Note: This article is mostly about my reflection on using AI. For how to use AI properly, other people wrote it much better than I do :). Check out our AI Native Summer Calendar!
Background
I actually used AI in my development flow quite early in my career. I received an invitation to the GitHub Copilot Technical Preview since 2021 (one year even before ChatGPT was released), and was fascinated by how AI can predict pretty accurately what I was going to type next. I could simply press Tab to complete my job. I got to experience the magic of AI quite early, and that technically qualifies me as an early AI adopter. That said, my opinion on AI remains mostly unchanged since. It is nothing more than a fancy autocomplete that can save me some time Googling and reading the documentation.
So what has changed?
Working on an AI product
I am very lucky to just recently join the Cometa team at primeNumber. One of our product's goal is to enable our customers to best leverage AI for their data utilization. It's essential for us to understand and get up-to-date about the latest models and their capabilities.
Another pivotal moment for me was seeing my coworker performing a Tony Stark-like maneuver. He spoke to a mic, the audio was transcribed into a prompt, then a Claude Code session was spawned to perform the task. It was very close to a scene straight out of a sci-fi movie. Even though I turned to him joking "Are you too lazy that you couldn't type out some text yourself?", internally I knew that I needed to re-evaluate my judgement about AI more seriously. In a true Show, don't tell fashion, I got convinced.
Claude Code is actually different
Note: For the rest of this article, I will refer to Claude Code as Claude
Part of the reason why I only considered AI to be a fancy autocomplete so far in my career, is that I have only used ChatGPT and similar tools outside of my editor. Usually, my workflow in the past was: I already have a specific question in my head, I then open a new ChatGPT session, feed it the relevant context and get back an answer. Sometimes I would ask for a code snippet instead, and then I would tailor it to fit into my code base. I was faster with AI for sure, but certainly not 10x faster as some claimed to be. Some newer tools came, such as Cursor, but I only saw it as an embellished version of Copilot.
And then Claude Code got released and changed everything for me. It was the first tool that brought me to this "a-ha" moment: AI isn't just a conversational buddy anymore, but a full-fledged operating system that can interact with the surrounding environment through tools (e.g. the file system, the command line). Now instead of having to distill my intention into specific questions, I can order Claude to, for example:
- Go read a GitHub issue, say about implementing an API (you can ask Claude to use
ghCLI to read the issue by the way!) - Point it to relevant files and folders. e.g. where are the routes, the OpenAPI documentation is located here, this folder contains some sample tests (later I discovered you could add a memory file for such task)
- Submit a PR when it is done
And it would do almost exactly how I would do it myself. Granted I only trust it with small-scoped tasks, but it is still a completely different way of doing my job.
Later, I also stumbled upon the claudelog blog (highly recommended!) and got fascinated by the idea of You are the main thread. In essence, you think of yourself as the main thread that can spawn multiple child processes to perform tasks in parallel. Think of the example issue I mentioned above, now imagine Claude can get it all done in the background while you are busy doing other (more interesting) works.
The good parts
Speed
A little bit more about myself, I am pretty efficient when it comes to using the computer. I type decently fast. I use neovim (btw!) and can move around files & make edits very quickly. I also have several shortcuts and hotkeys for opening Slack, replying to emails, etc. For most of the workflows that can be automated, I would probably have a bash script for it already (well maybe not for making a coffee).
However, there is no way that I can realistically compete with AI in terms of just raw speed. Unfortunately my brain isn't like a CPU where I can just spawn child processes to read multiple hundreds-of-LOC files, and grasp the pattern simultaneously. I also physically cannot turn an API specification from the design document to an OpenAPI schema in less than 5 seconds. Accepting defeat with humility means I have to rethink about the role of a modern software engineer. Constantly asking myself questions such as: what are the things that AI now can do better than me? Or how can I make the most use of AI to optimize my productivity?
Thinking out loud with AI
What I have found is if I can articulate my plan to Claude, it can perform the task pretty accurately. This is the common theme you could also notice from other articles in our AI Native Summer Calendar, like using AI for editing or making the calendar itself using AI!. This is expected since LLM models are extremely good at pattern recognition. You can level this up by utilizing a memory file, where you keep the common knowledge about the project (e.g. testing patterns) so that Claude can generate code closer to your expectation.
It's also worth highlighting that this isn't a novel action. Before AI, we were already constantly thinking about the current and the next step when doing a task. Now we only elevate that one level higher: articulate the plan into smaller, ideally independent steps. AI has actually challenged me to be more clear when explaining my thought process, which conveniently is also an extremely important skill for communicating with teammates!
The not-so-hidden costs
Skills atrophy: Use it or lose it
At primeNumber, Ruby is the primary language, which is something I had never had any experience with. Using Claude, I was able to generate decent quality code, possibly due to Rails being a highly opinionated framework and has stayed relatively stable throughout the years, but I also noticed that I hardly absorbed any knowledge. I tried to review carefully every piece of the generated code, but often find myself still forgetting the syntax, or simply being not able to comfortably rewrite it myself.
This aligns with what presented in one of MIT's study regarding using AI for writing essays, suggested that overuse of AI can erode your critical thinking skill. The author also pointed out that, "The task was executed, and you could say that it was efficient and convenient. But as we show in the paper, you basically didn’t integrate any of it into your memory networks.", which perfectly matches with my experience, but with coding. As someone who deeply cares about skills mastery, I now need to always stay alert and not to fully "outsource" my thinking to AI.
Note: I didn't use AI to write this article :)
The cognitive loads are still there
Another thing I also noticed is, although my "output" has drastically improved quantitatively, I'd still need to spend lots of time reviewing the code, as you cannot trust AI to do everything yet. Not only that, the moment the commits (signed by me!) are pushed, my coworkers will still have to comprehend those, and I have to take full responsibility for it. It feels like my duty shifted from writing the code to mostly reviewing. As engineers, we all know how much more difficult it is to understand someone else's code as opposed to your own!
Back to the idea of You are the main thread, even though I was able to "perform" four small tasks concurrently, I often find myself needing to switch context between reviewing AI's outputs. Unfortunately, context switching kills productivity, which completely defeats the purpose of using AI in the first place!
Key takeaways
It is about staying open-minded
I believe, when it comes to AI, your mileage may vary. While I'm about 30% faster with AI just by my gut estimation, some open-source developers actually get slowed down by it. I think the correct mindset here is to always be open to new ideas. Try new things out to see if it works for you, and don't forget to share if it does.
After all, don't forget to have fun
People have been saying AGI is 5 minutes from now since 2022. Luckily, humans are pretty bad at predicting the future. So while waiting for the machines to wake up and take all of our jobs, I think it's important to remind ourselves that we can have fun in our jobs. I certainly am still having fun smashing my Vim combos on the keyboard. I still very much like to refactor that function that I find ugly, chisel every variable and every loop on my screen, and then ask Claude to do the boring search & replace for me afterward. Whichever part of coding is fun to you, don't let AI rob it away!