_My views have evolved since I first wrote this post in 2023. _
Original post
A few days ago, I stumbled on this blog post, titled “Why I will likely never use AI programming tools”.
But the act of thinking and figuring out a problem, solving it, writing some code, fixing bugs…
Especially creating complex systems that interact with each other, with a lot of moving parts! (Hi, gamedev)
It’s fun to me.
This is exactly why I like using AI programming tools. Same reasoning, opposite conclusion.
AI tools allow me to completely evacuate the boring parts of programming: the boilerplate, the dumb loops, the language idiosyncrasies. When you consider AI tools - like the aptly named Copilot - as “helpers” that generate pseudo-code, you can focus on the fun parts: the problem solving, the architecture, the design.
The AI doesn’t write code for you, it writes code with you.
I feel like ranting against the AI is like ranting against the invention of the compiler. Yes you can find joy in writing Assembly code, but… well.
Now maybe that will change in 6 months, or a year or two. Maybe the AI will be so good that it will write code for you. Maybe I’ll be out of a job before 2030, but if that happens, I’m sure going against the flow won’t help. For better or worse.
2024-06-04 Update
From a Bluesky user:
Google explaining the “glue pizza” search result by stating EXPLICITLY that the wrong answer was given BECAUSE nobody had ever asked that question in the training data tells you that what we have isn’t artificial intelligence. It’s a server that copies, changes slightly, and pastes. Not intelligence
Awful for factual answers, perfect for programming : AI is a giant copy-paste machine, so this is exactly what you need for code 90% of time. Even if your program is groundbreaking and 100% original, most of its snippets are not.
- Source
- Glue pizza reference: Google Is Paying Reddit $60 Million for Fucksmith to Tell Its Users to Eat Glue
2024-12-24 Update
I cancelled my Copilot subscription.
It’s equally useful for some things (boilerplate code, unit tests, api discovery in an unfamiliar language or framework), and terrible for everything else. So in the end, I don’t think it saves me any time. What’s worse is that the most interesting thing it offers — code/api discovery — makes me lazy and not willing to understand what “I” write. I consider using Copilot for these tasks to be a net loss.
Many people have written much more articulate thoughts on it, so I’ll left it here. I’m glad I gave it a thorough try, though.
2025-08-06 Update
- Status at home: no AI auto-completion, though I still use Copilot free tier as a (wasteful) rubber duck. I don’t miss it when I’m out of tokens, though.
- Status at work: Copilot is also useful as a rubber duck for large, complex projects. Code completion is still a mixed bag, and I still estimate the net gain as zero.