I would replace the start menu with a Raycast clone. Most people want to open files, ask a question or run an app they know the name of, not scroll a list of things they might never use. If you add in semantic search users don't even need to know the app name.

It's also a perfect use for local System One models that can classify prompts and forward them to the right module - calculator for math, distrobox for shell commands, LLM for natural language questions and normal search for everything else. Not using a fixed classifier because plugins should be able to add their own question types.

The article mentioned WinFS and relational filesystems. That's another good idea, integrate metadata so you can support use cases like finding all pictures of your dog or all playlists with Ed Sheeran songs (to remove them). Everything bound to the file entries instead of ID3 tags inside the file or app specific databases. It would be a lot more reliable and probably faster.

IIRC bcachefs is already a relational database but I don't know how good the query APIs are. If we can't do full extendable tags then at least add user defined groups/tags like Work and Games and Photos to make finding things easier. People remember what a file is better than where it's stored.

> I would replace the start menu with a Raycast clone. Most people want to open files, ask a question or run an app they know the name of, not scroll a list of things they might never use. If you add in semantic search users don't even need to know the app name.

So I’m looking at the Raycast web site, and aside from the UI being a little different, Raycast seems like it is itself a clone of the already long-existing KDE Plasma start menu and KRunner? I guess the only thing missing from the KDE one is “AI”. Is there something I am missing about how Raycast is some innovation not already realised in the Linux desktop?

Good suggestions yes. Strong ideas. Ultimately the desktop has been content to let the apps take over more and more, while not extending itself. Actually making the OS something that actual activity happens on feels absolutely essential. Making a Raycast/Vicinae (the pretty great open source port) integral feels like it would go a long way towards this end, of offering a consistent way to expose & access the toolbox.

I'd love to see an agentic OS where the OS and apps can both talk in terms of MCP things like resources, tools, skills. Yes these are primarily geared towards agents, but they are also acceptable machine-to-machine interfaces to script against too (so long as the type signature remains compatible).

Android's Intents feel related, similar-ish. It's a pity no one bit on PK's WebIntents. That could have been an interesting way to create abstractions for intents, for activities, for stuff happening on the computer. Notably the xopc actually had activities and a journal that let apps contribute meaningful intent to the OS as a whole. But mostly it just ended up as a history viewer there. Android mostly uses intents as a launcher. But this idea of having the activities of the system as available abstractions could go much further.