FOSDEM 2025

February 2, 2025

I went to FOSDEM in Brussels this weekend. It was my first time attending, and now, at last, I know what it's all about. FOSDEM was 'the one that got away' for the past few years with other trips and events coming first - not this year! My main takeaway was a deeper appreciation for open source and a resolve to integrate it (even) more into my life. On a personal note, I’m typically occupied with responsibilities at conferences, so having the freedom to attend talks and explore such a varied schedule was a rare and welcome experience.

Some notes on talks I went to:

  • From Queries to Pints: Building a Beer Recommendation System with pgvector
    • This was my favourite talk!
    • I learnt the name for an 'Embedding Model' (I've never studied machine learning). I have been looking for this term to better research some things I'm keen to do in my self hosting storage console project.
    • I really liked the description of the process of going from some random data, through processing it, to querying it in fun ways. I think the process outlined in this talk is pretty generally applicable and a fun way to explore all sorts of data.
  • Building a watt-meter esp-rs and a rocket backend
    • The "one new thing" 'rule' for side projects where you only try one new technology in each project is a way of life.
    • I appreciated the close attention to detail in the costing of the project. This is often overlooked, even in projects which are started solely to save money.
    • This talk felt very much like the spirit of FOSDEM, a bunch of interested people living their lives and scratching itches along the way.
  • A database for your program state
    • The takeaway here for me was that the eventing state store could be helpful in our Language Server implementation of Regal. Rather than placing jobs on queues from all over the place, we might instead be able to do things on change events.
  • Pick My Project! Lessons Learned from Interviewing and Writing 20+ End User Case Studies
    • I've spent a good bit of time working on the content on The OPA Ecosystem. This talk was about the data behind Cilium's comparable content where they showcase user stories by use case etc.
    • The suggestion to focus on specific user needs was in website copy was also food for thought, it's not about what it does, but what it does for the user.
  • CLI Design for Designers and Developers
    • This got me thinking about how tricky 'above the fold' output is in a CLI context where output is read backwards if it's long and forwards if it's short.
    • I also liked the suggestion of documenting the main contact points and paths in --help output, rather than just listing all available commands and options.
  • Kernel support for Mobile Linux: The missing 20%
    • I really appreciated that people are working on making more open mobile phones. I was interested in this talk how the speaker (Luca Weiss) made complex tasks sound approachable, and that just having a little more demand would make some of the grunt work more appealing.
    • I bought a secondhand Fairphone 5 to see what I might actually miss.
  • History and advances of quantization in llama.cpp
    • Some hot takes in here e.g. if we had more memory bandwidth, CPUs would be more viable for model inference & the promise of INT8 quantization methods.
  • Every ISP Needs To Use A QoE Middle-Box On Their Network
    • This talk found me about a year after I had issues playing GeForce NOW on a 1gb DOCSIS connection and first dug into Bufferbloat, CAKE etc. It's an important topic and I hope that some people go and install OpenWRT after watching this talk.
    • This is an issue close to my heart and remedies for Bufferbloat are still important when working from my parents in rural Scotland with pitiful upload bandwidth.
  • LLM Tool use in vLLM
    • LLM tool use is a use case I'm really interested to play with myself, but I think it's simple enough I can build my own system for it as the main factor seems to be a suitable model.
  • Go in the Nix ecosystem: vulnerability scanning and experiments towards a next-gen builder
    • I was not aware of GOCACHEPROG from 1.24 and learnt a bit about it in this talk and the extensibility it offers to reproducible build systems like Nix.

Other talks I went to and learnt from, but don't have any specific comments on.