Re: What's your view on AI coding assistants?

[haj]

Like several others here, I'd call them somewhat useful.

Sometimes I got stuck reading a specification and asked ChatGPT to explain a paragraph. Those results were almost always helpful-

I was really surprised that ChatGPT got me a working skeleton for a PDL::PP component. I had never tried PDL::PP before and this was really helpful. On the other hand, I could have also asked here or on IRC #pdl, and etj or someone else would have pointed me to pptemplate which does a better job. Actually, etj did that when I asked on IRC the next day.

Sometimes they failed at concrete tasks which I consider boring, but not actually difficult. When they produced wrong results and I pointed that out, they apologized at length and provided a "now correct" solution, which almost always was just as wrong or even worse than the first one. Their inability to say "I can't do that reliably" is really annoying. As a recent example, I asked them for vertex indices (how the vertices are wired together) in an icosahedron. The AI (I forgot which one I used for this) got the coordinates of the vertices right, but for the edges and faces... well. Sometimes it wrote "20 faces" and produced 18, or 12, or 22. When it produced indices for 20 faces, I occasionally produced the images for fun... i1, i2, i3...

The AIs guesswork was accompanied by "I appreciate your feedback. Here's the correct implementation", "I appreciate your patience. Here’s the correct... ", "Thank you for your continued patience. Here’s the corrected code", "I apologize for the confusion and appreciate your patience." "You're correct that in an icosahedron, each vertex connects to five faces.", "I appreciate your patience, and I sincerely apologize for the inaccuracies.", "I appreciate your continued feedback and apologize for the errors." ... and eventually I gave up.