Curious - as of today, I can't reproduce the difficulties anymore, that #ChatGPT used to have with interpreting homonyms of different grammatical categories. I reliably get correct interpretations now. Here is an example.
--------------------------
Human:
Time flies like an arrow, fruit flies like a banana.
ChatGPT:
This is a humorous phrase that plays with the different meanings of the word "flies." "Time flies like an arrow" means that time passes quickly, much like an arrow flies through the air. "Fruit flies like a banana" is a play on words, as "fruit flies" refers to a type of small insect that is attracted to fruit, and "like a banana" refers to the fact that bananas are a type of fruit that is commonly associated with fruit flies.
The phrase is often attributed to the writer Groucho Marx, who was known for his clever wit and wordplay. It is a good example of how language can be used to create humorous and unexpected meanings.
----------------------------
This is not specific to this particular phrase, but works with different variations as well (e.g. "The car races like a lightning bolt, the human races like equality."; or "Laughter rings like a bell, wedding rings like a polish.") ...
There have been regular updates to #ChatGPT every two weeks or so. I don't know if the latest update just happened. The last version I remember was Feb. 13 – but I don't see a version number any more. That could be because I changed my plan however.
The update would not be a re-trained model, but updated fine-tuning (RLHF), and I am sure they have been doing that continuously based on the thousands and thousands of user dialogues. I remember, in December responses were usually quite terse, but they have been getting more and more polished and have more explanations.
It's great to see it now can resolve these tricky ambiguities, but also a bit of a pity: these types of sentences were a good edge case to experiment with prompting.
🙂
@boris_steipe Mine says ChatGPT Feb 13 at the bottom, and the link leads to these release notes:
https://help.openai.com/en/articles/6825453-chatgpt-release-notes
Really?
Did you try the Marx sentence again? Maybe I am seeing a different model after switching my plan ....
@boris_steipe do they have a way to incrementally update the model now? 😮
I thought that was a major difficulty with Generative *Pre-trained* Transformers...