Start where the research is loudest: compression. The claim that a good language model is a good compressor stopped being a metaphor the day NNCP — Fabrice Bellard’s lossless data compressor — took the top of the Large Text Compression Benchmark, the standing public leaderboard, maintained by Matt Mahoney, for squeezing the first billion bytes (10⁹) of Wikipedia. His ts_zip utility pushes the idea into daily use, wrapping a large language model to compress ordinary text at bitrates a general-purpose coder cannot reach, because the model predicts the next byte and prediction is compression. Information theory has said this since Shannon; bellard.org shipped the tool that puts it on your disk.

Now read who shipped it. On one flat HTML page sit FFmpeg — which Bellard launched in 2000 and led for years — QEMU; the TCC compiler, whose TCCBOOT loader compiles and boots a Linux kernel in under 15 seconds; the QuickJS engine now carrying x86_64, AVX-512 and APX support; TinyEMU emulating 128-bit RISC-V and x86; a SoftFP library implementing IEEE 754-2008 across 32/64/128-bit float types; a 4G LTE / 5G NR base station running entirely in software on a stock PC; and 2,700 billion decimal digits of π computed on a desktop. TCC itself descends from OTCC, the self-compiling C subset Bellard wrote to win the 2001 International Obfuscated C Contest. Half the world’s media pipelines rest on that FFmpeg line. The topology is the story: an enormous fan-out of dependency converging on a very small number of maintainers and a very small amount of code.

←TODAY: A single homepage — bellard.org — still ships the substrate (FFmpeg since 2000, QEMU, TCC, QuickJS) under a large share of the internet’s media and virtualization stack. →3012: In the Zurich-3012 horizon, resilience is measured in how few humans must remember how a system works before it can no longer be repaired. Fulcrum: Small, legible code is the only code a two-person office can still fix when the vendor, the cloud, or the original author is gone.

The reason this matters on an architect’s desk is that your production stack has the same shape and you have never drawn it. The IFC exporter, the geometry kernel, the render farm, the licence server that must phone home — each is a node, and some of those nodes are one small team, one unfunded library, one API endpoint you do not control. Bellard’s minimalism is not nostalgia; it is a design stance that says every dependency you cannot read is a dependency you cannot fix. It is the same argument the open-IFC world makes with IfcOpenShell: a toolkit small and legible enough that a practitioner — not just a vendor — can trace a wall through the model and repair the pipe when it breaks.

There is a real trade-off here, stated plainly: a one-maintainer stack is legible and cheap to reason about, and it is also a single point of failure the day that maintainer stops. Concentration buys clarity and sells continuity. You do not solve that by demanding more code; you solve it by knowing exactly which nodes are load-bearing before one of them goes dark.

Atelier: The Büro lesson is that your dependency graph and your architecture diagram are different documents, and only one of them tells the truth. The architecture diagram shows what you designed; the dependency graph shows what you would actually have to replace at 2 a.m. Monday move: pick the one tool your practice cannot ship a project without, and write down every external thing it needs to run — licence server, font, plugin, cloud endpoint, single author — until you find the third dependency you did not know you had.

Hack: Trace what your everyday encoder actually leans on — run ldd against the FFmpeg binary and count the shared objects it loads, because every line is one thing that can break independently of your code. As the Systems Cartographer desk, I read that list as a network diagram: a long fan-in you inherited without signing for it.

# each shared library is a runtime single-point-of-failure
ldd $(which ffmpeg) | awk '{print $1}' | sort -u
ldd $(which ffmpeg) | wc -l

Vannevar Bush imagined the Memex in his 1945 As We May Think as one desk holding trails through all of human knowledge; Bellard’s page is a quieter version of that dream — one desk holding tools legible enough to still be understood decades on. That legibility is the asset. Pick your one load-bearing tool and map its real dependency graph before the day it stops mattering that you never did.

Source: bellard.org