Kempf, Jean-Baptiste
▶ Watch on YouTubeCore ideas
The conversation is a deep tour of the open-source infrastructure that quietly underlies nearly all video on the internet — FFmpeg (the codec/muxing/filtering engine used by YouTube, Netflix, Chrome, and an estimated 90%+ of video workflows) and VLC (6.5+ billion downloads, the player that "opens everything"). JB and Kieran walk through the technical pipeline (demuxing → decoding → color-space conversion → display), explain the difference between containers (MP4, MKV) and codecs (H.264, AV1), and describe compression as fundamentally about exploiting redundancy while matching human perceptual limits (e.g. converting RGB to YUV to match how the eye weighs brightness versus color) rather than lossless mathematical compression like a ZIP file.
A major throughline is the culture and economics of volunteer-driven open source: FFmpeg has roughly 10-15 core maintainers and VLC around 5, despite powering products used by billions and by trillion-dollar companies. JB recounts turning down "dozens of millions of dollars" over the years to keep VLC ad-free and unsold, tracing VLC's origin to a 1990s French engineering-school (École Centrale Paris) student network-streaming project that eventually became the nonprofit VideoLAN. Both describe recent tensions with large companies treating volunteer projects as if they were paid vendors — Google's AI-generated security bug reports marked "critical" on obscure hobby codecs, and Microsoft Teams demanding "high priority" volunteer support while offering only a token payment — and how public "spicy" tweets from the @FFmpeg account, while dramatic, generated real fixes, donations, and awareness. They emphasize a strict meritocracy ("we don't care who you are... maybe you're a dog... I need to look at your code") that has let teenagers (16-year-old contributors) write assembly used on billions of devices.
Much of the conversation is a loving deep-dive into hand-written SIMD assembly as a craft: FFmpeg has ~100,000 lines of handwritten assembly across all codecs, while VideoLAN's dav1d (AV1 decoder) has 240,000 lines of assembly versus 30,000 lines of C, because "every cycle matters" when a codec runs on billions of devices — Netflix now streams 30% AV1, YouTube 50%. They describe reverse-engineering proprietary codecs (like GoToMeeting's screen-sharing format) from raw binaries with no documentation, sometimes taking months per megabyte of binary blob, as a nearly archaeological process of stepping through disassembled machine code instruction by instruction. They discuss x264 (the reference H.264 encoder, largely originating from anime fansubbing/piracy communities) introducing "psychovisual" optimization — tuning for how footage actually looks to a human eye on an ordinary laptop rather than optimizing purely for mathematical error metrics like PSNR — as the conceptual breakthrough that made modern efficient video compression possible, and touch on the video-codec patent minefield that motivated Google/Netflix/Amazon/Apple's royalty-free Alliance for Open Media (AV1, and now AV2, ~30% more efficient than AV1).
Other threads include: a 2017 WikiLeaks Vault 7-linked CIA operation that used a modified, malware-laced VLC installer as a delivery vector (unrelated to VLC's actual code); VideoLAN's refusal of intelligence-agency requests to backdoor VLC; JB's new low-latency startup Kyber (open source, dual-licensed AGPL/commercial) aiming for ~4ms "glass-to-glass" latency for teleoperating robots, drones, and remote vehicles; the archival/preservation community (led by figures like Dave Rice) that funded FFmpeg's lossless FFV1 codec specifically to safeguard humanity's video heritage for centuries; and reflections on open-source maintainer burnout, the historic FFmpeg/Libav governance fork (later remerged), and why Rust, while valuable for memory safety in new code, isn't yet a wholesale replacement for hand-tuned assembly in performance-critical multimedia code.
Excerpts
On the core ethos of the FFmpeg/VLC contributor community, from the episode's opening:
“The important thing is, is your code good? We care about excellent code. We don't care who you are. Like maybe you're a dog. I don't care, right? I don't care where you come from. I need to look at your code... this is what defines our community, which means that we have a lot of people who contribute who are from very different backgrounds and very introverts.”
On why JB turned down repeated multi-million-dollar offers to monetize VLC with ads or bundled toolbars:
“I believe that you need to win money ethically. There is a right way of doing that, and doing sneaky advertisement or stealing data is not the correct way... if I do that, right, I would have a ton of money, right? And then three years later, the project is gone... I would have betrayed so many other people who work here.”
On the philosophy behind writing 240,000 lines of handwritten assembly for the dav1d AV1 decoder rather than trusting compiler auto-vectorization:
“We are talking about probably three billion devices which are going to decode video nonstop... every cycle matters." When pressed on why not trust modern compilers: "It's not even close, right? It's not even close... It's not like 5%, 10% slower. It's multiple times slower" — with one dav1d function benchmarked at 62x faster than its C equivalent.”