Discussion about this post

User's avatar
Ben P's avatar

I'm not a real programmer; I'm a statistician who writes a fair amount of code in the course of my work. I only know one language, R. Here's the general situation where ChatGPT is helpful:

I want to do a thing. I know I've done the thing before. I don't remember off the top of my head how to do the thing, but I do remember it was a pain in the ass and I had to spend a long time reading help docs and stackexchange to get it to work right. When I see the code that does it, I'll be able to understand exactly what it's doing and why it works.

That's where ChatGPT shines. Examples:

- I'm making a plot and on the x-axis are people's names. I want the names oriented diagonally, with the first name on top of the last name, so that none of the text overlaps and it doesn't leave me with losts of unnecessary white space.

- I have a data set of Supreme Court cases from the last two decades. I want a separate data set for each unique 9-justice lineup, and in each dataset I want a new character vector created that contains each of their names.

- I have a simulation that generates pretend high school students and their scores on a standardized test. I have a system in place for generating their scores randomly, via a multi-step process that allows for systematic differences across districts and schools and classrooms and SES levels. My simulation takes forever to run because I did it all using "for" loops, which are slow. I want it vectorized, which will make it go much faster, but I'm not good at this and I need help.

Note the many conditions:

1. I know precisely what outcome I desire.

2. I know that my outcome is possible, easily achieved, and I have a general sense of how to do it, I just can't come up with the syntax off the top of my head.

3. Once I see the code, I'll be able to read it and understand what each line does.

4. I know how to check to make sure it did what it was supposed to do.

5. These applications are my own personal projects; they don't introduce security risks and they aren't going to make life miserable for someone else because they aren't components in some larger project and no one else will be running the code.

The techno-futurists would look at these conditions as unfortuate limitations to be overcome by more advanced AI. My use is so... "narrow".

Of course, the less "narrow" our application of AI, the more problems it will cause. This will always be the case.

Expand full comment
mpsingh's avatar

Have you seen opinions regarding claude, it seems genuinely impressive, I want to see where is fails however

Expand full comment
1 more comment...

No posts