Carmack on Bellard: the closest software has to a one-person benchmark

What happened

John Carmack tweeted about Fabrice Bellard. The post hit 871 on Hacker News, which sounds unremarkable until you read the comments: it's almost entirely developers taking turns listing the things Bellard has built — single-handedly, mostly in C, mostly while holding a day job, and mostly without ever giving a conference keynote about any of it.

Carmack — who is himself the closest thing the industry has to a folk hero for solo output — doesn't hand out praise like this often. When the person most often cited as a one-man software department publicly identifies someone *else* as the outlier, that is a data point worth pausing on.

The immediate cause for the post appears to be casual: Carmack noting that Bellard's body of work continues to be, in his words, hard to comprehend. But the thread underneath functioned as a community-sourced CV, and the CV is the real story.

Why it matters

If you've watched a video on the internet in the last fifteen years, you've used FFmpeg. Bellard wrote it. If you've used Docker, KVM, Android emulators, or virtually any cloud VM that isn't running on bare hypervisor silicon, you've touched QEMU. Bellard wrote that too. TinyCC — a C compiler small enough to run in a boot sector — Bellard. QuickJS, the embeddable JavaScript engine that's become the default choice for anyone who needs to run JS without dragging in V8: Bellard, 2019. BPG, an image format that beat JPEG and gave HEIC a run for its money before HEIC's patent lawyers showed up: Bellard. LTE base station entirely in software — the kind of thing telecom vendors charge seven figures for in hardware: Bellard wrote one, alone, called it Amarisoft, and it became a real company. He also held the world record for computed digits of π using a single desktop PC.

That list is not a survey of one career. It is a survey of multiple industries. Video codecs, virtualization, compilers, JavaScript runtimes, image formats, cellular infrastructure, numerical computing. Pick any one of those and Bellard's contribution would qualify as a lifetime achievement. He has done all of them, and the work is in production at companies whose engineering orgs have more headcount than the Wikipedia summary of his projects has paragraphs.

The HN thread is interesting less for what it says about Bellard than for what the comments reveal about the audience. Senior developers were not flexing. They were comparing notes — "I just realized FFmpeg and QEMU are the same person" appeared in three separate top-level comments. The pattern of "I've been using his tools for fifteen years without knowing" is not a quirk; it is the defining feature of infrastructure that actually works.

This lands in the middle of a very loud industry argument about productivity. The AI-tooling pitch is that a developer with the right copilot can ship 5x, 10x, 100x more. Some of that is true; some of it is marketing. Bellard is the inconvenient control variable. He has been shipping at his pace, in C, with the same fingers, for twenty-five years, and his output remains the upper bound. No copilot in production today has shipped anything close to FFmpeg.

What this means for your stack

First, an audit. There's a non-trivial chance that the dependency tree of your current project contains FFmpeg, QEMU, or QuickJS without you knowing it. If you ship video, transcoding, or media pipelines, FFmpeg is almost certainly in the path — sometimes wrapped, sometimes shelled out to, sometimes invisible behind a SaaS layer billing you per minute. If you build dev tooling, QEMU is the thing underneath your local emulator, your CI Linux runners, and large swaths of Apple Silicon cross-compilation. If you've recently considered embedding a scripting language and rejected V8 as too heavy, you almost certainly looked at QuickJS. The base layer of "things one person built that everyone uses" is larger than the base layer of "things consortia built that everyone uses," and that ratio is worth thinking about when you choose your next dependency.

Second, a hiring filter. The Bellard pattern — deep ownership of an entire problem domain, willingness to write the unsexy parts (codec tables, instruction decoders, register allocators), no public performance — is rare and identifiable. If you're hiring senior ICs, the person whose GitHub looks like a graveyard of one-person libraries that actually work is statistically more interesting than the person with a thousand stars on a tutorial repo. The signal is finished things, not popular things.

Third, an honest reckoning with what "productivity" means. Bellard's secret is not faster typing or AI autocomplete. It is taste, scope discipline, and the willingness to read specifications — telecom specs, video codec specs, ECMAScript specs — end to end and implement them. Those are skills the current toolchain neither teaches nor measures.

Looking ahead

The interesting question is not whether Bellard's pace is reproducible — it isn't, for almost anyone — but whether the conditions that produced him still exist. He did most of his foundational work before "developer" became a thought-leader role and before every interesting project came with a Discord and a roadmap. If your stack today depends on infrastructure written by one person who refused all of that, it's worth asking what you would build if you had three uninterrupted years and no one watching. Carmack's tweet was, in the end, a reminder that the answer for at least one human being is: most of the modern internet.