(2024-05-23) Schroeder Saving The Desktop

Karl Schroeder: Saving the Desktop. This week Microsoft announced its new line of Surface devices, which have a Neural Processing Unit (NPU). They’re building their Copilot AI into the Windows operating system. One of the new features is Recall, which records everything you do for later reference

for years the tech giants have been boiling the frog (us) by increasingly monitoring everything we do.

Chinese company Huawei has been banned from doing most business in the US, partly because of concerns about it taking over the American computer hardware business, but also because of worries that it might be building spyware into its systems

If like me you’re not American, you are always aware that the desktop OSes you use (and most of your software) are built by US companies. Granted the political instability that appears to be gripping America right now, we in other countries find ourselves in the uncomfortable position of having to consider future scenarios where we treat American software companies the way we’ve been treating Huawei: with suspicion or even outright ban.

What I am doing (as always) is building second-order or indirect impact scenarios—as Jerome Glenn pioneered with his Futures Wheel approach to foresight

Linux is a bad desktop OS because the desktop part is bolted onto a clone of a 50-year-old command-line-driven multi-user server OS that only a programmer could love. Linux, like UNIX before it, was designed and built to be used by software engineers

It’s not that Linux isn’t ready for the desktop. It’s that it isn’t a desktop OS.

A Promising Alternative

BeOS. It was first released in 1995.

BeOS lives on, however, in an open-source version called Haiku.

BeOS could do things in 1995 that Windows, Linux, and MacOS can’t do in 2024.

new development tools have allowed developers to port Linux desktop apps to Haiku. This is huge—since Haiku conforms to the POSIX standard, programs written for UNIX or its derivatives are fairly easy to compile into native Haiku versions

The OS itself has been rock solid, although it doesn’t recognize my sound card, and some apps have crashed

a lot of the available software is more at the ‘technology demo’ stage than fully operational

LLM OS?

There is at least one project to build an open-source alternative to Copilot. Check out Fabric or Langchain and you’ll see that we’re rapidly entering a new era in personal computing; like Copilot and GPT-4o, these systems bring together swarms of Large Language Model agents, and using techniques like chain-of-thought they do real-world reasoning. But they do it on your computer, on your terms, privately, and transparently.

What an average user needs is something that, like Copilot, just works. As I argued above, I don’t think Linux will ever be the platform where that can happen

Haiku, on the other hand, is ideal for juggling complex multi-media tasks. It could easily be the foundation for a trustworthy open-source personal AI. And that would be truly empowering.

For that to happen, Haiku has to mature. It needs to become stable enough for daily use (in my experience, it already is) but also be usable on a wide variety of modern systems, and that means more device drivers for, for instance, my sound card.

The foundation for an LLM OS must be laid. This means building core logic into the OS for using NPUs, which is not something one person will be able to do. It would need a team effort.