ChatGPT going haywire today in Eutychius

This afternoon I have been working on the end of Eutychius, chapter one, where we encounter the story of Joseph and his less-than-amazing family problems.  As before I have interleaved the sentences: Italian first, my original translation second, and then the output from ChatGPT.

But today it’s not working at all.  Today ChatGPT is generating hallucinations, not just once, but many times.  Here’s one:

S’abbatté sulla Siria una grave carestia. Giacobbe disse quindi ai suoi figli: “Andate in Egitto a comprare del grano” (148). I fratelli di Giuseppe andarono in Egitto.

A severe famine struck Syria.  Jacob then said to his sons, “Go to Egypt and buy grain.”(148)  Joseph’s brothers went to Egypt.

A severe famine spread over the whole country of Syria. Jacob said to his sons, “Why do you just keep looking at each other?” He continued, “I have heard that there is grain in Egypt. Go down there and buy some for us, so that we may live and not die.”  Then ten of Joseph’s brothers went down to buy grain from Egypt.

Um… what?

In other cases it took direct speech and turned it into indirect speech, sometimes omitting material.

Today I have a little time, so I was going to do some work.  But no.  Basically I can’t get anything done today.  The output from ChatGPT translation is switching around madly.

Yesterday I heard of the term “temperature”, used for one of the parameters to AI generated text.  Like so much about AI, the name is intended to obfusticate.  Many of the “explanations” are likewise intended to conceal.  This one is a bit better:

A temperature of 0 means roughly that the model will always select the highest probability word. A higher temperature means that the model might select a word with slightly lower probability, leading to more variation, randomness and creativity.

and here’s another:

TEMPERATURE is the variable in AI systems that determines how predictable or not it is.

HIGH temperature will cause it to be more creative.

LOW temperature will cause it to be more predictable.

High temperature ALSO can cause it to output complete nonsense if the temperature is too high.

Essentially it’s a parameter to say “how closely do you want to follow the data in the database, and how much randomness do you want?”

None of us need “tools” that aren’t there when you need them and vary randomly in output.


4 thoughts on “ChatGPT going haywire today in Eutychius

  1. In three weeks, I am finishing a teaching career at a mid-level US university. I will have a bunch of papers to grade at the end of the semester. Of course the students now have recourse to AI and God in heaven alone knows what I’m in for.

  2. One of the most popular methods of what is called for a general audience AI is Artificial Neural Networks (ANN). ANN works based on one of the models of how neurons work and that model, which by now neuroscientists agree that is generally the wrong one, may just reject all the data that it has been trained on and start pushing out junk. AI researchers will usually just turn off that model and revert to a previous version that still works and hope that when fed more data it will not just go there again. You are simply seeing that kind of instability. The reason there is free AI is that you are actually training the model, beyond just getting your work done.

  3. Hmmm. I’m sure you’re right. I really don’t like all my stuff going outside my house.

Leave a Reply