Adventures in collaborative open source development
Motivation Like most people, I've been playing with agents to see where they're helpful, where they're not, and what kind of workflows are a good match for me. One area I've found friction is in iterating on a piece of code - written by me or otherwise. I can describe the relevant section and my question/request in command line chat, but it would be better to do so directly…
Thanks to the Debian 64-bit RISC-V port it's really easy to build a sysroot appropriate for cross-compiling Clang/LLVM and its separate test suite . Either use my rootless-deboostrap-wrapper script or the command I documented in LLVM's cross-compilation instructions , being sure to see the note on working around a Ninja dependency issue . For a bootable QEMU image, Debian-based recipes…
Quite some time ago I shared a script and methodology for performing a cross-architecture debootstrap in a rootless way . I had a short note on producing an image bootable in QEMU, but it was fairly minimal. This page provides a cookbook / quick reference on producing such images across various Debian target architectures supported by QEMU. The goal is that the starting point here "gets the basics…
Simple goal: take advantage of my home router's WireGuard support and have one of my external servers connect using this, and pass all traffic from a certain user through that interface. Create WireGuard credentials This part of the note won't be that useful to you unless you're using a Fritzbox router. But if you're me or someone suspiciously like me you may want to know to:…
Last month I wrote up a fairly long piece on per-query energy consumption of LLMs using the data from InferenceMAX (note: InferenceMAX has since been renamed to InferenceX). Much of the write-up was dedicated to exploring what you can actually conclude from these figures and how that interacts with some of the implementation decisions in the benchmark, but I feel the results still give a useful…
shandbox is a simple Linux sandboxing script that serves my needs well. Perhaps it works for you too? No dependencies between a shell and util-linux ( unshare and nsenter ). In short, it aims to provide fairly good isolation for personal files (i.e. your $HOME ) while being very convenient for day to day use. It's designed to be run as an unprivileged user - as long as you can make new…
How much energy is consumed when querying an LLM? We're largely in the dark when it comes to proprietary models, but for open weight models that anyone can host on readily available, albeit eye-wateringly expensive, hardware this is something that can be measured and reported, right? In fact, given other people are doing the hard work of setting up and running benchmarks across all kinds of…
Although analysing performance by way of instruction counting has obvious limitations, it can be helpful (especially when combined with appropriate analysis scripts) to get rapid feedback on the impact of code generation changes or to explore hypotheses about why code from one compiler might be performing differently from another - for instance, by looking at instruction mix in the most executed…
Recently I jotted down some notes on LLM inference vs training costs for DeepSeek and I wanted to add on an additional datapoint for training cost based on the recently released Olmo3 models from the Allen Institute for AI ("Ai2"). The model family has 7B and 32B parameter models, with 'Think' variants available for 7B and 32B but so far only a 7B 'Instruct' non-reasoning…
I recently had reason to do a quick comparison of the performance of the Hetzner AX102 dedicated server and the high-end 'dedicated' CCX53 VPS on Hetzner Cloud and thought I may as well write up the results for posterity. I'm incapable of starting a post without some kind of disclaimer so here comes the one for this post: naturally the two products have major differences in terms of…