A program that uses the ChatGPT API to generate code. A little bigger than Hello World, but not too much bigger. The interesting thing is that the program uses the output of one completion to set up the next.
Corresponding blog post
Generate a full web app with ChatGPT
On this page
I was looking for an interesting project that used the ChatGPT API and was wrtten in Python. I found smol-ai/developer which is described as the thing that builds the thing.
Given a prompt that describes a program in detail, it creates the code and writes out the files.
I was having trouble understanding the prompts, so I pulled them out into a separate file and reorganized and renamed the functions. Since I was more interested in the mechanics of writing in Python than in generating code, I removed the files that weren't germane to my goal.
The point at which I diverged from
smol-ai/developer
is at the branch named smol-ai/developer.
You need
- To clone this repo
- An OpenAI API Key
stored in
.env(See.example.env) - A Modal account
Run tepiton coder like this:
$ modal run main --prompt "write a 'hello world' in Python"The original used Modal,
and I had just started playing with it.
It seemed to take a little of
the pain of learning the Python ecosystem.
Using it feels like when I started using Git.
I'd just use the commands as incantations, without
really understanding what was going on.
That's where I am with Modal now. All that stub.function()
stuff is just magic to me. I hope I'm using right. but 🤷.
For what it's worth, in working on this project I've used about $0.01 of the free tier.
- Got rid of all the code except for
main.pyto concentrate on learning Python - Got rid of the devcontainer stuff. I do want to learn about devcontainers, but not right now.
- Moved the prompts to
prompts.pyso I could see what was going on. I gave the goofy names that made sense to me. - Added [pretty printing(./utils.py)] code to see what was going on
- Not interested in tokens at the moment
- Refactored a bit
You can see a more detailed walkthrough at How tepiton coder works.
The goal is to generate a multifile program, so the the first thing is to ask GPT to come up with the list of files it's going to generate.
The system prompt sets the task, The user prompt
is the prompt you provide.
Given the files needed, and what the user wants to do, the next step is to find the shared dependencies. With more complicated requests, this would generate a list of things that might be affected by the program.
We now have a list of files and a list of shared dependencies. We loop through the files and generate the code for each one.
The generated code ends up in the generated directory.
I'm learning about LLMs and using ChatGPT seems to be a convenient place to start. I've played with the OpenAI API in JavaScript, but it seems like all the cool kids are using Python.
The interesting thing about this project, of course, is how it uses the output of one completion to set up the next. That was the main thing I wanted to learn.
Prompts are weird. A prompt that works one day, doesn't quite work the next. My inclination when something doesn't work is to make the prompt more specific, but somehow I end up making it louder.
The original project was designed for GPT-4, which apparently generate the code for a Chrome plugin. I'm using GPT-3.5 and my aims are much more modest.
I still don't like Python. Probably because I don't use it much, I find it difficult to read. My eye keeps scanning for curly braces. It's not clear to me when newlines are OK and when they're not. It's a new ecosystem where they do things differently than in the JavaScript and Bash worlds. It is a strange and colorful place.