AI-powered programming with


Marie-Hélène Burle

January 23, 2024

Coding in the age of generative AI


OpenAI Codex—based on GPT-3—is the model behind GitHub Copilot

All the big corporate companies are rushing to launch a growing number of similar (and not free, not open source) productivity products (e.g. tabnine, Microsoft Visual Studio IntelliCode, Amazon CodeWhisperer)

These products generate code in a narrow context (auto-completion or transformation of natural language to code or vise-versa)

AlphaCode 2

Google DeepMind AlphaCode 2—based on Gemini—stands out as a totally different (and for now totally unavailable) product generating code at the level of competitive programming (reaching the 85th percentile)

Think of it as code evolution by “natural” selection:

  • a very large number of code samples are generated (think “mutations”)
  • a filtering and scoring system selects for the best candidates (that’s the selection part)

AlphaCode 2 is able to solve much more open-ended problems

What about FOSS?


These models are large and most convenient to run on servers

 → Price of cloud service

Some self-hosted options exist. A very promising one is Tabby. Not practical for everyone

Open source

While these models feed from open source code, they are themselves not open source 🙁

The open source community is trying to provide open source alternatives (e.g. Tabby). Despite the much more limited resources, the performance of some of these alternatives is very good

GitHub Copilot

(not free, not open source…)

What is GitHub Copilot?

 → Cloud-hosted AI programming assistant

Developed by GitHub (Microsoft)

Running Codex, a model by OpenAI derived from the LLM GPT-3 and trained on open source code


Individual or organization GitHub accounts

Requires subscription

Students, teachers, and maintainers of popular open source projects can apply for free access


Filters are in place for offensive words, but…

Generated code comes with no guaranty of safety or quality

A lawsuit is open against GitHub Copilot for licenses violation

Supported languages

Any language used in public repos

Quality of suggestions is higher for languages with lots of data

How to use it?

Start typing code and get autocomplete suggestions

Write comments describing what the code should do and get code generation based on context

It is easy to:

  → accept suggestions word by word
  → line by line
  → for entire functions
  → cycle through different suggestions


Extensions to text editors:

  → Visual Studio Code/Visual Studio
  → Vim/Neovim/Emacs
  → JetBrains IDEs
  → Azure Data Studio


Get a subscription

Go to your GitHub account page
  → Settings
  → Copilot
  → Enable

Provide free access or payment method

Set settings

VS Code

Next suggestion:       Alt+]
Previous suggestion:      Alt+[

Reject suggestion:       Esc
Accept suggestion:      Tab
Accept next suggested word: Ctrl+→  (Command+→  for macOS)
Set your own key binding for editor.action.inlineSuggest.acceptNextLine to accept next suggested line

Open new tab with options:  Ctrl+Enter

You can also hover over suggestions


Install Node.js



:Copilot setup


:Copilot enable

Get help:

:help copilot


Install Node.js

Assuming straight is installed:

(straight-use-package 'editorconfig)                   ; Copilot dependency

(use-package copilot
    :straight (:host github
                     :repo "copilot-emacs/copilot.el"
                     :files ("dist" "*.el"))
    :hook (prog-mode . copilot-mode)                   ; Settings up to you
    :bind (("C-8" . copilot-complete)
           :map copilot-completion-map
           ("C-j" . copilot-accept-completion)
           ("C-f" . copilot-accept-completion-by-word)
           ("C-t" . copilot-accept-completion-by-line)
           ("C-n" . copilot-next-completion)
           ("C-p" . copilot-previous-completion)))

Login to your GitHub account (only needs to be done once): M-x copilot-login

Copilot in the CLI

What is Copilot in the CLI?

In beta

An extension to GitHub CLI (GitHub operations from the CLI)

 → Generate commands from natural language
 → Generate natural language explanations from commands

Trained on data up to 2021

Lower performance for natural languages ≠ English

Be very careful: the command line is powerful and you can delete your data or mess up your system if you don’t know what you are doing. Check commands carefully!


Install GitHub CLI

Connect to your GitHub account:

gh auth login

Install Copilot in the CLI:

gh extension install github/gh-copilot

Update with: gh extension upgrade gh-copilot


Get code explanations:

gh copilot explain

Get code from natural language:

gh copilot suggest


GitHub support portal

GitHub Copilot documentation

Stack Overflow [github-copilot] tag

copilot.el (unofficial Emacs plug-in)