I am making it my 2025 resolution to come back to writing a bit more. And to start, let’s a have a chat about my latest conference experience, ACCU 2025.

Back in Bristol

I have been a regular speaker at ACCU since 2018, a streak only interrupted by the cancellation of the 2020 edition (although one may argue a cancelled edition doesn’t count). Over the years my natural french distaste for anything british has been mellowed by this annual ritual and the lovely people who attend it.

Bristol was really nice this early April week and for the first time I made the trip by rail rather than air. Not from Stockholm, mind you, but from Paris. The Eurostar was of course late and I had to make a run to Paddington but I managed my connection to GWR with a couple minutes to spare. A similar experience on the way back had me running to Heathrow due to another train delay, but again by some luck my SAS flight was delayed half an hour so I could return safely into the cold of Stockholm.

The Good Parts

There is one major topic about this year’s edition I would like to get into, but first let’s talk about all the thing that make ACCU great. For those who have never been, it is a bit hard to pinpoint exactly, but there’s a special je ne sais quoi that sets this conference apart. There is some sort of cozyness to the experience that I do not find in other places. Familiar faces show up every year, of course, but it’s more than that. I usually get a feeling that this where the various generations of the british computing world gather, from the young graduates to the most venerable programmers I ever met (I believe one speaker was maybe double my age, and I’m turning 41 tomorrow).

This being a tech conference, I of course have to give a shout out to a few of the great talks I saw. The second keynote about C++ exceptions by Khalil Estell demonstrated some first hand research and hard numbers of the cost of exceptions and dispelled quite a few myths. The closing keynote by Matt Godbolt reminded me a lot of the now famous Rich Code for Tiny Computers talk by Jason Turner (it’s famous right? 255k views is pretty good in my book) and a nice side-by-side comparison of how more modern dialects of C++ can make your life easier.

Another very safe bet when picking which session to go next, Kevlin Henney had two talks I would recommend catching up once they are live, the first about project management and agile, and the second about programming languages and how the market for them seems to have more or less solidified in the past years.

Finally I would like to give a special shoutout to Matthew Dodkins’ talk that will sadly not be available on replay (I believe he requested it not to be filmed). I felt this was a very brave and touching performance from a speaker who was clearly passionate about his topic but visibly suffered from impostor syndrome. I hope the feedback he gets encourages him to do it again, and in general I think this is the kind of talk that makes ACCU stand out from other conferences by bringing a very personal and human touch.

And speaking of emotions (or the lack thereof), this is a perfect segway to the main reason I decided to get back to trip reports.

AI. God damn AI

This edition had not one, but two keynotes about generative AI and LLMs. Not just talks (where you can decide between that and 2 or 3 other topics) but full keynotes. You show up the first morning, everyone in the main room, the lights go out, the stage lights up, and you get an hour of sales pitch for some generative AI tool. And two days later, you do it again. I obviously have several issues with this.

AI is a polarizing topic to say the least. Writers and artists I have had the chance to work with hate it, but even in the tech sector it’s not an uncontroversial topic. With that mind, my expectations for a talk about AI would be that it at least tries to address some concerns and criticism that have been made about it. But that didn’t happen.

Both speakers were employed by either AI companies or companies that sell AI tools. That in itself would be enough to make me suspicious of their claims, because of course their boss wouldn’t give them the right to put their logo on the slides if they spoke ill of the tech they’re making money of.

The discourse during the talk was mostly in line with my worries. The common critics brought up against AI were not addressed (or brought up at all, really). The discourse was very FOMO: “if you are not using AI today, you will be left in the dust tomorrow”. That is not a keynote, that’s a sales pitch, and not a very good one. Since we’re here, let’s talk about the things I would have liked them to bring up and discuss or challenge.

First, the price. Something like Claude Code costs about 5 dollars per session. That’s not a lot for a company, but that’s not nothing. And here’s the thing, none of those companies are making money at the moment. The news articles keep talking about their revenue, which is impressive, but the costs are even bigger. Their models are getting more expensive to train and run each iteration, and yet they keep promising it’ll get cheaper in the future. When? How? No idea.

Secondly, the code they generate isn’t very good. The C++ examples shown in one of the keynotes for example generated code with a bunch of naked new and delete. We’ve had unique_ptr for 14 years now. Another live example shown that you need to first run the AI, then immediately run clang-tidy to fix the decade(s) old idioms it’s using. One of the issue is that the models learns from a large dataset that is mostly unedited (something akin to parsing GitHub or StackOverflow or whatever website they can grab). And as it turns out, the average or median code on the internet isn’t very good. And with more and more “vibe coding” being pushed out to the world, I don’t expect the corpus will get better over time.

This point was partly countered in the talks by suggesting this should be used for code you don’t really want to read or maintain. And of course they had to mention build files. Those who have been following my articles and talks for some time might recognize my usual catch phrase “and furthermore, I think your build should be destroyed”. I spent some amount of energy convincing people that the reason why their build pipeline is bad is because they never really cared about it, often it boils down to copy-pasting some CMake found on the internet and praying it will compile and link, then move on.

This attitude, to me, is the crux of the problem and something we as engineers should talk about. Is it ok to give little to no care about quality of implementation because it’s done cheap and “good enough”? The Spotify app running on my computer at the moment is 8 processes for a total commit size of 840 MB of RAM. It’s also not even playing a song. Would you say this is good and efficient software? This isn’t even to say Spotify codebase is done by AI. AFAIK the issue is that the interface is a chromium app, which are notoriously inefficient at what they do, but they are easy to make because it’s all Javascript and HTML. Generative AI “coding” just another nudge on the quantity vs quality slider, allowing us to generate bloatware at scale. It’s also a really weird to be doing a keynote at a programming conference extolling the virtues of giving up coding.

Finally, in one of the examples given during the talks, one model refused to work and replied to the prompt that “using an AI assistant is detrimental to your growth as an engineer you should be understanding what you are making”. This was brought up as a joke and the only takeaway was “it’s been flagged as a bug, it will be fixed”. Again, the AI used in the demo brought up a more important point than the entire talk it was used in and this was not addressed or reflected upon at all. We are already seeing some employers hoping they can fire all their juniors and replace them with only seniors and LLMs. Maybe it’s time to have a talk about how our profession is gonna work if there’s less and less entry level positions to gain experience from.

I could continue on about the ethics of AI, or the fact that their CEOs are on record saying that their business model only works if they are allowed to train on any and all content, including copyrighted ones, without paying any royalties or getting the author’s permission but I think I made my point.

Closing thoughts

I don’t want you to come off this article thinking ACCU is a bad conference that you shouldn’t join. I still think ACCU is an amazing conference that is well worth the trip. I’m not even sure the organizers had a big say in the keynote topics (and certainly not in the fine details). My point is more that a topic like generative AI should be given more depth and thought in a keynote that somethings that felt like a sales pitch.

See you in Bristol next year, I hope!