Does ChatGPT know what it is saying?
As a high level description, ChatGPT learns the linguistc structure of a language by using
lexical analysis to parse text into tokens
and then learning the frequency of how those tokens are organised.
It doesn't learn facts or logic, it only learns the way the sequence of tokens in the text that it has
learnt from.
ChatGPT uses a uses a
Generative adversarial network (GAN) to generate responses by ordering tokens in a similar way
as it has learnt is the most likely. It effectively generates a response by merging the most likely
token sequences. For example, this becomes evident when we ask ChatGPT about a topic that has divided
opinions in a manner that makes it hard for ChatGPT to realise that it is the same question.
First, to introduce the topic, below is a screenshot of ChatGPT's answer when asked in English what
Völuspá is.
The answer is a good summary.
However, asking in Icelandic and then in English whether the Poetic Edda is written in Icelandic gives
polar opposite answers.
Below is a translation from Google Translate of its Icelandic response:
"Yes, Eru Eddukäði was written in Icelandic. It is the oldest manuscript collection of Eddic poems
owned by the
Árni Magnússon Foundation in Iceland, which is also known as the Codex Regius. This manuscript was
composed around 1270 and is among
the best known and most important surviving manuscripts of the Eddus."
When answering the question in English, ChatGPT gives the opposite answer and says that it was written
in
Old Norse in Iceland.

As seen in the above screenshot, the question is posed in Icelandic and then immediately in English, yet
ChatGPT is unaware
that it has given factually opposing replies; it doesn't understand the question or the answers; they
have no meaning to ChatGPT.
Furthermore, since the questions and responses are in different languages, they don't have a similar
token sequence; they only share a few tokens,
such as the tokens for "Edda" and "Icelandic". So Chat GPT cannot see any connection between the two.
We can fix this discrepancy by asking it to translate the English question to Icelandic or vice versa.
See below:

Albeit the Icelandic name of the Poetic Edda is "Eddukvæði", not "Póetíska Eddan". So the original name
is lost in ChatGPT's translation. "Póetíska" is English slang and not an Icelandic word. Eddukvæði means
Edda's poems. Edda here means grandmother.