BSD's first 30 tapes show why open AI weights keep winning

The thirteenth installment of F/OSS Comics, posted October 30, covers the moment in early 1977 when Bill Joy bundled a Pascal compiler with ex, a line editor built from em, and mailed roughly thirty tapes to other universities. That bundle became the Berkeley Software Distribution. It was not an operating system. It was a pile of software you installed on top of Unix you already licensed from AT&T.

Read that again, because it is the part the AI industry keeps getting wrong. BSD’s first release was a compiler and an editor, shipped by one graduate student who handled every order himself. The second release, 2BSD, added vi and the C shell and shipped nearly seventy-five tapes. Joy copied them, addressed them, and mailed them. No sales team. No procurement cycle. No per-seat negotiation. Just tape and postage, because Berkeley charged only for the media and distribution, and only to people who already held an AT&T Unix source license.

The license was the moat, and it leaked

The comic is careful about the legal structure, and that care is the whole point. You could not get BSD without an AT&T Unix source license. AT&T, per the dialogue in the strip, was barred by its antitrust consent decree from selling computers or Unix as a product, so it licensed source code to universities for a modest fee and let the code travel. Berkeley’s Free Speech Movement culture, per the Salon piece the comic cites, is said to have pushed Joy to charge as little as he could.

The result was a distribution channel that scaled on goodwill and near-zero marginal cost. Thirty tapes became seventy-five. Then a VAX arrived in 1978, and the story accelerates: graduate student Ozalp Babaoglu, working with Professor Domenico Ferrari, was asked to add virtual memory to Bell Labs’ 32/V. Richard Fateman’s Macsyma work needed an address space bigger than physical memory, and 32/V did not use the VAX’s virtual-memory hardware. Joy helped Babaoglu integrate and debug the code through the holidays at the end of 1978, and by January 1979 most bugs were fixed. Joy ported ex, vi, and the C shell to the VAX while Peter Kessler and Marshall Kirk McKusick ported the Pascal system. 3BSD shipped in December 1979.

DARPA noticed. Its ARPANET sites had been running mismatched machines and operating systems since 1969, and porting software between them was expensive. Unix’s portability, proven by Berkeley’s distribution to research institutions, made 3BSD the candidate. That is how a student’s tape-mailing hobby became the substrate for defense-funded networking research.

What this has to do with model weights

The AI parallel is not decorative. Every argument now raging about open-weight models, from Llama derivatives to the licensing fights around Mistral and DeepSeek releases, is a rerun of the BSD tape economy. The winning artifact is rarely the most capable one at release. It is the one that is cheap to copy, legal to modify, and small enough to mail to a colleague.

Consider the mechanics the comic documents. BSD’s early releases were not standalone. They required a paid AT&T license, which meant the community was gated by a commercial relationship the community did not control. That gate is exactly what today’s open-weight licenses attempt to recreate: you can have the artifact, but the terms come from the vendor. When Meta publishes weights under a community license with usage thresholds, it is doing the AT&T move, granting source access while keeping the commercial ceiling. When a lab publishes under Apache 2.0, it is doing the Joy move, charging for distribution and nothing else.

The historical record here favors the Joy move, but slowly and indirectly. BSD did not beat AT&T’s Unix in the 1970s. It took the VAX port, the DARPA contract that followed, and a decade of accretion before the Berkeley lineage mattered commercially. AI labs announcing open weights in 2026 should expect the same lag. The capability gap between open and closed models narrows over quarters, not weeks, and the ecosystem effects compound only after the tooling around the weights matures.

The compute bottleneck is the new VAX

There is a second lesson hiding in the virtual-memory episode. Babaoglu’s task was not to write a kernel from scratch. It was to make an existing kernel use hardware it was ignoring. The VAX had virtual-memory capability. 32/V did not use it. The fix was integration and debugging, done by two people over a holiday break.

That is the shape of most useful AI infrastructure work right now. The frontier labs train the big models. Everyone else makes the existing stack use the hardware it is ignoring: better kernels for inference, quantization that fits a model into memory it was not designed for, scheduling that keeps accelerators busy. The comic’s heroes are not the people who invented Unix. They are the people who ported it, tuned it, and mailed it out.

The winning artifact is rarely the most capable one at release. It is the one that is cheap to copy, legal to modify, and small enough to mail to a colleague.

The AI economy has a version of the AT&T constraint, and it is compute. Access to training hardware is gated by contracts and allocations the open community does not control, the way BSD was gated by the AT&T source license. Open weights without open compute are a partial release. The labs that publish weights but rent the only viable clusters are running the AT&T playbook with better marketing.

What to watch

Watch the licensing terms on the next round of open-weight releases, not the benchmark scores. A permissive license with a large model is a tape worth mailing. A restricted license with a slightly better model is a source license with a ceiling. The BSD story says the permissive one wins eventually, and that the eventual winner is usually not the first mover but the one whose distribution survives the gatekeepers.

The comic ends on ARPANET noticing 3BSD. That is the moment to track in AI: not the release, but the first time a government or large institution standardizes on an open artifact because porting to it is cheaper than building around it. When a national lab or a defense program picks open weights as its default substrate, the tape economy has arrived again. Until then, the mailing continues, and the postage is still the interesting number.