LangDev’26

Share
LangDev’26

Apologies for not having published the 10th blog post on MyFPL yet. Last week I was in Spain for holidays but also for the MODELS 2026 conference, or more specifically: the LangDev’26 conference that was co-located with it. And before that, I was preparing for this conference, which left me with too little time to have post 10 ready.

LangDev is an annual conference for language development enthusiasts, typically consisting of 1.5-2 days of talks and a social dinner on the evening of the first day. Historically, LangDev can be seen as a continuation of the Code Generation conferences held in Cambridge until about 2015. Topics range across various aspects of language development, from experience reports from industry to “hard” topics like compiler optimization. Giving live demos as part of your talk is explicitly and heartily encouraged. It’s a non-academic conference: the talks are “just that”, without needing to write a paper first, and no proceedings are published afterward. This year, all talks were video’d (thanks to some sponsoring) — see the link below.

💡
You can find the program with the presentation abstracts here.
You can find the videos of all talks here.

Fifteen of the talks lasted 30 minutes, with the other ten lasting 15 minutes. A 15' timeslot amounts to an extra challenge for the presenter(s). Various esteemed writers made it clear that writing a short, concise text with the same amount of “meat on the bones” is a lot more difficult than just spending however many pages you like. The same is true for talks — maybe even more so. I guesstimate that halving a talk’s length roughly doubles the required amount of time and effort for preparing it — at least to be able to deliver it with the same quality.

There’s an upshot for the audience, though, in my opinion — several upshots, in fact:

  1. The presenter is forced to be brief. Choosing what content is most important, is key. Remove any unnecessary stuff. This naturally – though not always, of course – leads to talks with more focus, and possibly higher quality.
  2. The brevity also helps the audience to focus. There’s less urge to just open your laptop and do some work, thinking you’re able to simultaneously follow the talk — newsflash: research shows you can’t anyway.
  3. Talks you don’t like or understand so much take only half the time to finish.

The downside is of course that interesting topics only get half the time as well, but hey: there’s always the (coffee) break. I think I’d actually like it when more than 33% of all talks lasted 15 minutes.

I won’t go over all the talks separately, but I’ll highlight a couple that stood out to me. (I’ll skip the ones involving LionWeb, since I have a “slight” conflict of interest there.)

First of all, I’ve posed the question whether ALEF – from the Dutch Tax Agency: “de Belastingdienst” – is open-source already many times in this context before. Arjan Oortgiese was able to answer that with a resounding “Yes!” It’s good to see that software that’s useful and already paid for by the taxpayers is (finally) made available to those taxpayers.

Obviously, there was a lot of talk about the use of GenAI in the context of DSLs and modeling. I’m not at all a fan of GenAI, for all the usual and obvious reasons. So, I really appreciated the talk “Domain Knowledge Debt in Metamodeling: When LLMs Act as Oracles” by Alfonso Pierantonio. In it, he explained that research shows that having an LLM just spit out a metamodel – as if it were an oracle – creates something called domain knowledge debt. It’s like technical debt, but it happens in your own brain: because you didn’t actively involve it with coming up with a metamodel, that didn’t really have the chance to become a part of you by expanding and deepening your mental model about that domain — deskilling happening in realtime, as it were. (It was also interesting to see that we, as practitioners, have a tendency to come up with technical solutions for what is essentially a sociological problem: responsible use of the tools at are disposal, in the context of a job that comes with a real-world responsibility.)

On the other hand, it’s also becoming pretty clear that having LLMs “speak DSL” is very promising. Effectively, by forcing an LLM to express deliverables in a DSL – which you’ve taught it – the domain can become verifiable. As long as the DSL comes equipped with a checker that tells the LLM what it did wrong, an LLM will converge to something that’s at least not syntactic and type-wise garbage. Also, because DSLs tend to occupy a higher level of abstraction than conventional GPL code, the resulting DSL content is shorter, and also lends itself to verification by domain experts, by “unit tests”, or ideally: the combination of both. ALEF with RegelSpraak and TestSpraak is a great example: have an LLM formalize “legalese” into rules and accompanying tests.

It was also interesing to see Nelumbo – presented by Wim Bast of the Modeling Value Group – in action. I’m not sure I’d be able to use it in real life, on account of me not being smart enough to grasp the sky-high level of abstraction. In any case, it’s a good example of what beautiful “stuff” sticking relentlessy to your principles can lead to.

The final mention goes to the panel session, prepared, presented, and moderated by Arjan Oortgiese and Johan Blok — the latter unfortunately in absentio due to illness. We discussed a couple of – relatively thorny – questions in a constructive and interactive way. Such sessions can be very inspirational, and this one certainly was. I hope we’ll have one next LangDev as well.

The final-final mention goes to Pedro Molina and Javier Troya for organizing LangDev’26. Hopefully, LangDev’27 will be co-located again with MODELS 2027, which takes place from September 26th to October 1st in Budapest. Not only is Budapest a beautiful city — and for me, it’s also reachable by train in 15-16 hours.