Finding Joy In A World Of Clankers

Note: the following piece is about my personal experience of programming with LLMs. It’s not at all intended to be a blanket philosophy of LLMs and humanity, or reflective of the massive sociological, ecological, and cultural problems that LLM usage is creating in our society.

Note: I did not write a single world of this with an LLM. I use semicolons and em-dashes in real-life. Stop it.

For most of the last 4 years, I’ve hated LLMs. Well, that’s not quite fair; I hate the people who talk endlessly about LLMs, about this prompt and that skill set, about the latest way to burn billions of tokens without contributing anything useful to their field or human life. What I hate most of all is the gleefulness with which they proclaim the death of programming, the atrophy of their skills (which I often doubt ever existed in the first place), as if they’ve spent their whole careers in software anxiously awaiting a time when they no longer had to actually practice their craft at all. I have no understanding or sympathy for these people.

But I don’t really hate LLMs. For most of their still very short history, I’ve hated the experience of working with them, but that’s not quite the same. For most of the last few years, LLMs have produced shockingly bad code, and it’s been much easier to just write it yourself, if you care about quality at all. This has meant that most of the time, I exist in basically one of two modes: total apathy about what I work on, incapable of valuing the software that I’m creating, because I’m using an LLM, or, if I do value what I’m working on, not using an LLM at all in any capacity. Most of the time this has meant a total refusal to use LLMs on my personal projects – not because using an LLM would invalidate them in some way, but because I was incapable of using LLMs without giving up on the quality and care that I want out of the things that I make. I want to be proud of my work, and for most of the history of LLMs, I haven’t been able to use them while being proud of my work.

But over the last year, that’s changed to a degree, and I want to write a bit about why.

First of all, the models have gotten a lot better. This results in two main things in my experience: firstly, I don’t spend 99% of my time, when engaging with the LLM, in fighting with the horrible code that it produces. I find that if I express how I think code should be written carefully enough, and if I write enough of the code myself (and of course carefully review the model’s output), I’m able to produce the software that I want, to the same degree of quality that I would expect of myself if I was writing all of it by hand. This is quite satisfying, because I don’t have a lot of time. I have young children, and a lot of personal and professional responsibilities, and the fact that I’m able to get personal programming projects done again is satisfying. I love programming, and it makes me happy that I’m able to do it more with the limited time that I have.

Secondly, and perhaps because of the first, I feel like I’ve gotten a lot better at using LLMs while not giving up control. As I mentioned above, for a long time in the early days of LLMs, I was either vibe coding out of total apathy and nihilism, or else just not using LLMs at all. Over the last year, I feel like I’ve been able to carve out a middle ground, where I use an LLM to perform some tasks in my codebase, while still writing code myself (and of course, carefully reading the model’s output). I feel in control of the quality, I feel in control of my own understanding of the code, and I feel more like I’m doing the kind of work I did before LLMs, just amplified – I’m able to get the ideas in my head out into the files faster than before, and by and large, that’s a satisfying experience.

All of this isn’t to say that working with LLMs isn’t still incredibly frustrating a lot of the time. They’re still incredibly stupid in a lot of ways. I don’t really trust people who don’t share this frustration. The people who are YOLO’ing hundreds of pull requests a day into production, while talking about how nobody reads and writes code anymore, while gleefully proclaiming the end of software engineering as a discipline – I hate these people. I feel genuine bitterness in my heart when they talk about the craft that I fell in love with and practice every day. In order to earn my trust these days, you need to be able to acknowledge how terrible LLMs and the software they create often are. A few people that do this really well, and that I’ve found a lot of inspiration in over the last year, are Mario Zechner and Armin Ronacher. Thanks for building pi, thanks for caring deeply about your craft no matter what tools you use, and thanks for bringing sanity to this whole stupid online conversation, guys :)

Addendum: Tool Bias

I don’t want to fall into the trap of saying “LLMs aren’t the problem, they’re just a tool, it’s the people that are the problem.” Of course this is true to a degree – if you produce garbage, that’s your fault, whether you used an LLM or not. But tools have biases, and the bias of LLMs is to phone it in and not care about your work. When you have a hammer, everything looks like a nail, and when you have an LLM, everything looks like an excuse to hit approve on the PR and scroll twitter for a while.

I don’t know the solution to this. But I want to acknowledge that it exists, and that dealing with the perverse, anti-human tendencies of LLMs is going to be one of the great struggles of future generations.

Addendum: Craftsmanship

I think that I would be remiss writing all this without leaving some thoughts about craftsmanship, or humanism in this world of clankers. After all, the point of all this isn’t software. Nothing we’re building actually matters in an ultimate sense; the value that it has comes only from the degree to which it improves human life or, like art, reflects the human experience and soul in some way. Some people argue that LLMs destroy the point of writing code by hand because they can achieve the same result but faster; I disagree because I don’t think that “the result” is fully determined by the bytes of code in text files; it also in some degree comes from the effort and care put into it by a human.

I have no interest in listening to music created by an AI, because the entire point of music is that it’s a communication medium between humans. It doesn’t matter if it sounds the same. If an LLM was capable of producing the most perfect fugue (which in some cases, they now are), I still would never listen to it instead of Bach, because the point is that Bach, a fellow human now long dead, is communicating with me through his music. The sound isn’t the point; the person is.

I think there’s a degree to which this is true of software as well, but it’s not quite the same. There’s a place for generated software in a way that there’s not a place for generated music. Music is “pure” art, whereas software’s value, I would argue, is only partially artistic, but also partially utilitarian. I think that the process of building software with LLMs is also more human than a lot of other applications of LLMs. After all, I can write code alongside the LLM, edit the code that it produces, carefully refactor until the code reaches just the state I had imagined it in my head. This is not quite true of art, at least as I’ve seen it used.

I write this to say that I don’t think that using LLMs in and of itself destroys the human craftsmanship of a piece of software in the same way that it does of music or visual art. Of course, if you use an LLM to generate software without yourself caring about that software, without caring about the particular lines of code and data structures that it uses, without knowing yourself how it should be built, naturally that reduces the human value of the software (and I think that’s bad!). But I don’t think that LLM usage in and of itself destroys the human value. You can use LLMs and be a diligent craftsman, or you can use LLMs and be a slop monster. Your users, and your fellow programmers, will be able to tell which you are by the care evident in what you create.

Addendum: Realism

If I could press a button and destroy LLMs from the face of the earth, I would. But I can’t. I think their existence is inevitable, at least in the profession that I’ve chosen to use to pay the bills, so I’m trying to think and write about how to find joy in that world.