Fabrice Bellard

Fabrice Bellard (born 1972) is arguably the most consequential programmer most people have never heard of. He created FFmpeg, the media engine behind essentially every video played on the internet; QEMU, the emulator underpinning modern cloud virtualization; the Tiny C Compiler; QuickJS; and, for good measure, once held the world record for computing digits of pi on a desktop PC. He almost never gives interviews — his work speaks entirely through shipped code.

The quiet pro-AI vote

So when Bellard turned to AI, no manifesto appeared — just software. Starting in 2019 with GPT-2, he built TextSynth, a commercial LLM and image-generation inference server written in plain C on his own LibNC library: one binary, no Python, custom 8-, 4-, and 3-bit quantization, sold under commercial license and running the textsynth.com playground. His NNCP neural compressor took first place on the Large Text Compression Benchmark, and 2023’s ts_zip used an LLM to squeeze text beyond what conventional compressors can reach.

That body of work makes Bellard a distinctive data point in the debate: a systems purist of the old school — the kind of figure John Carmack has called almost certainly a better overall programmer than himself — who examined large language models and judged them worth a decade of his scarce, legendary attention. Where others argue about whether the technology is real, Bellard reimplemented it from scratch, alone, in C.