0
Framework OS
By Joe Petersonin
Have you ever tried to buy a fitted suit online? You can't tell if it fits, can't see the stitch work, almost nobody does this so whose review do you trust?
Yeah that's what it's like trying to pick a linux distro as a Windows or MacOS normie.
Why not go with Ubuntu? Well, it doesn't support Nvidia drivers out of the box. What about Pop OS? That's an Ubuntu fork with Nvidia drivers! The problem is that those drivers change frequently and break your computer. Okay, well, fucking update them then, you stupid idiot. True, or I could just use MacOS or Windows.
WAIT -- An AI harness can handle Linux's troubles. In fact, functional networking is all Hermes needs from the operating system to fix the rest.
Karpathy made the case that a ~1B param model with strong reasoning could discard its knowledge-containing weights and simply make heavy use of RAG. It's a barebones vision, and judging by Nvidia's 2026 Computex presentation, with its grand emphasis on CPU-GPU-Storage bandwidth, Huang seems to agree (at least directionally). The model needs to *think* and the cornucopia of external knowledge carries the rest.
Operating Systems can now surely head in the same direction. Whatever drivers, features, or aesthetics the user wants of their operating system can now be organised by their AI harness.
Consider the gentleman in the video below, who between 1:20 and 3:30 describes a driver issue that he ran into when installing Bazzite OS. A Linux driver loaded too early during setup, so Bazzite would see the media driver pop up but fail to form a connection. He could manually load the driver through the terminal and force the connection but the WiFi driver was initialised before the USB adapter was ready. It's an esoteric, niche, edge case that prevented Linux from attaching the driver during setup. He fixed the issue by creating a service that attached the driver later in the boot process.
What do you think you would find easier: convincing a normal MacOS user to figure that or convincing them to quit smoking at the age of 80?
It's a non-issue with Hermes; frontier models can figure it out.
Open source operating systems are weird and painful for people who do not have domain expertise in (or a mind for) this realm of problems. But AI harnesses could be the best thing to happen to open source adoption since Linus Torvalds or Windows 11. Harnesses reduce the switching cost from a skill issue to a word-of-mouth issue. People are going to come to trust (and soon thereafter, rely upon) their AI agents for their interactions with computers completely. It'll be a shift akin to the move from TUI to GUI operating systems.
The operating system is going to be as duopolised in four years as the furniture business is, today. There will be big players who ship good-enough, modular things, but most users are going to possess a menagerie of items from different makers. Brand loyalty will die a dismal death for two reasons: 1. Humans will barely interact with the operating system (their agent will become the primary consumer) so humans won't care, and 2. humans will prioritise smooth transitions between hardware so they will simply port whatever crazy-quilt of a setup they already have over to new silicon and let their agent continue being a software homemaker.
Neither artificial nor wet intelligence is going to pay Microsoft $100 or $200 per year for an esoteric system the human doesn't understand and doesn't interact with, while there are free versions standing at the ready, Not when the human's trusted harness tells them to swap to the open source alternative because that would give the agent better finesse over its goals. Users are going to give their passwords, credentials, and bank details over to their agents soon enough. They won't care at all about Windows or MacOS.
Nvidia's new announcement on Monday (1 June 2026) -- that Nvidia will finally begin competing competently with Apple Silicon on local inference, is an accelerant to the change. Local intelligences will handle 'all things computer' for those people who are only tolerating computers today. The proper operating system for the future is a good, solid, stable foundation layer that amicably plugs into third-party modules. The open source space is going to explode because of a translation in adoption more so than supply (in the near term). Harnesses will tinker with, and improve on, code bases that they use to get their jobs done.
Managers and executives at Microsoft probably think that their alliance with Nvidia for RTX Spark is going to cement their position in home PCs. But they don't control the bandwidth between their customers and their competitors anymore. Models are trained by foreign labs and Nvidia has placed the harnessed agent at the centre of their business strategy. Customers will be tempted, by their agents, to save the annual Microsoft tax and switch to an open source alternative that their agents will happily manage in full. Nvidia's GPU-CPU unity and vendor collaboration just gave the open source space a tight target to attack.
On Monday, Windows' heartbeat ceased and we're now in a transient grace period before its cells stop metabolising.
Yeah that's what it's like trying to pick a linux distro as a Windows or MacOS normie.
Why not go with Ubuntu? Well, it doesn't support Nvidia drivers out of the box. What about Pop OS? That's an Ubuntu fork with Nvidia drivers! The problem is that those drivers change frequently and break your computer. Okay, well, fucking update them then, you stupid idiot. True, or I could just use MacOS or Windows.
WAIT -- An AI harness can handle Linux's troubles. In fact, functional networking is all Hermes needs from the operating system to fix the rest.
Karpathy made the case that a ~1B param model with strong reasoning could discard its knowledge-containing weights and simply make heavy use of RAG. It's a barebones vision, and judging by Nvidia's 2026 Computex presentation, with its grand emphasis on CPU-GPU-Storage bandwidth, Huang seems to agree (at least directionally). The model needs to *think* and the cornucopia of external knowledge carries the rest.
Operating Systems can now surely head in the same direction. Whatever drivers, features, or aesthetics the user wants of their operating system can now be organised by their AI harness.
Consider the gentleman in the video below, who between 1:20 and 3:30 describes a driver issue that he ran into when installing Bazzite OS. A Linux driver loaded too early during setup, so Bazzite would see the media driver pop up but fail to form a connection. He could manually load the driver through the terminal and force the connection but the WiFi driver was initialised before the USB adapter was ready. It's an esoteric, niche, edge case that prevented Linux from attaching the driver during setup. He fixed the issue by creating a service that attached the driver later in the boot process.
What do you think you would find easier: convincing a normal MacOS user to figure that or convincing them to quit smoking at the age of 80?
It's a non-issue with Hermes; frontier models can figure it out.
Open source operating systems are weird and painful for people who do not have domain expertise in (or a mind for) this realm of problems. But AI harnesses could be the best thing to happen to open source adoption since Linus Torvalds or Windows 11. Harnesses reduce the switching cost from a skill issue to a word-of-mouth issue. People are going to come to trust (and soon thereafter, rely upon) their AI agents for their interactions with computers completely. It'll be a shift akin to the move from TUI to GUI operating systems.
The operating system is going to be as duopolised in four years as the furniture business is, today. There will be big players who ship good-enough, modular things, but most users are going to possess a menagerie of items from different makers. Brand loyalty will die a dismal death for two reasons: 1. Humans will barely interact with the operating system (their agent will become the primary consumer) so humans won't care, and 2. humans will prioritise smooth transitions between hardware so they will simply port whatever crazy-quilt of a setup they already have over to new silicon and let their agent continue being a software homemaker.
Neither artificial nor wet intelligence is going to pay Microsoft $100 or $200 per year for an esoteric system the human doesn't understand and doesn't interact with, while there are free versions standing at the ready, Not when the human's trusted harness tells them to swap to the open source alternative because that would give the agent better finesse over its goals. Users are going to give their passwords, credentials, and bank details over to their agents soon enough. They won't care at all about Windows or MacOS.
Nvidia's new announcement on Monday (1 June 2026) -- that Nvidia will finally begin competing competently with Apple Silicon on local inference, is an accelerant to the change. Local intelligences will handle 'all things computer' for those people who are only tolerating computers today. The proper operating system for the future is a good, solid, stable foundation layer that amicably plugs into third-party modules. The open source space is going to explode because of a translation in adoption more so than supply (in the near term). Harnesses will tinker with, and improve on, code bases that they use to get their jobs done.
Managers and executives at Microsoft probably think that their alliance with Nvidia for RTX Spark is going to cement their position in home PCs. But they don't control the bandwidth between their customers and their competitors anymore. Models are trained by foreign labs and Nvidia has placed the harnessed agent at the centre of their business strategy. Customers will be tempted, by their agents, to save the annual Microsoft tax and switch to an open source alternative that their agents will happily manage in full. Nvidia's GPU-CPU unity and vendor collaboration just gave the open source space a tight target to attack.
On Monday, Windows' heartbeat ceased and we're now in a transient grace period before its cells stop metabolising.