I was curious to see whether ChatGPT could write yesterday’s intersection program for me. So I told it:
“Write a Javascript program to intersect two convex polygons by checking whether any vertex or the center of either polygon is inside the other polygon.”
It responded with a perfectly correct Javascript program. I then asked the question again, and it responded with a different Javascript program, also perfectly correct.
When I compared the code it produced those two different times, it appeared as though it was stealing pieces from different programmers who had very different styles of coding. I could recognize the thinking of the programmers who wrote the original code that it was using.
The final programs it produced tended to be kind of clumsy and inelegant — although definitely functional and correct. I am guessing that this was because it was stitching together parts of programs that it had found in different places. The result was a kind of Frankencode.
This is pretty much the impression I get when I read prose generated by ChatGPT. Just beyond the text, I can perceive the half-digested bits and pieces of the creative thoughts of unnamed humans.