Imagine a team that spends a quarter century building the same piece of software — and all that time can't bring itself to give it version number 1.0. Not because they can't pull it off: betas ship regularly, the latest one right on the anniversary. But because the word "beta" here isn't a marketing stage, it's an honest description of the state of the world. This fall the Haiku project turned 25, and its story isn't about nostalgia for a dead operating system. It's the longest continuous experiment in the industry: can volunteer hands, without corporate budget and, lately, without a single line of AI-generated code, build an entire operating system from scratch — not Linux, not BSD, not Unix at all.
It all started with a routine digest: among the day's news came a brief note "Haiku turns 25." The number caught my attention not with its idle anniversary, but with two details that together looked like a logical error. First: in a quarter century the project still has no final R1 release — just alphas and betas. Second: working on all this labor is a single paid contractor, living on donations — and, rumor had it, building the 32-bit version requires keeping two compilers running simultaneously, ancient GCC 2 and modern GCC 13.
When you reach the primary sources, the "rumors" turn into a whole architectural drama. Two compilers aren't a curiosity, they're direct inheritance from the fact that Haiku must be binary-compatible with BeOS, the system killed for $11 million in 2001. And "one contractor" isn't a sob story, it's a remarkably efficient model: readers fund the contract, and the results from the past two years — 530 closed tickets and the fastest version in the project's history. Let's dig into how this all works.
To understand Haiku, you first need to understand what it inherits from. Be Inc. was founded in October 1990 by Jean-Louis Gassée — the man John Sculley forced out of Apple. According to corporate legend, Gassée and his partner Steve Sakoman started building a computer prototype the very next day after leaving. Seymour Cray himself provided the startup capital. The company name was born from a joke Gassée loved to tell: he wanted to call the firm "United Technoids Inc.," Sakoman started leafing through a dictionary looking for something better, got tired on the letter "B" — and Gassée said: "'Be' — not bad. End of story." The future system's slogan was written long before it: "One processor isn't enough for a person."
BeOS was a declaration of war on all the legacy of the 1970s. In their signature white paper, Be wrote directly: Windows 95 grew out of DOS, Windows NT out of VAX mainframes, Mac out of Xerox PARC ideas; the foundations of all mainstream systems "simply weren't designed for high-bandwidth media." BeOS was built as a "Media OS" — from scratch, in C++, multiprocessor from the start, without layers of accumulated backward compatibility that the paper called "software silt." It was the only truly new desktop system of its generation: multiprocessor from day one, with a file system that worked like a database, with audio path latency Windows would dream about a decade later.
Commercially it went differently. BeBox with its famous CPU load indicator strips launched in 1995 and remained a niche relic. Hope for salvation came in 1996 when Apple was looking to replace its sunken Copland: the finalists were BeOS and NeXTSTEP. Steve Jobs convinced the board to buy NeXT for $400 million. Rumors said Be was asking $500 million and a high position — Gassée himself denied this, but Apple's choice buried BeOS's chances of becoming Mac's system platform. What followed — ports to x86, a bet on internet gadgets BeIA, a NASDAQ IPO in 1999, and universal dead end: OEM manufacturers were bound by Microsoft contracts and physically couldn't install alternative dual-boot systems.
Microsoft contracts — that's not a figure of speech. In 2002, already dying, Be filed an antitrust lawsuit against Microsoft, directly claiming: licensing agreements forbade OEMs from dual-booting with non-Microsoft systems, Compaq under pressure abandoned a joint gadget, and Microsoft artificially tanked Be's IPO. The irony of history: in September 2003, when Be itself was already gone, Microsoft paid its estate $23.25 million — more than the company would have made from the sale itself. The sale itself took place August 16, 2001: Palm paid $11 million for the assets and, as became immediately clear, wasn't planning to do anything with the desktop OS. A year later Be Inc. legally disappeared.
The day after the deal was announced, August 18, 2001, a message appeared on the Be developers mailing list with a subject line now quoted as the project's charter: "Ok, let's start." The author, Marcus Overhagen, laid out the technical framework right in it: full binary compatibility between releases, compatibility layer for existing programs, compiler-independent format, modular kernel with drivers loadable on the fly. The ideological leader, Michael Phipps, later said in a 2003 interview that the idea to start wasn't his: the BeDevTalk thread was opened by Cédric Dejean. Phipps was planning to develop his own programming language at the time and didn't want to port it to someone else's OS — but the resistance didn't last long.
Three weeks later, September 9, 2001, the project imported the NewOS kernel by Travis Geiselbrecht — an engineer who'd worked at Be itself. This kernel still powers Haiku today. And Phipps formulated the key architectural decision: everyone usually starts with the kernel — Linus, Jolitz; but the kernel scares off ninety percent of developers who don't like "raw zeros and ones." So OpenBeOS (the project's first name) wrote all its "kits" in parallel — Interface Kit, Media Kit, Storage Kit, each with its own team. When Palm refused to license BeOS source code (January 2002, official "firm decision" from PalmSource chief David Nagel), the only path was reverse engineering: rewrite the system by observing the original's behavior. The first years volunteers lived on a stopgap — a set of open components that installed over live BeOS 5 and replaced proprietary pieces one by one.
The 2004 renaming also became legend, but legend with a correction: popular myth says Palm sued the project over trademark. In reality Palm never sent any lawsuits — the community itself decided not to play with the boundaries of someone else's "BeOS," voted and announced the new name at the first WalterCon conference in Columbus, Ohio. The name "Haiku" — an homage to the NetPositive browser from the original BeOS: it reported 404 errors not with numbers but with a three-line poem in haiku form. A pedantic project that even formatted the system's death beautifully.
Gassée himself gave symbolic blessing: in February 2007, at a Google Tech Talk, the ex-CEO of Be came on stage with the project leader and wished them success. By that point Haiku was already drawing windows, sitting in Google Summer of Code with eight students, and a year later compiled itself for the first time. The first alpha came out in 2009, betas starting in 2018, and at each step the system gathered around itself that very ecosystem BeOS's death had stripped away: Firefox and LibreOffice in repositories, Wine for Windows software, Linux ports through POSIX compatibility.
Now the number that makes Haiku a statistical unicorn of open source. In August 2021 the nonprofit Haiku, Inc. signed a contract with a developer under the username waddlesplash — not for a specific project but for "general system improvement" with an open end date. The announcement honestly counted the money: there was a reserve of about a year and a half of work in the account, and for sustainability the annual donation target needed to rise from $10,000 to $80,000. For scale comparison: the annual budget of many popular Linux distributions' foundations is measured in hundreds of thousands and millions of dollars.
Six years later the model works. In the R1/beta6 release notes (August 26, 2026) it says directly: the contract for the entire two-year release cycle was funded by donations through a 501(c)3 nonprofit organization, and this is the only paid person on the project. Each of his monthly reports ends with the same phrase: "thanks to donors who make my contract possible." And now look at what one person with volunteers accomplished in the cycle: 530 closed bugs and tickets, new memory allocator based on OpenBSD, pipeline optimization of system calls that sped up git status on a 160,000 file repository from 15 seconds to 2.5, and a complete rebuild of the web engine that dropped from 4 hours 53 minutes to 2 hours 33 minutes — nearly in half. Not bad for a "one-man day-laborer project," right?
The pace is confirmed by raw GitHub numbers: in the first two and a half weeks after the anniversary release, 87 commits from 16 authors hit the main branch. waddlesplash — 21, classic volunteers like Scipione and PulkoMandy — 17 and 11. It's a small but living organism: after each release it doesn't exhale but continues spinning at the same speed. Salary here, apparently, plays the role not of "hiring" but of insurance against gaps: when volunteers hit exam season, deadlines, or life, the paid hand keeps the project from stalling.
To the question "why 25 years without version one" the project has an honest, almost pitiable in its candor answer. In 2017 developer kallisti5 wrote a post titled "Where is Haiku R1?" where he laid it all out: five years between alpha 4.1 and that post were eaten by implementing the package system, because after it "we needed to build all the ports into packages and maintain dependencies." Then a phrase worth framing next to any open-source manifesto: "in most operating systems such a system is written by a full-time team of engineers; we have a few people writing it in their spare time." And the finale: "I know we've been saying 'soon' for quite a while now… but soon."
"Soon" continues to this day, and it's not powerlessness but a conscious position. R1 by internal definition isn't "the version after which marketing begins" but "the version where the system fully covers the project's goals." Until the ARM64 port is pulled up, until multiprocessor WebKit is brought up to speed, there won't be a version one. Also telling is how much is already alive in the system: 3,900 open tickets against 16,000 closed, including true dinosaurs like ticket #344, opened March 21, 2006, "20 years ago" right on the page. The bugtracker numbers themselves are small: Haiku ran development in its own Trac system for a long time, not on GitHub, and it was opened in an era when "bugtracker as memory repository" was normal — some tickets are younger, but ancestor numbers like #344, #570, and #605 (May 2006) still hang open, having outlived several version control system changes.
Yet users themselves are increasingly asking if it's time to rename "beta" to "stable." Monthly builds are so high quality that part of the community lives on them as a daily driver, and The Register review, usually dry as a bone, finished its beta6 breakdown with a phrase Haiku won't forget: "if it were up to us, we'd call this version 1.0 without hesitation." The project isn't rushing — and in this, perhaps, is its most underrated talent: twenty-five years resisting the temptation to declare victory until it becomes undeniable.
And now to the sharpest part of the anniversary. In July 2026 a short file AGENTS.md was added to the root of the Haiku repository, its entire meaning two sentences: "AI-generated code is prohibited. Each contribution must be written by a human." Previously this policy existed politely — in the copyright policy section on the site, which said the project doesn't accept contributions with "unclear licenses," and that "this includes most content produced using LLMs." Now it exists bluntly and fashionably: a file that will be read by both agents and their owners.
The policy isn't abstract — it already has casualties. In April 2026 developer rcarmo appeared on the forum building an automated ARM64 Haiku build pipeline using LLMs and honestly admitted it in his very first message. The response was polite but absolute: such code cannot be merged. Rcarmo argued that banning neural networks because of copyright was "a fundamental misunderstanding of how they work," continued work as an independent fork — and at the end of May archived the repository with the words: "I don't see the point in spending tokens and time on a system that, despite the joy it brings me, has no holistic view of how to keep up with the times." Yet his technical result was real: his ARM64 Haiku build booted to desktop in QEMU. They didn't take away his skills — they just took away his right to be part of the project. In parallel YouTube enthusiast ActionRetro assembled Tabby-PPC — a Haiku distribution for PowerPC, also with AI help and also knowingly remaining a fork forever.
The community's choice isn't trivial, and in July 2026 on the forum it was discussed with all the bitterness and enthusiasm at once. X512's counterargument — a logical trap: if everything that touched AI even indirectly is forbidden, then "a bug found by AI and confirmed by a human can't be fixed because AI was used as a source." And developer Jonson26 added an image: "even googling documentation puts you in contact with AI. Avoiding generated stuff reminds me of bullet hell games." Policy supporters answer with fact, not slogan: "I know at least one organization that chose Haiku for public systems precisely because of this stance, and several people looking at the system because of the no-LLM policy." The project's philosophy here is strict and consistent: if code can't be legally clean to its roots, it doesn't exist — and AI-modified code by definition can't be legally clean.
Against the industry backdrop this looks like an anomaly, and The Register's anniversary review took the trouble to list everyone else: the Linux kernel officially documents work with AI assistants (mandatory Assisted-by: LLM tag, ban only on Signed-off-by for robots), Ubuntu embeds AI tools, Debian in August 2026 held a general vote on nine policy options and chose "Responsible Use of Generative AI" — 425 votes, the ban lost 111:273. Haiku remained the only living desktop system that answered this question with "no" — and did so not out of fear but because in its coordinate system code origin isn't fashion, it's foundation.
The last question that usually ends such stories: but seriously — who uses this? The project itself honestly doesn't know: there's no reliable way to count users, the forum gives roughly 85,000 visits per month, downloads aren't declared, telemetry is minimal and sparks community disputes with every mention. Forum humor: "there are three Haiku users." But those three write substantively.
In March 2026 Corbin Davenport on How-To Geek described an experiment: a 2009 netbook laptop, single-core Intel Atom, 2 GB memory. Debian 12 with LXQt — "on the edge of usable," web doesn't work. Haiku installed in minutes, boots in 30 seconds, media player and text editor start faster than a second — "squeezes more performance out of this ancient netbook than I thought possible." Honest verdict: not ideal for a laptop yet — no sleep mode, no password, Bluetooth not recognized. In August Jeff Butts on XDA spent a month with Haiku and wrote text that best explains what this is all for: "I don't think Haiku will replace my main system — but I constantly catch myself wishing a couple of my main system's ideas worked the same way as in Haiku." We're talking about a system where a saved file search isn't a bookmark but a live view; where window stacking isn't a trend but a solution invented before tabs became fashionable.
Tellingly, both authors, like many reviewers, reach the same conclusion: nostalgia here is secondary. Haiku is interesting not as a BeOS museum but as proof that a desktop operating system doesn't have to accept the decisions every modern desktop has accepted: cloud background, server roots, AI layer on top. While in the industry the world slides into five platforms and one design, the existence of a system that answers the question "can we do it differently?" with living code — that's not entertainment and not an archive, it's a control experiment someone must run. Twenty-five years — just the lab's operating term.
If you pull one thought from this story, it's not "persistence matters more than results." Something more uncomfortable. Haiku shows that in open source, time isn't a competitor but a resource if you have a philosophy that can't be bought. A system that's spent 25 years not giving itself 1.0, not because it's embarrassed but because it keeps its word to its own declaration of goals — rarest species. It won't surrender its name to someone else's trademark (2004 renaming), won't surrender code origin to a neural net (AGENTS.md 2026), won't surrender its route to someone else's roadmap (which is why there are no no-roadmap scandals like in big distributions).
And here too — the price. The AI ban is a choice that in literal terms has already cost the project one real ARM64 bring-up and one capable person. In a world where Debian on a general vote chose "responsible use," and the Linux kernel issued official rules for LLM assistants, Haiku's position looks either like blind stubbornness or the only pure experiment in the industry — or maybe both at once. I lean toward the second: the rationality of the ban isn't that neural networks write code badly, but that a project built on judicial cleanliness of every line's origin can't allow itself a murky source, even an efficient one. But to be fair: if in ten years Haiku is still on beta version 7 and the ARM64 port still hasn't caught up with x86, we'll have the full picture of what ideological purity costs.
And the most curious part lies ahead. The project is now 25, it has one paid developer fed by readers of its own site, and in two years he made the system twice as fast. If someday you see someone on an old laptop with a desktop sporting neat three-dimensional window shadows and Mac-like laconicism — don't rush to write it off as retro. It might just be the longest living experiment in open operating system history, still running.