Who's got two thumbs and decided this year's big project would be to get 16 governments to sign a multi-million euro contract together to establish a regional bikesharing system? This guuyyy. The good parts are that I'm essentially in the lead for deciding how the system will look like, as I've written of the requirements specification and partnership agreement. There's a blend of pragmatism, idealism and knowledge of the abstruse ways bikesharing works that I'm proud to have been able to put into it. The downside is that when you run a pack of 16 governments in a particular direction, there's a significant risk of the pace being defined by the slowest of the pack. And the biggest municipality in this ordeal has been plodding along with a breathtaking lack of competence even after escalating it to the highest level possible. I think we'll find a way? But I'm not as sure about that as I'd like to be.

I vibecoded a game this week! Inspired by mike's game I wondered if there is a game like that that I'd like to play and make. Being a bit of a public transit ("OV") and of course mapping nerd, I ended up with the idea to take GeoGuessr and apply that best-of-5 game mode to guessing the location of Dutch train stations based on the name alone: OVGuesser. If you're really good at triangulating you might get some points, but I've made it kinda hard on purpose. The combo of Replit and Claude 4 meant I had a prototype up and running in, like, half an hour? Most of the rest of the time was me finnicking with the UI and UX to get it to look and feel like I wanted it to.

It did break spectacularly when I shared it (first on Linkedin, because who isn't on Linkedin for distractions?). Since I never said I wanted the game to be scalable, the AI decided to give each game session its own sessionID stored locally. Which worked fine when I played it solo but caused every single person to be unable to progress past their first guess to the next station. Cue me being unable to panic-fix that issue. In the end I realized pretty quickly that scalability must be the issue. Refactoring AI-generated code is...not great at the moment. Claude 4 is very cheery and happy about everything it makes, and for sure not critical enough about its own capabilities. In the end I just got Gemini 2.5 Pro to act as the sceptical senior dev, double-checking everything to make sure the refactor actually solved the problem. That worked much better than I expected.

I'm still kinda surprised that the word vibecoding was coined just this year. It is very obviously an imperfect system? But man is it fun and quick and good at side projects like this.