How I write software for fun in 2026

This post has been all hand-written, like in the golden days of the past.

Big news, software development has changed a lot in the last year.

So I thought it was time to update how I actually write software nowadays.

First of all, there is really no excuse for not using Nix anymore. Your coding harness can insta-shot Nix configurations and fix all sorts of Nix errors, so do yourself a favour and set up Nix on your computer. You can read this blog post from 2023, point your agent at my setup, just do it.

Once you have Nix and home-manager all set up, it’s time to write some code!

So open your terminal and start typing rails new… Just kidding :)

Actually, talking about terminals, you should install Ghostty, it’s made by Mitchell Hashimoto, and it’s super great.

Then you want to install pi, it’s the simplest and greatest agentic harness ever written by man. I like that the creator is a normal human being who still has doubts, and can express them without the pressure of trillions of dollars of backing on his shoulders. Watch this talk which explains the philosophical principles behind pi, then skim through its website. You won’t regret it.

Now comes the fun part: I don’t want to be paying a $20 monthly subscription for software that I write for fun. And also I don’t want to use my work laptop to write personal projects, which only exist to make me enjoy building software. So what gives?

Luckily China exists, and in the last few years they have released open-weights models that are basically (for my fun intents and purposes) equivalent to SOTA models by American labs. But I still feel a bit funny actually connecting my laptop to DeepSeek’s servers in China.

Open-weights. A-ha!

There are many providers based in Europe or North America that host these open-weights models, and using pi you can just use them and tell them to do things. My personal favourite provider at the moment is Fireworks.

The great thing about these Chinese models is that some of them are extremely cheap.

deepseek-4.1-flash-1017 (a 552B-parameter model) costs this much:

On average a short prompt costs me $0.002, that’s incredibly cheap. And even though this is the “standard” version, it’s really fast, averaging 250 tokens per second.

Compare that with an equivalent model from Anthropic, claude-sonnet-5-5, which is much slower, averaging 140 tokens per second:

Moreover, you can easily switch to another model whenever you want, just check the list of models Fireworks supports: https://fireworks.ai/models

There you have it, hope this was useful! 👋