincing syntheses being penalised, and responses saying it doesn't know or using accurate facts being preferred. And so it also does the same testing process with the "how likely is it my trainer will like this completion?" metric. That explains the behaviour in the second session. In the first, it slipped by that metric.