Brennan Kenneth Brown’s love letter to SDF is nominally about a bulletin board system. Read it through an AI lens and it is an argument about durability. SDF, the Super Dimension Fortress, has run continuously since June 16, 1987, when Ted Uhlemann (“iczer”) plugged an Apple IIe’s 300-baud modem into a phone line and named the board after the anime Super Dimension Fortress Macross. Today it is a multi-user NetBSD machine funded solely by members as a 501(c)(7) nonprofit, with tens of thousands of accounts and a $36 lifetime ARPA membership. Brown joined this past week. The piece is worth reading for the guestbook Perl, but it is worth thinking about for what it says about the stack we are now building around large language models.
The contrast is not subtle. SDF survived hardware failures, “countless DDoSes,” a termination by its own ISP, and multiple relocations, per Brown’s account. It never went commercial. Stephen Jones, the longtime sysadmin, wrote in a 30th-anniversary letter that newcomers may see the place as “outdated and ‘retro,’” but that “if you get involved, you’ll see it is quite alive with new ideas.” That is a 39-year-old system making a claim most AI infrastructure cannot make about a model checkpoint from eighteen months ago.
What SDF actually runs
The specifics matter because they show what durability looks like in practice. SDF serves you.sdf.org pages from a ~/html directory, Gopher and Gemini from a single ~/gopher directory served over two protocols at once, and runs a roster of services on subdomains: irc.sdf.org, jabber.sdf.org, matrix.sdf.org, git.sdf.org, mastodon.sdf.org, pixelfed.sdf.org, lemmy.sdf.org, and toobnix.org for PeerTube. ARPA members can pick a vanity domain from about 60 choices, including pdp10.org, twenex.org, chaosnet.org, multics.org, and freak.net. The hardware includes NetBSD servers on DEC Alpha and AMD Opteron, plus live retrocomputing environments: a TOPS-20 system on two XKL TOAD-2s, a Symbolics Genera machine, and an ITS system. Free accounts on all of them.
None of this is efficient. Brown’s own CGI guestbook spawns a new OS process per visitor click, which is precisely why traditional CGI was abandoned. He knows it. “Things like ‘scaling’ and ‘efficient resource usage’ can be thrown out the window,” he writes. “Who cares?” The guestbook stores entries as single lines separated by the \x01 control character, uses a CSS honeypot to catch bots, and relies on a .htaccess deny rule to keep the data file out of the web root. It is janky by 2026 standards, and he loves it.
The AI parallel nobody wants to draw
Here is the take. The AI industry is currently building the most disposable software stack in the history of computing, and it is doing so while claiming to be the most important thing happening. Model weights are re-released on a cadence measured in weeks. API endpoints get deprecated. Agent frameworks are rewritten before their tutorials finish rendering. A developer who built a workflow on a 2023 model version has, by now, likely rebuilt it twice. None of that is a durable community. It is a treadmill with a subscription.
SDF is the counterexample. It predates the modern internet, and it is still answering SSH connections at tty.sdf.org. Its members get a shell, bash through zsh, editors from ed to emacs, ping, traceroute, dig, whois, and finger. They get voting rights on system features and policy. They get quotas of roughly 150 MB and 5,000 files each for home, mail, webspace, and gopherspace. That is a governance model, not a product roadmap. The 2019 survey History, Evolution and Future of Public Access UNIX, hosted on SDF itself, describes the system as “a fully articulated social community,” with special interest groups for ham radio (SDFARC), Python, Gopher enthusiasts, and Minecraft players alongside “digital artists, computer historians, web developers, gamers, and of course many, many programmers.”
Model weights are re-released on a cadence measured in weeks. SDF is still answering SSH connections at tty.sdf.org.
The AI economy has no equivalent. There is no $36 lifetime membership to a frontier lab. There is no member-funded nonprofit where you get voting rights on which capabilities ship. There is no place where a researcher can log into the same environment their predecessor used in 1987 and find it still running. Instead there are usage-based bills, rate limits, and the quiet understanding that the tool you depend on may be retired or repriced by the next earnings call.
What AI builders should take from a 1987 BBS
The lesson is not “use Gopher.” It is that the most valuable technical systems are the ones that outlive their creators’ business plans. SDF’s ARPA membership buys outbound telnet, ssh, sftp, ftp, snarf, wget, ytalk, and external IRC, plus gcc, perl, php, python, ruby, and elisp on the box. It buys UUCP and ClariNET feeds over TCP or dialup. It buys a real DID option for VoIP at $15 per quarter. These are commitments. They are the opposite of a deprecation notice.
AI builders who want their work to last should be asking the SDF question: what happens to this when the funding runs out, when the model is superseded, when the vendor pivots? The answer for most current agent infrastructure is “it stops.” The answer for SDF, 39 years in, is that a member can still ssh in and run mkhomepg to publish a page. Brown’s guestbook is live at bren.sdf.org/guestbook.cgi, note the http, not https. It will probably still be there in a decade. The same cannot be said, with confidence, of anything shipped this quarter.
The outstanding question is whether the AI tooling ecosystem can produce its own SDF: a durable, member-governed, non-commercial substrate that researchers and builders can rely on across model generations. Nobody has built it yet. The 150,000 BBSes that once operated in North America are down to a few hundred, per Brown. SDF is one of the survivors because it was never trying to scale. That is the part the AI industry keeps missing.