This is a post about using LLMs to write software.

Some people dispute the idea that there are any notable negative externalities to the way that LLMs are being developed, deployed, marketed, and sold today. I am not talking to those people; we live in different realities, and if I’m wrong about that, I’m definitely wrong about my negative stance on LLMs.

Given that there are significant negative externalities to, say, Anthropic’s Claude Opus 4.5, OpenAI’s GPT 5.2, and the models they are currently working on to replace or augment those, I have found it very difficult to understand how people who I generally think have similar value systems as I do can emphatically defend paying these companies money, glazing their products online, and so forth.

I’ve tried arguing my position based on the facts, but it’s useless. The facts are either so obvious that nobody bothers to disagree, or so contentious and unsettled that I’m not sure myself.

So what do we actually disagree on?

I hear mega-boosters say that these tools help them produce code faster than ever; that they can write more software than they ever imagined possible. They say this as if that made it all worth it; as if speeding the planet towards uninhabitiability was a worthwhile sacrifice for computers that program themselves.

Is that what they believe?

I used to think software could change the world. The way I use computers reflects this: decentralized where I can, self-hosted where I can’t. I spent a decade of my life contributing to documentation for Rust, promoting it everywhere I worked, and doing the little bits I could to make the language and ecosystem better. I thought I was doing something real, something to make the world a better place.

I don’t think I believe that anymore. Software is important for some things, absolutely, but I don’t think that more software implies more better.

Is that what they believe?

If so, I get it, but I also think that they are catastrophically, historically wrong.