I'm quite happy refactoring code from time to time with Qwen3.6-35B-A3B, specifically the Unsloth quantized UD-Q5_K_M version.

[…] wants me to clean up the code to look less "AI-generated". Let me think about what's going on here. The user wants me to clean up the code to look less "AI-generated". Let me think about what's going on here. The user wants me to clean up the code to look less "AI-generated". Let me think about what's going on here. The user […]

This model is "antiquated" for modern standards, but you're reading a blog written by someone who "still" uses Sonnet 4.6.

From Aider to Pi

While I've been an Aider-person for almost 2 years¹ now, I've started messing around with Pi. I've even managed to set it up with Nix on both my NixOS and macOS computers thanks to pi.nix by Lukas Leeb.

Aider's architecture makes it harder to LLMs to arbitrarily run code because it creates edits and makes you consent to running commands. On the contrary Pi's a tool that's meant to run commands and perform edits on top. This gives it a large vulnerability surface.

Sandboxing

Given this vulnerability risk, I spent some time learning how to sandbox things properly on NixOS and macOS. For NixOS I'm using Landrun which in turn uses the Linux Landlock API. On macOS, I'm using sandbox-exec, which Apple claims to be deprecated since 2017, despite software vendors heavily relying on it:

The sandbox-exec command is DEPRECATED. Developers who wish to sandbox an app should instead adopt the App Sandbox feature described in the App Sandbox Design Guide. The sandbox-exec command enters a sandbox using a profile specified by the -f, -n, or -p option and executes command with arguments. […]

March 9, 2017²

The sandbox syntax is only poorly documented, but the /System/Library/Sandbox/Profiles directory on macOS has plenty of good examples. You can concatenate strings and define helper methods. It's almost a full Scheme-like language! I ran Pi with sandbox-exec a few times and checked for sandbox violations logs with log stream --predicate 'sender == "Sandbox"'. The macos_sandbox_profile.sb file in my dotfiles holds all the rules needed to run Pi on macOS.

I've made a small command called sand-pi which opens Pi with either Landrun or sandbox-exec depending on the current OS. Note that Pi can't make outgoing network connections on both NixOS and macOS, except to talk to my llama.cpp server on my local network. I realize that there may still be loopholes. Maybe an LLM could store exploits in my code that non-sandboxed tools may accidentally run.

Use cases

I see uses cases for Aider and Pi all the time. LLMs excel at tedious refactoring tasks whose results are reviewable and verifiable for developers. John Goerzen from The Changelog shared in a recent article that updating Debian builds is deterministic and humans aren't adding a lot of value there:

[…] I could see a seasoned developer saving hours, maybe even days, by turning over the mundane tasks of managing trees of cascading dependencies over to a coding tool — and verifying and directing the process.⁸

One of my favorite use cases is to update Python modules with missing translation strings:

print(f"hello, {name}")
# ->
print(_("hello, {name}").format(name=name))

You could use regular expressions or Rope in Python to fix this. But while re stands around revving its s/".+?[^"] ####!!@@ /(…/g, the more advanced LLMs complete the mission.⁷

Running LLMs on your own hardware

I've been happily using OpenRouter for most of my LLM adventures. You can switch between models and try out everything without being bound to a single monthly subscription.

I don't want to enter any direct contract with companies like ClosedAI or Misanthropic. Using their fully integrated services means that you can buy LLM output below cost, for which you in return pay with your soul.⁶

Running my own LLM on my RTX 3090 Ti graphics card gives me freedom and privacy.⁵ I bought the graphics card a while ago, so the only cost is electricity³.

It's the closest thing I've got to having my cake and eating it. If the means of production are software production capabilities, then I own the means of production. Perhaps this lets me express sympathy with Luddites, who were skilled machine operators and opposed harsh labor practices, and weren't against automation an sich.⁴