On AI Coding and Its Discontents - Cal Newport
“Writing your own code, slowly but surely, and using LLMs for narrow or particularly annoying tasks (say like writing tests or throw-away scripts), is the best way to produce the highest quality code, since it’s the only way to properly understand it.” […]
This last year has been exhausting. The PR departments of the frontier labs have done an excellent job convincing us that AI developments are occurring at an astounding, world-changing rate. But if you zoom out, it becomes clear that almost every “breakthrough” since last summer has concerned the narrow domains of computer code and math, which are defined by highly structured languages and come accompanied by massive amounts of specialized training data.
And yet, even in this best-case-scenario setting for AI, we’re still struggling to figure out how to actually use these tools in a way that makes sense in the long run.
Writing production code is tricky in that it can not fail. In this engineer’s case it broke production twice.
The thesis of this article is that it’s ok to use AI to write code, which if it breaks does not cause to too much of an issue. Which has been my experience too. I have primarily used Claude to build personal tools, and my website and some scripts and what nots at work. Also, how much of code out in the world is truly code that can not break?
But I know of people who are talking to their agents and having it build things.
The one thing I agree with is - that we are all trying to figure out the steady state of using AI in development work.
The Human-in-the-Loop is Tired
If you're feeling like your work is simultaneously more productive and less satisfying, you're not broken. The feedback loop is broken. And I think we need to start treating that as an engineering problem in its own right, not a personal failure.
The people who are finding their footing right now seem to share a few traits: they have strong opinions earned through practice, they can distinguish between principles that still apply and habits that were just bandwidth constraints, and they're willing to evolve their workflow without abandoning their standards.
Understanding is the new bottleneck by Geoffrey Litt
You can learn what the agent is doing to make sure you can be an active participant in the creative process.
I have been in this scenario at work and in personal projects as well, where I don’t understand what is going on.
What I usually do is ask the model to create documentation explaining the changes. Not for everything though.
Geoffrey talks about three methods to understand better -
- Explanations
- Micro-worlds
- Shared spaces
Explanations is a skill that I plan on using going forward.
What do coders do after AI? - Anil Dash
But there are people who have spent decades honing their craft, committing to memory the most obscure vagaries of this computer processor or that web browser or that one gaming console, all in service of creating code that was particularly elegant or especially high-performing, or just really satisfying to write. There's a real art to it. When you get your code to run just so, you feel a quiet pride in yourself, and a sense of relief that there are still things in the world that work as they should. It's a little box that you can type in where things are fair. It's the same reason so many coders like to bake, or knit, or do woodworking — they're all hobbies where precisely doing the right thing is rewarded with a delightful result.
Compared to those who see this just as a job, are not passionate about it, and seem like people who will be replaced by AI.