The Speech Company That Gets Paid to Label Its Own Training Data

Lex Fridman starts with science fiction. Dan Kokotov does not hesitate: the greatest sci-fi novel is *Dune*, then *Children of Dune*, then *God Emperor of Dune*. The six Frank Herbert books are one chronological story and should be read in order. The later books by Herbert’s son, he says, are not worth the bother. Lex has never read them — one of his bigger regrets, with a new film incoming — though he did play the 1990s real-time strategy game. Kokotov remembers that game from after he left Russia; in Russia, in 1991, Pong was about as good as it got.

The philosophical hook is book four. Leto has merged with a sandworm and become a tyrant who oppresses humanity for a long stretch, on purpose. People had gotten too comfortable; civilization had stagnated. The suffering is meant to force an explosion outward, new worlds, renewed momentum. “You need a little bit of pressure and suffering… to make progress, not get too comfortable,” Kokotov says. Lex, who left Russia as a child himself, is a self-described fan of suffering. The David Lynch *Dune*, confusing and controversial, still captures the book’s mood better than most adaptations. Both men are skeptical of the new trailer.

Lex has used Rev for years because Upwork transcription was a pain: browse profiles, guess quality, chase formats. Rev takes audio, video, or a YouTube link and returns a standard caption file a day or two later. He groups it with tools that erase a pain you barely named — iZotope RX for mouth noise and gear hiss, Adobe Premiere, AutoHotkey, Emacs. “I just like talking to people who create things.” He notes, carefully, that Rev later offered to sponsor the show; this conversation was not bought. Integrity, he says, is not for sale.

Rev.com is the human, end-user service. Rev.ai is the ASR brand, like WordPress.com versus WordPress.org. CEO Jason was an early Upwork employee and wanted to kill the browse-and-pray loop. Buyers on Upwork were not experts, so they could not pick well; freelancers spent half their time dressing a profile they were also not experts at. The bet: pick a few standardizable verticals, take the file, hide the details, return the result. They started with translation, then added audio transcription, later adding rush processing.

The vertical had to be computer work from home, not TaskRabbit, and simple enough for a clean interface. Programming projects are unique; five hours of audio is five hours of audio, a five-page document is a five-page document. Early translation specialized in birth certificates and immigration paperwork, where quality is obvious. Transcription is priced per audio minute. Translation, when it was the main business, was per page or per word. Focus now is speech-to-text and language services generally.

Lex wants to test Russian. Kokotov thinks subtitles currently run English out, not the other way. After COVID, Lex plans to sit in Paris with the famous couple who translate Dostoevsky and Tolstoy. He has Russian speakers he wants on the show, including mathematician Grigori Perelman, who speaks no English, and — he is determined — Vladimir Putin. His Russian can understand; it cannot be witty. “I’m already an idiot in English. I’m an extra level of awkward idiot in Russian.”

The interview shape he is circling is a three-person dance: questions in English, live translation, answers in Russian, translation back. He wants to start with literary translators, who live in the nuance, then ask what real-time speech could do. Kokotov’s father writes poetry as a life passion beside a day job. In America he began translating English verse into Russian and later the other way. When Louise Glück won the 2020 Nobel, he was one of the few who had already put a book of hers into Russian, and Radio Liberty asked him on air about the craft.

The hard part is word choice. Russian has inflections and genders English lacks, which is one reason machine translation between the two is uniquely painful. English, meanwhile, has a much larger vocabulary, so the emotionally right word often does not exist. Russian, Kokotov says, plays with words more. Literature is the extra level of difficulty above an interview, and an interview is already hard.

Rev is a two-sided marketplace: customers on one side, “Revs” — their word for freelancers — on the other. Too few Revs and jobs slip. Too many and people log into an empty queue and leave. Kokotov says they have been tuning that balance for years and are still refining it. Scaling, here, is marketplace scaling, not a single algorithm.

He does not like “gig economy.” In music, a gig is a real term; in labor it sounds amateurish and light. What they think they are doing is improving work-from-home: strip geography and the clock. Work-from-home parents who refuse nine-to-five but need income; people who want a cabin on a mountain and a paycheck. The old world forced a choice. Most Revs are in the U.S. because the work is English audio. Management still flies out to meet them — recently Utah — and finds a crowd almost anywhere.

The people are mixed: students, socially anxious folks who do not want an office, a former high-powered professional on a break who took the work partly to *see* the gig economy. Lex has noticed some Revs are oddly good at machine-learning terms in his MIT lectures, capitalization and all. They Google. For a subset, research is the fun: they say they get paid to watch a documentary and learn while they type.

A typical Rev day: log into a workspace, see a list of audio, filter by length, subject, or country of accent, then open a purpose-built editor. The starting point is an ASR draft. The human’s job is correction. On a clean podcast like this one, the draft is decent. A lecture recorded from the back of an auditorium, speaker far away, lots of crosstalk, and the engine collapses; the transcriber may throw the draft out and type from zero.

Real-time is “pretty difficult.” The company asks everyone to try a day as a transcriber because the work is harder than it looks: fast talk, accents, mess. On good audio Kokotov estimates two to three times real time — an hour of tape costs two or three hours of human. Lex notes the meta joke: whoever is captioning *this* episode is listening to them guess the speed, and may be swearing that it takes longer. “You’re way wrong. This takes way longer.”

ASR, Kokotov says, is the ML problem of turning speech into a word sequence. Domain matters. Siri is limited vocabulary — play a song, order pizza — and it adapts to your voice. Unstructured conversation is another world, and in that world Rev thinks it has the best engine. Word error rate adds insertions, deletions, and substitutions: extra words, missing words, “Apple” heard as “able.” On their internal suite — podcasts, videos, interviews, lectures — they are around 14%. Humans, most people think, top out near 2–3%. The gap is still large.

Lex asks if they are trying to beat YouTube and Google. Google’s ASR and NLP teams are first-tier. Rev scores itself against Google, Amazon, Microsoft, and smaller rivals on that homemade suite, and wins on its own tests. Human captioning used to be a dollar a minute and is now about $1.25. Temi, the consumer machine transcript plus an editor, is about 25 cents a minute. Upwork can be cheaper; quality is the bet you lose. Lex confesses the programmer disease of writing a Python downloader when he did not caption often enough to justify it. Kokotov’s line: if your bottleneck is clicking Rev’s download button, you are at Elon Musk volume. Zapier, Dropbox, and an API exist for people who actually are.

Three percent is the grand vision. The path is data. “Our business model is being paid to annotate the data, so it’s kind of like a pretty magical flywheel.” They train on final transcripts plus audio. They also have, in principle, every edit a Rev makes in the editor — unused, so far. Lex remembers semantic-segmentation labeling for driving: polygon clicks, repairs, order of points. A few CVPR-scale papers tried RNNs on that click trail; nobody pushed it. Kokotov’s intuition is that time-to-edit and the repair sequence are signal. That is the exciting work of the next couple of years.

Rev.ai is the last step of a path: humans only, then Temi for consumers, then ASR inside Rev’s own human pipeline, then the engine as a platform. The AWS analogy is building blocks, watch what people build, make the popular paths easier. Kokotov’s own dream is meetings. As a manager his day is meetings, and he often wants to know who said what, in what context, and cannot retrieve it. Indexed, archived, clip-shareable meetings would change the work, especially on Zoom. Lex wants the same for podcasts, where search is still keywords in titles.

The conversation veers into what it means to build something people love. Mechanical Turk’s idea felt like Rev; the interface, Lex says, still feels like a two-person committee that meets monthly. Even the API docs are dated. Google’s Maps and YouTube APIs are accurate-ish and cover maybe half the natural questions. YouTube Studio took years to allow multi-upload, and creators felt unheard. Kokotov’s diagnosis: people get fixated on numbers that are already great, and the person in the chair is a later-hired PM whose mandate is plus 10 percent, not the person who built the thing.

“You have to have a vision. You have to want people to love your stuff, because love is ultimately the beginning of a successful long-term company.” He points at Apple keeping teams small. Ken Kocienda’s *Creative Selection* is the book he recommends: not a Steve Jobs hagiography, but an engineer’s account of demoing the iPhone keyboard, being sent back because a key’s corner radius was off. Small teams own the work, so they care. Elon Musk, Lex adds, does a similar thing talking to the lowest-level engineers at Tesla.

Lex is torn on Spotify. He likes the engineers’ ambition to empower creators, and hopes they would use something like Rev to make shows indexable. Discovery is still “find a host and subscribe.” He constantly wants to recover a product someone mentioned and cannot, unless the show notes are unusually good. Graduate students write him wanting to study conversations the way they study papers. Reporters already transcribe privately to pull quotes. A public transcript would serve both the serious reader and, he says dryly, the clickbait writer.

He tells the Ryan Hall story: an MMA conversation about forgiveness used Hitler as the hardest possible case, and headlines became some version of “UFC fighter thinks we should forgive Hitler.” Close-ish, much worse in the title. He would rather make the source easy to check. Buzzsprout had just proposed an RSS extension for transcripts so any client could show the text. Lex loves the open RSS “pirate radio” of podcasting. Matt, who helps him with ads, warned that Buzzsprout wants to host exclusively — the same lock-in he dislikes in Spotify.

Joe Rogan’s Spotify exclusive, $100 million, video added, RSS gone after December 1, is the live test. Lex did soul-searching, supports Joe’s choice, and would not take the same deal. He hates money enough, and believes in the open-source spirit of the feed. The other side of exclusivity, Joe told him, is ride-or-die: Spotify gives no notes and wants him to succeed. YouTube, even for PewDiePie or Veritasium, mostly offers basic help. It sees Rogan as a headache with controversial guests, not as someone who might change how news and nuance work.

Once platforms sit in meetings about which conspiracy is dangerous and which is vanilla, Lex says, it gets weird. Kokotov worries more about over-policing. Sam Harris asked Jack Dorsey how Twitter could *not* ban the sitting president. If that question is on the table, everything is. Kokotov’s pragmatic line: if a YouTube video can convince you the government has been taken over by aliens, taking the video down will not return you to sanity.

Lex believes in people’s intelligence and still thinks platforms should encourage kindness — carrot, not the stick of censorship. “Just don’t be an asshole” will never be a Supreme Court holding. Kokotov calls it an interesting ML problem: machine learning for niceness. Dorsey talks about “health of conversation”; Kokotov doubts how seriously Twitter is doing it. The incentive is the obstacle. Rage, or its upgrade outrage, drives engagement, and these companies are judged by engagement.

Lex’s “fundamental belief”: a metric of long-term user happiness — intellectual and emotional health, not short-term clicks — would make more money and be better for society. Kokotov agrees, with a condition: it takes a patient person Wall Street trusts. Lex’s model is Jobs or Musk holding the board hostage: my way or I leave, and they cannot afford the leave. There is a reason a third name does not come immediately to mind. In a COVID stretch of dread, watching rockets keep going up with humans aboard is, for both of them, a reason to wake up.

Lex asks the “worst possible question”: how does an engineer become a manager. Kokotov’s origin is childhood astonishment — a program comes to life that did not exist. Other engineering is bounded by physics; on a computer you can imagine any program. “It’s a little bit god-like.” The first love was probably a BASIC snake. As an individual contributor, feedback is sharp: the thing is coming together or it is not. As a manager the signal is diffuse. You hope you are motivating people. You do not get the line-count dopamine. Eventually you take pride in the team’s work. “You live off other people’s dopamine.”

Humans are hard to parse. Lex’s MIT group split in two: some people did better work after hard criticism; others were so self-critical that any negative note crushed them and only celebration of small wins moved them. Rev asks first-time managers to read *First, Break All the Rules*. The philosophy is manage by exception: no standard template for how you give feedback. Figure out what drives this person and tailor it. Asked for books, Kokotov comes back to the *Dune* sequence — six books, chronological, skip the son’s — and then the dystopias.

Kokotov is reading *Brave New World* and is struck by Huxley’s sorting: Alphas and Epsilons assigned from birth. He sees a milder version now — marriages clustering by intellect and socioeconomic status more than before, society stratifying — and a meme version on social media, Dawkins’s ideas in echo chambers. *1984* is the other pole. Lex’s favorite dystopian film is *Brazil*: authoritarian incompetence, a creaky system no one will challenge, things ambling along. That, they agree, is more plausible than an efficient evil.

Movies usually write government as extremely competent, good or evil, because competence is a storytelling tool. Reality is incompetence with unpredictable consequences. HBO’s *Chernobyl* is the exhibit: information distorts as it travels up, no one takes responsibility, laziness and fear accumulate into catastrophe. Nobody in the story is born a cartoon villain. *The Death of Stalin* is the comic version of banal evil — people trying to survive, close enough to reality that Lex found it almost too true. Tarantino’s *Inglourious Basterds* felt unserious to both; *Downfall* is the Hitler film they respect.

Lex thinks Stalin’s story is more needed than Hitler’s in the English-speaking world. Hitler is a caricature of evil that feels unrepeatable; Stalin is bureaucracy that might still be running somewhere — North Korea, overreach in China. He has talked to Princeton historian Stephen Kotkin, who has read more Stalin than Stalin, and is now writing about Putin. Lex has read every English Putin biography and still needs the Russian ones. His method for a hard interview is not to write questions until the last minute. Soak, enter the mind, extreme empathy — the Will Graham move from *Hannibal*, half your time in the head of someone dark, used in moderation.

Chris Lattner’s advice to Lex was to widen the guest range to the extremes so there is room for nuance later. Social media cancels a slightly tense sentence; people still hunger for uncut hours. In COVID isolation, Lex says, someone in Saint Petersburg is jogging to this tape, someone else is on a couch, and they will comment that they are in Moscow. It is a weird one-way friendship. He “knew” Joe Rogan for ten years before it became two-way, and it is still surreal.

Kokotov’s evolutionary gloss: millions of years on the savanna, attuned to whether the tribe liked you, because a bad relationship meant being left for lions. Voice carries that information — tone, pause, intent. On Twitter you connect to an abstraction, a tweet stream, and it is easy to assign evil and dehumanize. A real voice makes that harder. “You realize it’s a real person behind the voice.”

Meaning of life, asked at the end, is a question Kokotov does not let himself sit with often. The answer he will give is contributing to the strange evolving organism called humanity — by existing, by creating something, by raising kids who are unlike you. If you are not religious, meaning is in the living and the making. Lex says even a JavaScript button on a page has an echo of that engine. Kokotov adds the religious version: if we were made in God’s image, part of that image is the drive to create. Lex closes the episode with Wittgenstein: the limits of my language mean the limits of my world — a line that, after ninety minutes on speech becoming text, is not decoration.