This is a submission for the GitHub Copilot Challenge : New Beginnings
What I Built
My idea was to create a small Telegram bot using Copilot, similar to the many AI-powered bots I’ve built over the past year. The concept for the bot was inspired by the theme of the competition, “New Beginnings.” The bot itself was designed as a generic assistant to help users train and improve their personal growth with AI.
Demo
The bot is available at this link:
https://t.me/motivat1vation_bot
Keep in mind, though, that it’s running in a Docker container on my laptop, so it might be offline if my laptop (or I) is asleep. It’s better to build your own instance using the repository on GitHub.
Repo
https://github.com/vorniches/motivation-bot/
Copilot Experience
I’ve been coding alongside AI since the release of original GitHub’s Copilot, which predated all the ChatGPT hype. Back then, it was essentially an advanced autocomplete – though it often produced nonsense. That was during the early days of its public beta in late August 2022. At the time, there was nothing like it, and Copilot felt revolutionary.
Then, in late November 2022, ChatGPT 3.5 became publicly available. It was quirky and unpredictable – those early days of testing are unforgettable for anyone who was there. Later, ChatGPT 4 came out, API access followed, and working with AI steadily became more exciting and productive.
By the time GPT-4o arrived, I was writing 90% of my code by simply tasking the AI. It had become a natural way to create software for me, and with the introduction of o1 models, this approach was solidified. Coding became intuitive and almost effortless – but it wasn’t always this way.
Looking back to the GPT-4 era, I remember spending hours trying to get the AI to generate something reasonably coherent. I wasted time on tasks that could have been debugged in minutes by reviewing the code myself. Unfortunately, the modern Copilot feels like a return to those days.
If you’re planning to try coding with Copilot, your choice of models is limited – either GPT-4 or the better, but still far from perfect, Claude 3.5 Sonnet. With the latter, you can (albeit awkwardly) produce functional code directly in VS Code, but for some reason, it refuses to interact with project files if you start it up there first.
The main advantage of Copilot is that it can edit files.
While it doesn’t always do this perfectly, it’s a significant improvement over ChatGPT, which lacks full integration with the project environment. Unfortunately, that’s where the advantages end.
Debug feels good though
Coding with Copilot often feels like a step back to the “bad old days.” Its output seems unpredictably dependent on various factors, and you can’t rely on it to provide a reasonable response to a simple request. You often have to layer instructions on top of instructions just to stop it from modifying files it shouldn’t touch.
What’s most surprising is how inconsistent its behavior can be – what worked perfectly 10 minutes ago might yield a radically different response now. At times, I had to switch from Copilot Edits to Chat mode because it started replacing existing files or behaving erratically.
Conclusion
In conclusion, Copilot offers limited model options, operates inconsistently at the level of a code editor assistant, and provides a user experience that is more frustrating than positive. A year ago, this level of functionality might have been a game-changer. Today, it feels like something you’d rather forget, like a nightmare.
Source link
lol
