It is, but an already niche operating system that refuses to leverage the advantages of LLMs is only going to fall further and further behind. I say that as someone who has actively contributed to the BSDs since the 90s and would hate to see that happen.
The slop problem is very real. People are spewing reams of bad code that generates entropy faster than humans or LLMs can clean it up. LLM best practices are not yet obvious, and we're still figuring out how best to leverage them while maintaining code quality and human understanding.
... but LLMs are coming whether we want them or not. Models keep consistently improving, the slop keeps getting better, and human-only development is going to struggle to keep up.
In my day job, I'm driving an enormous number of tokens at Astra+Ultramax solving problems in Lean 4 that previously we never would have attempted simply due to the incredible human engineering cost; we review the high-level lemmas and let it prove the rest.
LLMs are the only solution I've seen over these decades that could finally close the feature/usability gap between mainstream and more niche operating systems -- software support, features, drivers, and other broad usability barriers. All of that has historically been far too much work for a much smaller developer community to reasonably undertake.
one of the reasons that BSDs are less popular is because of the community advice of "The best supported machine is a ten + year old thinkpad."
so yeah you're right there isn't a finish line, but you do need to keep up with current stuff enough to give a possible fan base a decent selection of hardware to run the thing on.
in other words : if the companies that the BSDs already had a hard time keeping up with hasten development via LLMs the gap will grow ever wider until a critical point is hit and the thing becomes irrelevant.
if linux is the only show in town that supports flashy-hardware-Z then it's going to drive audience migration.
the thing that has allowed openbsd to adopt new things so slowly is a rock-solid set of core competencies; the amount that bsd is entrenched into the internet itself makes irrelevance hard to imagine -- but it's still a facet.
I still hear people bitch that BSDs don't support WiFi, and that hasn't been true in years and years.
Think about how far Linux has/hasn't come in the last 10 years. The user experience is nicer, drivers "just work" more, there's more support for more stuff...but at the end of the day, the actual computing experience is mostly the same today as when I installed Debian back in 2016. It's polish, interoperability, etc that's improved.
Maybe VR takes off and takes over the desktop, and then openBSD can't keep up with that, but I don't think that's likely. So what are they trying to "keep up" with?
The number of OpenBSD contributors, since 2017 is going down. For all the praise the system gets here, the developer team is small. And feature-wise the system is a toy compared to Linux. It's like comparing DigitalOcean with AWS.
Ppl praise the stance of OpenBSD/NetBSD and what-not. That's commendable, but then close the tab, open their spotify and get back to agentic programming for their proprietary company source. But God-forbid if OpenBSD turns anything more than a museum OS, full of purity but with a user-base of five. Because the GPU doesn't work properly, the sound drive is clunjy, wifi might work, etc.
LLMs could port huge amount of drivers to both system with a fraction of the effort. Specs can be implemented easily and properly. Then of course, you need to test and that takes time, but it's technically easier and more ppl could contribute.
I get the feeling, but they're not OpenBSD a favor, if anything else I feel the exact opposite is happening.
Ironically with sloppy code we will be back to 2001 with non-concurrent OSS and shitty sound daemons as esd/arts, now merged into pipewire but with crappy patches around.
Meanwhile OpenBSD will get 9front's GeFS stabilized and it will work on a fraction of what BTRFS needs to function, and don't let me start on ZFS.
In this case, reading the link you're replying under, it's being able to mount a disk formatted with a popular filesystem. The alternative here is waiting for a human to type this relatively "solved" problem, as a procedural requirement, more than anything.
So, it can easily mean "things won't work in this OS". But, that's to be expected with FreeBSD, for legitimately good (from the perspective of FreeBSD) reasons.
Bugfixes, feature implementation, and hardware support will not be able to be added to OpenBSD with the speed that they are added to Linux.
Whether you like this new reality or not, it is an objective fact that LLMs are the silver bullet Fred Brooks told us didn't exist—multiplying development speed by 10x or more—and that if you're not using them you are hobbling yourself. This especially goes for a "security-focused" OS like OpenBSD, because AI is more diligent about finding vulnerabilities and can do so without rest. Meaning that if you're serious about security, you are using AI to harden your software... and if not, someone else is using it to hack you.
I'm sitting here watching astra do in a day what would have taken me at least 3-6 months. The code is verified correct. The model finds, reads, understands, and then integrates the latest literature into its designs and plans. It implements the code, proves it correct, and has come up with its own novel solutions on top of the state of the art. It then refines its implementation to hit the performance goals I give it. This is the reality today.
That it would've taken you 3-6 months because you're bad at what you do and don't know much. Being 10 times faster and better than 0.05 isn't particularly impressive.
I work in formal verification at a company whose products you've definitely interacted with today. On an average day, I throw thousands of dollars worth of tokens at the SOTA models and have them produce tens of thousands of lines of both proofs and code (alongside proofs of correctness).
We're solving problems now that we never even would have attempted previously. We couldn't have even staffed them if we wanted to. Even if we had the people, it would be years of development before we saw usable results.
If you can't understand how this kind of work would take a human with a few decades of experience 3-6 months, you're the 0.05 developer in my world.
As one example I can think of, there are a ton of devices that only work on Windows (or Windows and MacOS). The audio interface I have is one example: its "official drivers" are Windows and MacOS only. I know that someone is trying to make Linux drivers for it, using LLMs, because if there's one thing LLMs are very good at, it's RE and figuring out how hardware works. One of my other friends made an accessible equivalent to the Razer app for his Razer headphones because the official app is complete shit.
If you follow the provided link, you'll see that the objection is about copyright, not quality: https://marc.info/?l=openbsd-tech&m=177411620801633&w=2 As someone who has certainly had his work sucked up into LLMs without consent or compensation, this seems like a reasonable stance to me.
Sales Engineers (another type of SE) are all the way in on LLMs to help them do unnatural things with the products they are trying to sell. I apologize in advance for this.
The article does state clearly first that it was first drafted with AI, I appreciate the author's honesty. I wish all AI-assisted articles would do the same.
Apple offer LLMs, so they are already on board with the LLM arguments.
I have many frustrations with LLMs, mainly around the type of user whos brain shuts down, but Apple would be walking a very fine line if they offer a service that they also will sue you for using.
Perhaps they should have used the Foundation Models to build their code. Really put Apple in a difficult position.
I generally assume Wine cannot use LLMs considering developers that have been in contact with Windows code cannot contribute and LLMs almost certainly were.
No such reference to LLMs possibly being trained on Windows code however. Well maybe AI companies were careful to remove problematic data from the training set.
That would be a stupid application of the rule, it would mean no developer could use an LLM for anything that already has a proprietary alternative with heavily documented internals or leaked source code.
Not so sure, if the LLM has been overfitted and can just throw back something equivalent to the original code, then it's no different from a human having seen the code and writing something similar afterwards.
On another hand, I'd be curious about Microsoft's reaction if somebody let an agent write a "perfect" Windows emulator/compatibility layer eating away at Windows's market share (assuming that's possible; since it does not exist it's probably not currently). Would not do much for kernel anti-cheats but would help for productivity software. Microsoft could either choose to say this is not a white box implementation and take down the project but it would hurt anything LLM-related since all LLM-based projects could be targeted, or let it live and eat away at Windows (I suspect they'd choose the latter because it's the shiny new thing and would probably not hurt Windows that much).
I wonder what Windows would be like if it was split off Microsoft. Technically it has far sufficient revenue to keep being maintained and updated well, but I guess the quest for ever more revenue would still lead to it becoming more and more enshittified.
I wonder what it would become if it was split off Microsoft and given to two competing companies.
Right the sad thing is that Microsoft is a great place for an OS to live because they have the money to make the right calls... they just don't seem to have the will to do so or the ability to identify someone to own it. The company has devolved to a core cesspit of middle-management.
I don't want to read LLM written text (EDIT: just replying to parent comment, not implying at all that the tutorial might be written that way; it definitely does not look like it). It would be nice if people could just disclose when they used LLMs for their writing and how, then people who care can skip and people who don't can read. Of course this would have to come with not shaming the people who disclose actual use of LLM, otherwise we're back to square one.
Ah, sorry, I was only replying to the above comment but did not mean to imply that the tutorial might be AI-written; it very much looks human-written and generally a nice read :-)
reply