Every Cycle Matters: The Volunteers Who Decode Almost All Internet Video

Lex Fridman opens by calling FFmpeg the invisible backbone under YouTube, Netflix, Chrome, VLC, and Discord, and VLC a player that takes almost any format with no ads and no tracking. He is wearing the orange traffic-cone hat for the whole recording. The guests are Jean-Baptiste Kempf — president of VideoLAN and a lead maintainer of VLC — and Kieran Kunhya, a longtime FFmpeg contributor, codec engineer, operator of the infamous @FFmpeg account on X, and founder of Open Broadcast Systems.

Asked for the weirdest thing VLC opens, Kempf lists VHS capture cards, DVD-Audio with custom encryption, and game codecs that live in FFmpeg. Kunhya's favorite is a ten-second Star Wars game intro that someone implemented bit-exact from a single disc. At a VideoLAN conference they ran a contest for the worst file: an MKV whose every frame changed resolution, aspect ratio, and rotation, and another "video" that was nothing but black frames with SSA subtitles animated on top. Both played. There was also a file that was a valid ZIP and a valid MP3 at once.

A quarter of VideoLAN's website traffic arrives on the query "cone player." Kempf says people in remote parts of India and China recognize the cone who do not know the name VLC. An April Fools' plan to swap the logo for a caterpillar-construction mark drew about 10,000 angry emails. A play button on a television would have been the YouTube mark, he says; the cone is orange, loud, and weird enough to become a meme, and a meme durable enough to last twenty years. They did try a pancake in a DVD drive. It did not play. They have the video.

Fridman asks what happens after you hit play. First the player turns a URL, path, or disc into a byte stream. A demuxer then cuts that stream into compressed video, audio, and subtitle frames. The player peeks at each video frame to decide whether a GPU can decode it; Kunhya puts the share that cannot at about 45 percent, because codec variants and vendor capabilities differ. Hardware gets a black box. Software starts with entropy decoding — Huffman or arithmetic coding — then intra prediction in the spatial domain, a residual stored in the frequency domain, an inverse transform, and the add-back that makes a picture.

People forget how far the signal is crushed, Kempf says. PCM audio to MP3 is roughly 10x; video is 100x or 200x, and the target is 1,000x. That is not ZIP. ZIP tries to approach a lossless limit; video codecs deliberately degrade a signal so a human still likes what remains. They leave RGB for YUV — luminance plus two chroma planes — because cones and rods in the eye care more about brightness than about color detail. Halving chroma resolution, before any clever math, already halves the file, and most viewers never notice.

Modern codecs live in the frequency domain, splitting frames into blocks with a discrete cosine transform, which is why a bad decode shows squares. Each generation buys about 30 percent less bitrate for the same look, at an order of magnitude more CPU. Kunhya's warning sits under every later war story: every sentence in that pipeline is someone's lifetime of work, and there are books about each sentence. Compression is also asymmetric. Encoding can spend huge compute once; decoding has to be cheap enough for a phone, and recoverable enough that a UDP stream with lost packets can still join mid-flight.

A container, or muxer, is the file format that holds several tracks; a codec is the coder-decoder inside. MP4 is the household name. Kempf grew up on Microsoft's AVI and Apple's MOV, which became MP4; VideoLAN still has Steve Lhomme, who started Matroska as a more future-proof open container. Confusing MP4 with H.264 is not quite a crime. H.264's formal name is MPEG-4 Part 10, also called AVC, inside a meta-spec that also holds early MPEG-4 video and AAC audio. Kunhya's practical rule: 99 percent of MP4 files are H.264 plus AAC, and the rest are edge cases.

VLC and FFmpeg still do not trust the suffix. A `.mp4` only bumps the priority of the MP4 demuxer; the programs look inside the bytes. That habit is older than the consumer player. VLC began as the client of a late-1990s campus streaming system called VideoLAN, feeding MPEG-2 over UDP, where packets arrive damaged. "You don't trust your inputs," Kempf says, and the whole architecture was built for broken files. When people downloaded incomplete AVIs whose index sat at the end of the file, other players quit; VLC said the file was broken and tried anyway.

A codec's job, Kunhya adds, is to strip spatial and temporal redundancy and still leave a stream you can join and recover. Kempf's picture is a camera pan: the cloud in the background does not need to be resent every frame; a black wall can be "same as the top-left pixel." In theory the search window is unlimited. In practice memory and compute kill you — comparing 4K pixels two hundred frames back is not a hobby. AV1, AV2, and VVC are not one codec so much as a toolbox, switching tools when a Zoom call goes from a PowerPoint to a movie clip.

FFmpeg is the low-level library for codecs, muxers, demuxers, and filters, plus a command-line tool that has become a language of its own. It shows up in VLC, Chrome, smart TVs, OBS, and professional boxes; Fridman cites an estimate that more than 90 percent of online and offline video workflows touch it. VLC has been downloaded at least 6.5 billion times. Kunhya likes that a grandmother's home movie and a trillion-dollar company stand on the same stack, sometimes as a 3,000-character FFmpeg command. People now ask language models to write those commands. Imaging, Kempf notes, has ImageMagick but nothing with FFmpeg's breadth; video accidentally got a command-line After Effects.

His analogy for open source is a chocolate cheesecake. A bakery sells you the cake. A grandmother might give you the recipe. Open source hands you the cake, the recipe, instructions for building the oven, and the rules for changing and reselling the recipe. Software is just a very long list of tiny instructions; computers are not clever, they are fast. Since FFmpeg began, some two to three thousand people have contributed. The license is the Rousseau-style social contract of the group. You can always tighten a license, never loosen it without consent, which is why cloud vendors wrapping Mongo or Elastic provoked relicensing fights FFmpeg cannot copy: too many contributors, some of them no longer alive.

Kempf moved libVLC's engine from GPL to LGPL so third parties could embed the player without open-sourcing a whole game, and because Apple's App Store makes GPL nearly impossible; the iPhone and Apple TV builds use MPL. FFmpeg is LGPL in the same spirit: ship a closed app, but give back changes to the library. Relicensing meant treating the project as joint work under U.S. copyright. He contacted more than 350 people. One trail ended at a factory, where a father had to hear why a stranger wanted permission for five lines his dead son had written. Kunhya's map of the same community includes a contributor in a Syrian war zone with part-time electricity. "Maybe you're a dog," Kempf says. "I don't care. I need to look at your code."

Linus Torvalds is the folk devil of this culture. Kempf thinks Git may matter more than the kernel, and notes that Torvalds is usually harsh to subsystem maintainers who already know him. Azure still runs mostly Linux; Android is Linux; quality does not get a pass. VLC's core is five people, FFmpeg's ten to fifteen. A thousand names on the timeline and ten who remain is a one-percent stay rate. People change jobs, partners, and continents. "We are the ones who are going to maintain your code," Kempf says. It has to be excellent, not merely good. Online the tone is a subculture; at VideoLAN DevDays the same people are fun. English is nobody's first language, which makes American readers hear insult where a tired European volunteer meant brevity.

A Reddit meme claims Kempf refused tens of millions to keep VLC ad-free. He confirms it, more than once. The project did not begin as his invention. Students at École Centrale Paris, stuck on a 1980s IBM/3Com token-ring campus they fully administered, wanted a faster net for Doom. A Bouygues CIO suggested one giant satellite dish and MPEG-2 on the LAN — Network 2000, a 45-second leaky demo. Later Christophe Massiot and others turned that into VideoLAN, spent three years forcing the school to open-source it in 2001, and shipped a client that became VLC. Kempf arrived in 2003, pulled the work into a nonprofit, and by 2005 was one of two active developers who refused to let it die.

The money that did show up was dishonest: browser toolbars that were spyware, hijacked homepages, ads inside the player. "If Netflix arrived and said we want to put Netflix inside VLC, probably the story would have been different," he says. They did not. He is not against money — he has started several companies — he is against sneaking. The last offer was "obscene," dressed up as funding for some future open-source project. Selling out, he thought, would kill VLC in three years via a fork and betray everyone who had worked unpaid. In the 2000s, Kunhya reminds listeners, installing a program that quietly dropped spyware was normal, buried in license text nobody read.

Kunhya is careful: Google is one of the biggest supporters of open source. What changed is that it started using AI to file security reports against FFmpeg, announced how good the model was in the press, and applied a standard 90-day disclosure clock designed for salaried teams. One issue sat in an obscure 1990s game codec. His metaphor is a house padlock. The lock is not there to protect Fort Knox; using AI at scale to pick it, then telling the homeowner the lock is unsafe, is a move available only to the party that also has the engineers to send patches. The reports are wordy. On niche codecs they amount to "a denial of service by AI-generated bug reports," and everything is marked high. Another non-Google "vulnerability" was an integer overflow in a filter that could tint one pixel the wrong color, scored 7.5 in red.

Former FFmpeg developer Alex Strange, posting personally, described the incentive gap: a researcher gets a cute name, a logo site, a Google bounty, a DEF CON prize, and "some kind of secret security people orgy where everyone is dressed like they're in The Matrix." The volunteer who fixes it gets nothing. After the fight, Google started sending patches and added rewards for fixes. Microsoft Teams, meanwhile, posted on the public tracker that its issue was high priority because Teams has users. After a request for a long-term support contract, the offer was a few thousand dollars once. Kempf's diagnosis is an OSPO that never told a middle manager FFmpeg is not a vendor Jira. "If you want me to be your supplier, I'm very happy. I will send you a contract."

The spicy account worked. Donations rose substantially and still do not cover one full-time developer, but awareness did. When VLC could not update on Android for more than a year because of a Play Store bug, the message that produced a human was a tweet threatening to stop shipping to about 100 million users. The Windows Store took the same treatment. VLC is, Kempf believes, a top-ten Windows program, yet he has no Microsoft ISV contact, unlike Adobe or Spotify. "Unfortunately, we are so small that the only very strong power we have is blaming on social network, because it snowballs." Kunhya calls the genre a rap battle or WWE. When a thread with OpenBSD's Theo de Raadt went too far, Kempf got him on the phone and cooled it down. Ruikai Peng, sixteen, found a bug in git, fixed it in three days, minted no CVE, and built no logo site. "The kids are all right."

Why contribute? Kempf's first answer is that multimedia plays video, and video is cool; a large share of the community arrived because they loved anime. His standing advice is to work on something you love — he is on VLC because he loves movies, even when his wife hates the rewatches. The second reason is school: if you can write C and assembly that survives FFmpeg review, you will be a better programmer even in TypeScript. Andrew Kelley started Zig after that school. You have 16 milliseconds to put a frame on screen; miss one and the picture falls apart. You also have to learn computer architecture — pipelines, SIMD, ALUs, I/O — because a three-cycle difference on one CPU generation changes the output. Kunhya estimates FFmpeg running on the order of 100 million to a billion CPUs at any moment. You can tell your grandmother you help her watch films. You cannot tell her you built a portal for PG&E invoices.

He calls FFmpeg and VideoLAN a binary star: VLC is to FFmpeg as Android is to Linux. More than 80 percent of FFmpeg encode pipelines, he guesses, depend on x264, a VideoLAN project that became the software H.264 encoder of the internet. VLC feeds FFmpeg weird files and has historically funded development. A Windows VLC build is about 16 million lines of code; one million live in the VLC tree, about two million in FFmpeg, the rest in neighbors such as libdvdcss, dav1d, and checkasm. The X meme that "it's just FFmpeg inside" misses the loop. Eras of FFmpeg, in Kunhya's telling: Fabrice Bellard founded it; Michael Niedermayer's 2000s made native DivX, Xvid, Windows Media, and Real decoders so you did not need spyware codec packs; the late 2000s brought H.264 and HD; the 2010s belonged to reverse engineers.

Kostya Shishkov, a Ukrainian then in Germany, treated a binary as a specification. A one-megabyte blob is roughly a month of work; he did 20- or 30-megabyte blobs. GoToMeeting was VLC's top feature request, so Kempf posted a bounty. Kostya delivered in about two months, recognized DCT patterns from WMV, and filled the source with jokes on Kempf's name. The method is archaeological: find the decode module, hook it, dump YUV as a bit-exact oracle, then hunt entropy and DCT in a disassembler, often inside a VM because of DRM. For a long time you see nothing on screen; you debug in memory. MPEG-2 in the 1990s was not bit-exact across implementations, which Kunhya, citing Yuriy Reznik, calls one of the industry's great mistakes. From the 2000s on, Lex's decoder, JB's, and Kieran's must match bit for bit. FFmpeg, unlike browsers, won almost the whole codec map.

FATE — FFmpeg Automated Testing Environment — is a volunteer farm of compiler, OS, and instruction-set combinations. macOS alone fans out into iOS and tvOS; the board has included RISC-V, PowerPC, DEC Alpha, Visual Studio, Intel's compiler, Apple Clang, and GCC. Kunhya hosts the Macs at the top of fate.ffmpeg.org in his office. The C is complex enough that compilers miscompile it, and a one-bit error cascades through predicted frames into a visible glitch. VLC, riding that matrix, still runs on Windows XP and Windows 11, macOS 10.7 through the latest, iOS 9 through current, Linux, BSD, Solaris, and OS/2 — a platform with maybe ten users, one of whom maintains the port. A handful of people support more operating systems than Microsoft, Google, or Apple. The nightmare is iOS: a Frankenstein mix of old Xcode and SDKs to keep Arm32 iPad 2 boxes playing movies. That, Kempf says, is a choice not to force a hardware upgrade you could have optimized away.

SIMD assembly is the lost art. Scalar add puts five on ten and gets fifteen; SIMD adds five to sixteen numbers in one instruction, which maps onto a pixel grid. FFmpeg does not hide behind intrinsics. Registers are allocated by hand. Speedups are 10x to 50x, not 10 percent; one of Kunhya's format-conversion kernels is 62x faster than C, because 10-bit video does not fit natively on a CPU and must be packed. At Open Broadcast Systems he keeps a 5×5 or 6×6 matrix of every packing format to every other, each handwritten, each covering multiple CPU generations. Most companies would have stopped at C.

dav1d is the exhibit. When Alliance for Open Media launched AV1, even member companies said the format was too complex for software. Kempf, Ronald, Henrik Gramner, and Martin Storsjö built a decoder anyway: 30,000 lines of C, 240,000 of handwritten assembly, motto "every cycle matters." FFmpeg has about 100,000 assembly lines for all codecs; this one decoder has more than twice that. About 30 percent of Netflix and 50 percent of YouTube is now AV1, on the order of three billion devices, many without a hardware decoder. One or two cores decode 720p. Kunhya tweeted the mix — 79.9 percent assembly, 19.6 percent C, 0.5 percent other — as "what peak video codecs should look like," and watched software Twitter insist for two years that auto-vectorization would catch up. It has not. The gap is multiples, not five percent.

They abuse the machine: cryptography instructions for video, custom calling conventions inside dav1d so they need not spill unused registers to L1 and L2, runtime detection across AVX-512, ARM64, SVE, SME, and RISC-V. Loren Merritt's x86inc already let one assembly file target four x86 ABIs; dav1d goes further because calls stay inside its own binary. Kempf has never heard of another mass-market project that does this. High-frequency trading cares about instructions, but not on a billion laptops. For a Python-and-vibe-coding listener, the analogy is LLM quantization to FP8, FP4, even Microsoft Phi at 1.5 bits: you are memory-bound, you cannot always buy a bigger box, so you profile the code that runs billions of hours a day. Glue on the FFmpeg CLI is not worth 240,000 lines. The hot decoder is. "The core thing that you will not be able to vibe code," Kempf says, "are optimization for the hardware."

Kunhya's asm-lessons repo exists because textbooks teach assembly as grammar. He wants the FFmpeg way: solve a real vector problem, with a bar of high-school math and pointers. Sixteen-year-olds, including Daniel Kang years ago, have landed patches. He likens the work to flying a Spitfire: the stick is cabled to the ailerons, nothing in between, and you push the airframe past the designer's idea of possible. Language models know more assembly than he expected and still fail bit-exactness, because there is no Stack Overflow corpus. Henrik Gramner, in Sweden, knows Intel SIMD latencies so well that Intel tells VideoLAN to ask Henrik. Martin Storsjö, in Finland, writes ARM NEON on a phone's virtual keyboard while watching his kids in a playground.

On Rust they split. Kunhya values memory safety as a goal and dislikes the self-importance: Esperanto, or Sinclair's C5, the electric tricycle that was going to replace the car. Prototypes are easy; the last 1 percent of coreutils takes 99 percent of the time. One compiler, ABI breaks, and a rewrite that is only 85 percent as capable are not a migration plan. Kempf has shipped Rust modules in VLC and is writing his new company, Kyber, mostly in Rust. He calls it a better C++ for ownership, superb on greenfield parsing and networking, and poor at interop. His career rule is almost never to rewrite. Writing is an order of magnitude easier than reading; LLMs confirm it. You will not understand your own cleverest code in six months, and a newcomer who cannot see undocumented trade-offs will want a clean slate. That instinct is usually wrong.

Even a Rust rewrite of dav1d's C would collapse at the assembly boundary, because handwritten SIMD can jump anywhere in memory. What they actually need, he argues, is compile-time-checked assembly — the spirit of checkasm, which instruments SIMD at build time and compares it bit-for-bit with a C reference. Without that, 240,000 lines are unmaintainable. The 2011 Libav fork, which looked from Ubuntu's packages like a civil war, was about governance. Most of Libav merged back; review culture settled; FFmpeg got stronger. Burnout worries Fridman more than forks. Kempf points at curl's Daniel Stenberg on "AI slop," at the XZ maintainer who was harassed into giving an attacker commit access, and at libxml and the IANA time-zone database as single-person load-bearing walls. FFmpeg's ten or fifteen people are, relatively, a crowd.

Around 2009 or 2010, with four people near VLC 1.0, Kempf dropped new PowerPC builds after Apple's Intel switch. A letter arrived with powder and the line that he was a piece of shit who should die, PowerPC forever. Anthrax scares were still in the news. It was flour. His mother freaked out; they went to the police. He is, he says now, almost glad it happened then, because it taught him to take hate. Fridman answers that he has smiled thousands of times because of FFmpeg and never had a place to say thank you until the X account existed. Kempf still answers strangers on Reddit. The developers still live on IRC: no ads, no tracking, no threads, no emoji reactions, and it works.

H.264 still dominates internet video and Blu-ray. Kunhya, who contributed to the standard's open encoder, says the leap was timing — HD, Core 2, Nehalem — and a revolt against twenty years of PSNR worship. Minimizing mean squared error spreads a little blur everywhere, so "good" scores looked mushy. Hobbyists encoding personal video, mostly anime, built psychovisual rate-distortion from block energy and adaptive quantization that stole bits from busy grass and gave them to faces. ParkJoy, a no-expense-spared Swedish television HD film of people running by a river, remains the public sequence that sorts encoders: high-PSNR modes smear it; psychovisual modes look better and score worse. There was no VMAF. Loren Merritt refused a $30,000 reference monitor; he wanted it to look right on someone's laptop.

Kempf's torture clip is a Planet Earth flock that becomes millions of birds; on YouTube you can watch the encode fail. Anime banding was another Loren project. x264 is still the reference every AV1, AV2, VVC, and HEVC encoder is compared against. Chris Henderson at Warner Bros. used it for the entire *Fringe* Blu-ray box, betting his job on a free tool. Kunhya still buys discs for the most cinematic films. The encoder was born at the same École Centrale as VLC, started by Laurent Aimar, then carried by Loren, Jason, Måns, Henrik, Anton — most of whom never met. Before Crunchyroll, fansubbers ripped Japanese DVDs, translated them, and needed tools; Aegisub and ruby (furigana) stacked on English or French on the same frame. Gradients that are cheap on a screen and rare in live action became an encoding specialty.

From MPEG-2 through DivX, H.264, HEVC/VP9, and AV1, the concepts — inverse transform, intra, motion, entropy — stay. Each generation is roughly 25 to 50 percent more compression; AV1 can save 40 to 60 percent of H.264's bandwidth at a given look, at about two orders of magnitude more encode CPU. YouTube still encodes almost everything in H.264 and re-encodes popular titles in AV1. I-frames are complete pictures, like JPEGs. P-frames fetch blocks 5, 7, and 42 from the past. B-frames can depend on the future, so decode order is not display order — the fact that first blew Kempf's mind. FFmpeg's default H.264 GOP is about 250 frames, several seconds of prediction that still plays smoothly. Kyber often uses intra-refresh: one I-frame at the start, then a rolling refresh so bitrate never spikes. Thousands of people at YouTube, Netflix, and Meta do not write codecs; they hunt parameters for movies versus phone clips versus screen recordings versus editorial mezzanines. Google Video, before the YouTube acquisition, even ran VLC inside Internet Explorer via ActiveX. Today the joke is inverted: VLC compiles to WebAssembly and runs in a JavaScript VM.

WikiLeaks Vault 7 showed a CIA build of VLC that was the genuine binary plus a plugin, remembered as psapi.dll, which read the Documents folder, encrypted it, and sent it out. The social engineering was elegant: you watch a film for two hours without touching the machine; HD already makes fans roar; nothing looks wrong. Chinese operators used a similar trick against Indian users, got VLC banned in India, and forced Kempf into court to prove the real player was not the implant. Attackers reused a correctly signed DLL while pointing vlc.exe at a fake libVLC. There is, he says, not much they can do about impersonation. A German fake site has been reported for twelve years. Google knows the payload, Kempf says, but the binary is too large for its malware analyzer. In Germany the German-language clone outranks videolan.org. For three weeks the install does nothing; then a service wakes and pulls spyware and adware. Search will not protect you. Check the domain.

Sandboxing is hard because VLC is a small core plus about 500 plugins, including GPU drivers from Intel, NVIDIA, and AMD. A crash in a shader or in FFmpeg should not hand an attacker the user's files. Mobile apps are already jailed; a single desktop sandbox with every permission punched through is theater. They are splitting decode, demux, and filters into processes, Chrome-style, except the pipe has to carry hundreds of megabits per second, not a ten-megabyte page. At that user count, someone uses every feature. A high-school math teacher in the south of France landed a Bézier-curve puzzle filter in 2010 so bored spouses could click jigsaw pieces over a long French film. Five years later a user emailed that 16×16 was too easy. Kempf's commit raised the cap to 256×256. ASCII art output over SSH, he adds, is how you debug a multicast network: black or not black, from a router with no GUI.

Adaptive streaming, after all that codec romance, is mostly a CDN congestion problem: if a segment takes more than half its duration to download, drop a rung; going back up is the hard part. Viewers forgive video quality swings and even 30 fps sports; they notice at once when AAC falls over to spectral band replication and the audio goes thin. Live plus DRM plus every device is operationally hard, Kempf says, but 10–20 seconds of latency is not deep science. The deep science is what happens when you refuse those seconds.

Kyber is Kempf's next descent: from encoding files, to streaming, to interactive control of drones, humanoids, remote cars, and cloud games. Quality stops being the point; a millisecond matters when you drive. Waymo, he notes, still has remote operators for the fraction of the time autonomy fails. Kyber uses one QUIC/UDP connection, encrypted by default, carrying audio, video, and mouse, keyboard, or gamepad together, with server-side re-timestamping because clocks drift — one camera is easy, five or six is not, and robot-learning datasets have to stay coherent on replay. Forward error correction sends a few extra percent so lost packets rebuild without a round trip. The goal is 4 ms glass-to-glass, 240 Hz. They have 7 ms Windows-to-Windows or Windows-to-Mac, of which about 3.5 ms is NVIDIA's encoder and 2 ms Intel's decoder. Custom SDI hardware can go lower; he wants it on a Jetson Nano or an N150 over the public internet. The license is AGPL plus commercial. CES showed a 3D-printed rover in Las Vegas driven from France on a webcam and a small board.

AV2 is the next Alliance for Open Media codec, about 30 percent better than AV1; they demoed it in VLC at CES and plan dav1d2 (he wants to call it Devid). H.265 is HEVC, H.266 is VVC; ISO is private, ITU is the UN, so every standard wears two names. Generations average 30 percent, with screen content sometimes at 70. Multimedia is a patent minefield beside cellular RF. After H.264, HEVC sprouted tools useful 0.1 percent of the time so someone could hang a patent on them, then MPEG LA, HEVC Advance, and Nokia outside the pool. Uncapped fees could mean hundreds of millions a year for YouTube or Netflix — enough to justify inventing AV1. MPEG does not discuss patents in the room; AOM does patent review. AV1 even renamed a B-frame variant an S-frame, "I can't believe it's not butter." VLC sits in France, which rejects software patents; Kempf once calculated full licensing at more than €200 per user. Two intelligence agencies asked for a VLC backdoor. He said no, less politely: if they had to compromise the software, they would shut it down. Builds happen on never-networked machines that compile the compiler first, with double signatures, after a non-Western agency tried to push a fake binary onto their servers. There is no telemetry, so no government can ask who watched what. Legal VideoLAN is one person. Lawyer letters arrive weekly; fees stay under $10,000 a year. Worst case, someone takes $50,000 from the nonprofit. The source is already out.

The archive world, led in Kunhya's telling by Dave Rice at the City University of New York, treats FFmpeg as a Rosetta Stone for a thousand years. Britain's BBC Micro "New Domesday Book" needed reverse engineering after two decades. They funded FFV1, a lossless codec with GPU encode and bit-flip resilience so a tape error does not kill a whole GOP, and they teach FFmpeg commands to children in countries without institutions. C, Kunhya hopes, will be Latin: learned from the past, still runnable. Film from the 1930s–50s is already disappearing; 1970s–80s tape has fewer working heads than reels. Someone has to choose what to throw away. In a thousand years we may be drowning in AI slop; this is the transition from data scarcity that is worth keeping. FFmpeg will be here in a century, both men say. VLC maybe. Multimedia, Kempf defines, is a digital representation of several human senses, so odor tracks, stereo nostrils, and the haptic plugin VLC already has for 4D hydraulic seats all belong. Neuralink gets `ffmpeg -i` with input format human brain. Standards will bloom, die, and return the way 3D audio did. Dolby, once the company that defined sound, is in his view mostly lawyers now.

Kunhya's favorite FFmpeg tweets are "Talk is cheap, send patches" and "FFmpeg, nothing is beyond our reach," the second from a U.S. military satellite. The library compresses images on NASA's Mars 2020 rover; they wanted commercial off-the-shelf tech, so the project is multi-planetary. VLC shows up in Formula 1 paddocks, at ESA, at SpaceX launches, and at CERN, where analog cameras around the 27-kilometer LHC ring are captured into VLC and multicast site-wide. In 2010 Kempf and Laurent fixed a poorly documented parameter in an hour and spent the rest of the day among antimatter and colliders. Asked about regrets, Fridman quotes "Je ne regrette rien." Kempf answers that regrets are a tax on your mind: learn, but do not pay rent to a past you cannot time-travel. He objects, deadpan, to being compared with Johnny Depp. Fridman closes with Linus Torvalds: most good programmers program because it is fun. The cone stays on.