Desktop Is Dead
rusz.space
rusz.space
If it was meant to be provocative then well done you did that, but it leaves me feeling that there's no room for debate because you already know what we all "should" be doing.
It is an amazing show when Life slaps their face and tells them they know nothing, exactly everybody else.
A lot of these ideas have been not only had before but actually implemented and then thrown away, so it's worth studying why they didn't work out. It turns out there's some common themes that you didn't touch on.
1. What you're calling an item based OS was already implemented by both Apple and Microsoft in the 1990s. Apple's version was called OpenDoc and was a more purist version of the data first idea, Microsoft's was OLE and was less so, but they both attempted to rebuild the desktop OS as a primarily data-focused experience in which the app was less central. In OpenDoc you created a new document and then dropped app components into it, and the document, the UI and the file format were all managed by the OS rather than being owned by a specific app. Instead of apps there were pre-canned document templates called "stationary" (yes more paper metaphors).
This tech didn't work out for a few reasons. One was timing: Apple was bleeding money and OpenDoc got cut when Jobs came back, another was that it was very hard to write components for these systems. Those are fixable: the harder one was it was theoretically elegant but the UX was terrible, and it didn't solve a problem users really had. The app-first metaphor made sense to them.
2. Single-language/single-VM OS. Other than Smalltalk, Sun attempted to build a JavaOS and Microsoft ran a decade+ research project called Midori attempting to do the same thing in C#. These didn't work out because: (a) there's no language that is universally good at every task, in particular when it comes to efficiency, (b) the benefits of being single language was less than the benefits of being highly efficient and supporting existing apps written in any language, and (c) the process abstraction does a lot of different things and once you have it the benefits of a shared language mostly go away anyway.
3. Network transparency. So much has been written about this. Old dream, never worked. Network-exposed services are fundamentally different in several ways to in-process or same-machine components.
4. Tagged files. Already implemented since forever in every OS. It's there in macOS, Win 11 and Linux distros. People don't use it, largely because most don't really understand documents or files as first-class concepts to begin with. Nowadays even young people need remedial lessons in how to work with files when they start at university. The more intuitive interface turned out to be app centric, where the user starts an app and then sees their recent/shared documents along with a search bar.
5. Indexed filing systems. BFS gets lionized more than it deserves: its indexing support was very crude and barely usable. Most people who praise it have read about it but not tried to use it as a developer. I haven't either but I at least read the BeFS API docs, which is more than most people have done. Anyway Microsoft invested heavily in the idea of a powered-up filing system in the early 2000s (WinFS), but the project was a total failure and had to be scrapped. Most files are system files not user files, so indexing is of little value, but apps are hyper-sensitive to file access latencies. Doing any additional work on FS hot paths just makes everything sluggish. Also, UX: see (4).
6. Highly dynamic Smalltalk-style VM. Modern operating systems have gone in the opposite direction and lock everything down against changes. One reason is security, another is that apps that hack their way past carefully designed extension interfaces to modify arbitrary stuff cause a ton of system instability and broken upgrades.
7. Intent-based GUIs. This was the XML vision: full separation of data from presentation. It didn't work out partly for random web-specific reasons, partly because XSL:T sucked and was hard to use, but mostly because you don't get much value from separating the two. XML was justified by reference to future AI; the future arrived and the only markup it uses is Markdown i.e. purely presentational. Search engines don't rely on markup much and screen readers - where semantics matter most - seem to be satisfied with lightly decorated node trees.
Sometimes I like to while away some time and daydream about a better desktop OS, but I'm sure it can't be done without buying in 500% to AI and LLMs. For one, any talk of a new way to write apps or GUI toolkit is DOA without addressing the existence of web apps, which are what your user will spend all their time working with and which won't care about anything your new OS does. AI on the other hand could potentially knock up adapters and screen-scrapers on the fly. And if you have pervasive integration of LLMs then things like tagging or indexing become a lot less important to get right, as a sub-optimal UI is only sub-optimal for the AI which can brute force its way through.
If I were to design a new desktop OS today, I'd try and do an "intelligence first" design that assumed pervasive availability of advanced AI, and ask what that lets you re-imagine. I think the result would look nothing like the Alto.
I think if you were serious about something like this, you would go and sell to businesses first: productivity would be at an all-time high. But yeah, in today's world, AI everything everywhere. But I'd argue that makes an even bigger point for an item based OS, as everything is equally accessible for agents.
On (6), Plan 9 had a solution to a similar problem in file system mounts. All FS mounts were in a per-process namespace. Mounting file systems does not require special permissions like on Unix, because these mounts only exist within the process tree that creates them, rather than modifying global state.
Anyway, I'll think about this later, it's 11pm here.
The point about LLMs is that in reality you can't say "software should work like this". Maybe that was viable in 1990, but so much software has been written since that the designs, toolkits, paradigms: done. Nothing can be changed due to the massive weight of pre-existing software that will never be rewritten. I'm middle aged, and the only new OS designs in my entire lifetime have been iOS/Android where the immaturity of the mobile space allowed a clean slate start. But that was a one-off.
AI lets us imagine new things again because it's getting close to enabling something like "rewrite this giant codebase to use a new UI toolkit for me", or "take this web app, reverse engineer its backing REST API, combine it with that other web app and combine it into a new UI personalized just for me". And that then enables experimentation again: you don't have to wait for third party devs to support your cool new OS if you can get an LLM to hack in and crowbar a form of support via brute force.
Other thoughts: do files even matter anymore if you interact primarily via a chat box? Or is the only thing that matters recorded notes by/for the LLM which can then render whatever view of the user's knowledge base is appropriate? Do apps even matter anymore, or can good enough logic be assembled and written on the fly to suit any request? A lot of reasons why various ideas did or didn't work were to do with business models, but the low cost of inferencing fundamentally changes what works there.
I do agree that AI could be used as a rare reason to actually innovate in the OS space.
> do files even matter anymore if you interact primarily via a chat box? Or is the only thing that matters recorded notes by/for the LLM which can then render whatever view of the user's knowledge base is appropriate?
I'd say in an itemized OS with loads of links between items, AI would be more useful than in today's OSes. First, RAG would be faster as it wouldn't have to read everything, instead look for items that are relevant and traverse the tree of links, reading the nodes. Again, metadata. Apple has added stuff like automatic photo metadata (e.g., you snap a photo of a beach, and it automatically gets tagged with #beach; this is also being used by malware to exfil only documents) making AI even more useful.
> Do apps even matter anymore, or can good enough logic be assembled and written on the fly to suit any request? A lot of reasons why various ideas did or didn't work were to do with business models, but the low cost of inferencing fundamentally changes what works there.
This actually made me think of the secretary metaphor: Sure, a secretary can do a lot for you, but not everything. You still have the knowledge you need to act on, and will always know what you want to do more than the AI. If the itemized OS had a common API between app's backend and swappable views, a universal AI view would be very easy to implement.
On the previous comment:
(1) OpenDoc failed for a variety of reasons (https://instadeq.com/blog/posts/why-opendoc-failed-and-then-...), but I'd say the major reason why attempts at this fail is that they're another option of developing apps, and a bad one at that. If it was the only way, and was actually intelligently designed and documented, it might have had a better chance of succeeding.
(2) From what I've read (https://joeduffyblog.com/2015/11/03/blogging-about-midori/), Midori was a research project, and not a replacement. They started with .NET and C# because they own it, but then departed to custom languages. Their research included stuff like language safety (most of it now in Rust, and it seems they implemented some of it into C++ and C#). Also cool concepts like capabilities instead of permissions (2. post), programs that are multiple separate processes (like Smalltalk's; the lessons learned are now in C#'s await keyword), etc. The articles are a nice read.
> there's no language that is universally good at every task, in particular when it comes to efficiency
I'd say that Rust is now in a position that it is good enough for most tasks (even websites and AI), and is efficient enough for all of those tasks. Having your entire stack in Rust (or Elixir or most other languages) enables you to have more consistent software as everything is defined in one place only.
> (b) the benefits of being single language was less than the benefits of being highly efficient and supporting existing apps written in any language
The biggest benefit of being a single language is massive parallelism, especially if you're process/message oriented as Smalltalk or Elixir. Most modern software, unless really pushed (e.g., game engines) still do not use all cores available. There is also an argument to be made that context switching between smaller processes only on messages is faster than switching on modern OSes, and therefore would be more efficient. This would probably have to be implemented to be proven, though. (I'm not as confident on this level of OS design, feel free to disprove me)
> (c) the process abstraction does a lot of different things and once you have it the benefits of a shared language mostly go away anyway.
Yes, but to an extent. It is still hard to write safe concurrent code in a language like C++, and massive parallelism is hard with current black box binaries.
(3) My thought process was a) files mounted over a network act like they're local, b) in a Smalltalk system local and networked objects would appear the same, c) since Smalltalk essentially encapsulates a server, the entire stack is already network like communication. But yes, it would probably only work for certain things, like objects or documents accessible over a network.
(4 & 5) My point was that with an itemized OS, tagging and searching through items (non-system) would be better than with today's files. There is also a question of UX making it non-viable for most users. Searching files in Dolphin takes seconds, while doing the exact same thing in CLI with fd takes milliseconds. The end result should be a fast, accurate search engine for everything in your computer (local and remote; this would also augment AI to make it more useful).
(7) My idea was more swappable views, views shipped by the developer and custom views that a user could make. I'd recommend looking at User-created item views (https://alexanderobenauer.com/labnotes/009/). This also goes hand in hand with bootstrapping - the user changing their environment so it's easier for them to use. That, I think, is enough value to justify it. And if a developer doesn't think it is necessary, nothing should change for them, but the user should still have the option of making a custom UI (this also goes back to accessibility and AI using the apps).
Web apps are a problem, but most are easily replaceable or easily recreated. Specialized apps are a bigger problem (e.g., CAD, video editors). But you could also just link to the web, and link from the web.
Ummm most Smalltalks don't use all cores available, they time-slice a single OS process, aka green threads?
(Exceptions: SmalltalkMT and VisualWorks Matrix framework https://www.cincomsmalltalk.com/main/products/foundation/mat... .)
Also iirc Erlang originally did not take advantage of multi-core cpus or multiple processors for performance. The point of the massive parallelism was "process" isolation for high availability systems.
Only if you explicitly make that particular IO or message send async.
"When a block receives value, it waits to return until all of its expressions have been executed. For example, the following expression does not produce a value until all three clocks have been completely displayed.
[EasternTime display.
MountainTime display.
PacificTime display] value "When a block receives fork, it returns immediately, usually before its expressions have been executed.
[EasternTime display.
MountainTime display.
PacificTime display] forkpage 252 "Smalltalk-80: The language and its implementation." 1983
https://rmod-files.lille.inria.fr/FreeBooks/BlueBook/Blueboo...
~
> From what I've read (https://joeduffyblog.com/2015/11/03/blogging-about-midori/), Midori was a research project, and not a replacement. They started with .NET and C# because they own it, but then departed to custom languages. Their research included stuff like language safety (most of it now in Rust, and it seems they implemented some of it into C++ and C#). Also cool concepts like capabilities instead of permissions (2. post), programs that are multiple separate processes (like Smalltalk's; the lessons learned are now in C#'s await keyword), etc. The articles are a nice read.
Most of the stuff in Midori never made it out of the project at all, partly because a lot of the ideas weren't that great tbh.
This https://www.youtube.com/watch?v=37WgsoZpf3k and a bunch of follow-up videos on the channel shed the light on the technical side of things. As someone working on a somewhat performance-sensitive project, just one of the features - safe immutable and slice-eable buffers from the OS to enable fully zero-copy buffer handling already sounds like an excellent tool to have.
"I say that as someone who used Suckless’ st, dwm, surf with tabbed, slock and dmenu"
As I don't know or even recognize any of that jargon, Perhaps that is the point or the article and I am just old.
https://suckless.org/philosophy/
Many of the projects and libraries listed (st, dwm, etc.) are shown: