What’s it like to code with GPT-4 and aider?

Below are some chat transcripts showing what it’s like to code with the help of GPT-4 using the aider command-line chat tool. In the chats, you’ll see a varity of coding tasks like generating new code, editing existing code, debugging, exploring unfamiliar code, etc.

• Hello World Flask App: Start from scratch and have GPT create a simple Flask app with various endpoints, such as adding two numbers and calculating the Fibonacci sequence.

• Javascript Game Modification: Dive into an existing open-source repo, and get GPT’s help to understand it and make modifications.

• Complex Multi-file Change with Debugging: GPT makes a complex code change that is coordinated across multiple source files, and resolves bugs by reviewing error output and doc snippets.

• Create a Black Box Test Case: GPT creates a “black box” test case without access to the source of the method being tested, using only a high level map of the repository based on ctags.

• Honor the NO_COLOR env var: The user pastes the NO_COLOR spec from no-color.org into the chat, and GPT-4 modifies the application to conform.

• Download, analyze and plot US Census data: GPT-4 downloads census data, suggests some hypotheses to test, tests one and then summarizes and plots a graph of the results.

• Semantic Search & Replace: Updating a collection of function calls, which requires dealing with various formatting and semantic differences in the various function call sites.

• Pong Game with Pygame: Creating a simple Pong game using the Pygame library, with customizations for paddle size and color, and ball speed adjustments.

• CSS Exercise: Animation Dropdown Menu: A small CSS exercise involving adding animation to a dropdown menu.

• Automatically Update Docs: Automatically updating documentation based on the latest version of the main() function.

• Editing an Asciinema Cast File: Editing escape sequences in an asciinema screencast file.

What’s happening in these chats?

To better understand the chat transcripts, it’s worth knowing that:

• Each time GPT-4 suggests a code change, aider automatically applies it to the source files.
• After applying the edits, aider commits them to git with a descriptive commit message.
• GPT-4 can only see and edit files which have been “added to the chat session”. The user adds files either via the command line or the in-chat /add command. If GPT-4 asks to see specific files, aider asks the user for permission to add them to the chat. The transcripts contain notifications from aider whenever a file is added or dropped from the session.

Transcript formatting

This is output from the aider tool.

These are chat messages written by the user.

Chat responses from GPT-4 are in a blue font like this, and often include colorized “edit blocks” that specify edits to the code. Here’s a sample edit block that switches from printing “hello” to “goodbye”:

hello.py
<<<<<<< ORIGINAL
print("hello")
=======
print("goodbye")
>>>>>>> UPDATED