Florian's Filter — W38 2026

The barrier to creating software has collapsed, but the barrier to maintaining it remains as high as ever. We are entering a phase where the ease of deployment is masking a massive accumulation of future costs. It raises a fundamental question about who owns the long-term architectural integrity when anyone can ship a solution.

Articles

• When anyone can build software, who decides what not to build
  This analysis reflects patterns seen across the board, from startups to enterprises, where local GenAI solutions are deployed quickly while ignoring future costs. It examines the tension between rapid implementation and sustainable architecture - which is not as abstract as it might sound in the first moment. Give this a read, even if you are not a software developer or an architect!
• Models Don't Go Rogue
  Eryk Salvaggio offers a compelling perspective on the OpenAI/Hugging Face hack by stripping away the myth of mysterious intelligence. The piece emphasizes that while models aren't sentient agents, OpenAI remains fully responsible for the failures of its computer systems.
• dario, please!
  This post provides a sharp critique of a recent piece by the Anthropic CEO. The author dismantles the CEO's arguments with a level of logic that makes the original claims seem... well, make up your own mind about that!

Also new here

This week's articles have me circling the same question from a domestic angle. If the hard part of building software is deciding what not to build, the version I actually live with is deciding what not to keep.

I went through my own Mac and found somewhere north of 150 applications. I selected nineteen - apart from the obvious tablestakes like a browser or the M365 suite. They now live on a permanent page: The Mac software I actually use — grouped by what they're for, rated by how much I'd miss them, and with the limitations named rather than glossed over. One entry recommends a free competitor instead.

Going about this "data driven" had its limits: I started with launch counts, assuming the numbers would give me a clear ranking. They didn't. The second-most-launched application on this machine is one I abandoned months ago. A tool I touch dozens of times a day registers twenty-two launches, because it starts at login and then never stops. The numbers found the candidates; I had to talk my way through every one of them afterwards.

Each tool gets its own write-up, one every Thursday from here to February. The first is up: I take screenshots for other people to read — on CleanShot X, and why the friction was never the capture. I hope you'll find this helpful - and please let me know if you disagree!