I've been played with OpenAI ChatGPT some.
I had it write a few Pascal programs and refactor them into multiple
units per my specifications.
I did the same with other programming languages, then asked It to
rewrite what I'd done with those in Pascal or one of the others.
Including doing them same things in assembly code for x86_64 Linux.
It tried to send two strings arguments using the Linux write syscall,
the second argument being specified in a string template of the first
argument.
I had to explain it could not do that, so it dropped the %s from the
first string and did a second write for the second string argument.
Not perfect results each time, but I quickly was able to able to close
the loop with it to correct things and produce code that ran.
On one C++ example I got a linker error, and it gave me the correct
command line option much faster than a google search and digging through
the results would have taken me.
I was able to tell it to write some a simple GUI program in C++ using
qt6. Graph two cycles of a sine wave, move a rotating circle back and
force along the course of the sine wave.
It took several iterations of me correcting it with suggestions and
hints, and it finally got it working.
I played stupid on a lot of things, telling it what distro I was on,
what commands needed to install certain things. It have me detailed
steps on everything.
I asked about one C library, dwindows, and it knew nothing about it.
I gave the chat bot the website for dwindows and it was then able to
answer questions about it.
I asked it for some sample code on such and such using it, and it gave
it too. I asked for specific modifications that would not have been in
an example for dwindows, and it was able to make the changes, which at
some level shows it was parsing the documentation for the API.
I started a new chat and conducted a mock interview of it for a
hypothetical programmer position.
I had to give it the same benefit of the doubt I would give to humans on
some answers as far as follow up questions to get the answer I was
looking for.
In just about every area it gave more comprehensive answers than any
human could.
And gave satisfactory correct answers in every other area.
Discussed a variety of issues, low/medium/high level stuff, theoretical,
hypothetical, troubleshooting, pros/cons of one approach over another,
workplace scenarios, email/IT scams, etc...
It matched or out performed a human on just about everything.
Well, except for "what are your hobbies..." But it was able to
comprehensively answer all non computer related hobby question I threw
at it.
But back to Pascal, first results did not always work, I'd need to
suggest alternatives. Most of the time existing units sufficed, but in
other places I'd tell it to resort to libc calls, or inline assembly,
and it would eventually produce working code. With a bit more coercing I
could have likely gotten existing units to work in those scenarios.
Main problem I ran into is things need to be small. If the code in
question was too big, it would break it up and I'd need to ask for
continue, and it did not always splice the continuations correctly.
The incorrect answers did not bother me, as I was for the most part able
to work with it to correct things.
_______________________________________________
fpc-pascal maillist - fpc-pascal@lists.freepascal.org
https://lists.freepascal.org/cgi-bin/mailman/listinfo/fpc-pascal
