Cristos Goodrow: There Is No YouTube Algorithm, Only People Moving Through Video

Lex Fridman opens with the scale that makes the rest of the hour necessary. YouTube has about 1.9 billion users. People watch more than a billion hours a day — more than Netflix and Facebook video combined. It is the second most popular search engine after Google. Creators upload over 500,000 hours of video every day. A typical human lifespan, for comparison, is about 700,000 hours. What lands on the platform in one day is enough for a person to watch from birth to death.

Fridman’s absurdist question is whether, from birth, there exist trajectories through that space that maximize happiness, education, or growth. Cristos Goodrow, Google VP of engineering and head of search and discovery at YouTube, will not recommend anyone spend all their waking hours on the site. He will say there are great trajectories. His oldest daughter watched Tyler Oakley and the Vlog Brothers for years, with a real effect on her character. His younger daughter, a ballerina, can pause and rewind professionals doing the same routine her teachers assign. His son got through college linear algebra on 3Blue1Brown, in a way a chalkboard would struggle to match.

Fridman pushes on time horizon. Interests change; shows get dropped. If YouTube wants people to stay and grow over years, it cannot optimize only for this month’s engagement. Goodrow agrees. Introducing diversity — getting someone who is watching one thing to try something else they might like — is a problem his team has worked for all eight years he has been there. Random selection from billions of videos is the trivial version, and it fails: the draw is often not even in your language. The useful version sits in a narrow band, not too similar to what you already watch, and still something you might actually click.

Diversity of content, Fridman says, is almost impossible to define. He offers himself as the example: a 3Blue1Brown fan who did not know Derek Muller’s Veritasium channel existed until YouTube showed it to him. That is the first, easier step. You are watching math explainers; the system throws in a couple of other science or math channels instead of six videos from the same one. Goodrow says that was the first kind of diversity they shipped, years ago, almost everything operating at the individual-video level rather than the channel.

The bigger leap is embedding videos into clusters, then measuring a subtle conditional: among people who watch cluster A, how often do they also watch a very distinct cluster B? The relationship can surprise you. People who watch science also like jazz. Once that co-watching is measured, the system can occasionally recommend a jazz video. The likelihood that a science viewer likes jazz is higher than the likelihood they like backyard model railroads. Aggregate those hops, see whether people stay engaged, and you have a basis for a recommendation that is a real leap rather than a shuffle of the same shelf.

Fridman will not stay in the safe math cluster. He read Ayn Rand in college, was born in the Soviet Union, and treats both Objectivism and communism as ideologies worth exploring — communities that are certain they have the path to utopia. How does YouTube manage politics and ideas, both in what to recommend next and in what is allowed on the platform at all? Goodrow calls getting this right the team’s top priority. Freedom of speech is not the right to say literally anything; society already accepts limits such as libel. Where they can draw a clear line, they do, and they keep evolving it. Wherever you draw it, a borderline remains.

On that border they often do not remove the video. They reduce recommendations and try to raise what they call authoritative or credible sources. Ayn Rand and communism, he says, are two valid points of view people will debate. YouTube is not trying to settle which is right. It is trying to make sure the people expressing those views are authoritative and credible. Who defines those words in political philosophy is the live, heavy question. Susan Wojcicki, then CEO of YouTube, has a line Goodrow finds inspiring: do the job so that people twenty and thirty years from now look back and say they found the balance between openness and responsibility.

Trolls are the other gray. Mock-the-emperor comedy can be brilliant and can turn cruel. Creators, Goodrow has heard many times, have to become resilient to a certain amount of meanness. YouTube’s tools are comment ranking that tries to surface constructive conversation, and product features that let people block and manage. He hopes they get to a world where uploaders do not have to suffer it; he also thinks some ability to deal with it will remain part of being a creator. The whole project, as he frames it, is making YouTube great for kids and for society without pretending the responsibility is optional.

Fridman wants a split: how much of curation can be pure machine learning on the data, and how much needs people in a room deciding what truth and which ideals to promote. Goodrow’s experience is that you need both. Machine learning’s first hunger is data, and the data is made by humans. To find misinformation or borderline policy violations, the first step is getting people to decide which videos sit in which category. That labeled set is then extrapolated across billions of videos. Without humans you cannot start; without models the humans cannot finish.

A lot of the work, he says, is people at YouTube arguing through policies and checking whether the outcomes of those policies are the ones they want to see. Once there is consensus, evaluators and reviewers generate training data, and the models take it further. Fridman, coming from autonomous driving, notes that annotators are uneven — better at cars than at trees — and asks about bias in video reviewers. Goodrow is frank. Some biases they instruct: a bias toward scientific consensus, toward demonstrated expertise, credibility, authoritativeness. Other biases they try to remove, by sending the same item to many reviewers and by drawing those people from different backgrounds and regions. Unfair biases can still creep in from the training data itself, especially around protected classes, and they have worked on the models to reduce those.

Fridman finally asks what the so-called YouTube algorithm actually looks at. Goodrow splits the problem. Search borrows Google’s retrieval stack: does the query appear in the title, plus richer syntactic and semantic matching, plus a behavioral signal — is this video watched a lot after this query? Watch-next is a different system, and the first attempt that really worked was collaborative filtering. Observe which videos get watched close together by the same person. Draw a graph in which videos co-watched by many people sit near each other and rarely co-watched videos sit far apart. They call it the related graph.

The graph grows structure nobody programmed. Same-language videos clump. Sports clump. Most music videos clump. There is no language feature, no sports ontology. It is an outcome of how people on YouTube behave. That already solves a huge amount of the lowest-hanging fruit across millions of videos. A researcher once proposed a project on the theory that YouTube could not possibly recommend well to bilingual users. She was an academic in the United States: search a scholarly topic, get English, watch-next stays English — so YouTube must think she speaks only English. She is from Turkey. She searched for baklava, played a Turkish cooking video, and watch-next flipped to Turkish. She could not believe it. The English and Turkish regions of the related graph are adjacent but separate, because enough people binge in each. When she jumps, the graph jumps with her.

Fridman would love a map of his own clusters — years of cat videos and people falling over as a mirror. YouTube already ships watch history, and the team has at times experimented with showing people the topics they have lived in. The deeper point, Goodrow says, is that the recommender *sees* a user as exactly that history. You can think of yourself as a DNA strand of every video you have watched, or as a vector in the space of all YouTube videos. Nearby vectors are people with similar taste. A video one of them loved and you have not seen is a candidate for diverse recommendation. Two people can click the same video and get different watch-next lists, because their vectors are not in the same place. A new user looks bad because the vector has not formed yet.

Fridman has watched videos with fifty-seven views that he thinks are extraordinary, and wants to know how quality gets seen. Goodrow says it depends on the realm. News and medicine lean on journalism departments, fact-checkers, authoritativeness. The highest-quality prank or Minecraft video might be the one people enjoy and finish — or the one they still endorse when you ask them the next day. Since he arrived, the entertainment side of YouTube has been chasing better measures of quality, satisfaction, and enrichment.

The first approximation was viewers. Things can get a lot of viewers and still be junk, especially if people click and immediately abandon. So they moved to time spent watching, on the premise that time is related to value, even if not perfectly. That was a real step and still not enough. Goodrow has himself clicked through late-night television and ended up watching *Under Siege 2*. Asked the next day whether he was glad, he would have said he wished he had gone to bed or read a book or almost anything else. A few years ago someone got the idea to survey users afterward. That feedback trains the models not just to predict what you will click or sit through, but what, tomorrow, you would give four or five stars.

The signal list is already rich: clicks, views, watch time, relative watch time, likes, dislikes, comments, shares, and those surveys. Subscribe is another, and it is messier than it looks. User research found people who watch a lot from a few channels and never subscribe because they “can’t afford to pay.” The team has to explain that it is free — it just tells YouTube you care about this creator. Other people subscribe to many channels and watch almost none of the uploads. They wanted to give the person a high-five. Goodrow does this himself: amazing person, outside his interest area, subscribe as congratulations, never click the videos. Fridman admits the same. You cannot close your eyes and say they used it wrong. You have to embrace every way the world uses Subscribe, Like, and Dislike.

Titles, descriptions, and keywords remain a primary handle for both search and clustering. Understanding the pixels has been a project for as long as Goodrow has been at YouTube, including public computer-vision datasets. Ability is still crude. The system can tell a music video from a sports video, and that people are playing soccer. It probably cannot tell Manchester United from his daughter’s team. That coarse label is still useful: it informs the clusters, and it can add the word “soccer” when the title forgot to. Creators, he asks, please help with the title and description.

A few years ago YouTube live-streamed an important World of Warcraft competition. Type “World of Warcraft” into search and you would not find it. The title said match seven, team A versus team B. “Warcraft” might have been there; “World of Warcraft” was not. Fridman notes that being literal on the internet is uncool — wit and indirection are how great titles work, and models want you to just say what is in the thing. Goodrow agrees you have to play with both, and still: if you want humans to pick your video in search, put the query in the title. A results page full of “World of Warcraft” plus one coy exception looks, even to a person, like a search mistake.

Fridman pushes back: a title that just says “Winning,” with a sad orc thumbnail, would be a more beautiful experience if the algorithm were smart enough to know it was WoW. Goodrow’s realism is that some people will skip it, and the ones who do not will be delighted. Collaborative filtering itself often starts with people finding videos by searching, so discoverability is the first ant in the colony. He does not have a sense of how large the explorer class is versus people who watch the same thing forever. He does know there are people and machines trying to glue unrelated videos together because it is profitable, and the systems have to stay resilient to that. Clickbait thumbnails sit on a similar line. Derek Muller has said he will keep making high-quality videos with clicky titles. Goodrow’s analogy is his bookshelf: yellow identical Springer math texts versus pop-science books with provocative covers. A compelling cover is not a crime. All-caps, too many exclamation points, racy thumbnails that offend users — those get depressed for the people they offend.

The “I don’t want to see this” control is, Goodrow says, an absolutely strong signal. Fridman loves it the way Tesla owners feel they are training Autopilot, or Wikipedians feel they are building something together. YouTube does not always remind people that a like is helping the whole system, not just their homepage. Goodrow takes the note. Personalization underneath is very strong: two logged-in people on the same video get quite different watch-next and homepages, because sometimes you want more of this genre and sometimes you are done and ready for the latest from a channel you actually subscribe to. Success, at the limit, is that every surveyed watch is five stars, and if you asked whether it was the best video they had ever seen on YouTube they would say yes every time. He does not expect to hit that. He would take the asymptote. Coming back another day is the simple indicator that someone found value. Hours watched, by themselves, are not the point.

Asked how YouTube is changing society, Goodrow starts with openness: unlike older media, nobody at YouTube decides before upload whether your video is worth existing. Tyler Oakley has said he would not have had an audience without that. Colleagues from places with low literacy tell him YouTube can teach a job or a repair without requiring someone to read. After almost nine years on the job, strangers who hear where he works always say they love YouTube. Not once have they praised search or recommendations. They name a channel, a creator, a topic, a community. That taught him the product is the video and the connection; everything else should get out of the way.

Creator versus video matters for retention. If YouTube were only the next viral clip — a cat on a piano one day, children interrupting a reporter the next — you would have no idea what you were coming back for. Fans and creators are a relationship that can play into the future. Fridman asks about burnout, the highs and lows creators talk about, the fear that if you stop the party leaves. Goodrow says the first job is making sure YouTube’s systems are not causing that pressure. Research and observation: taking a break is absolutely fine. They have as many examples of people coming back more popular as of the opposite. The myth that a pause necessarily means the channel dies is one he wants to break. Looking after yourself is job one. Burned-out work is usually not your best work; a break can improve the videos that follow.

Fridman asks whether “the algorithm” is closer to a self-driving stack of patches and heuristics, or to one giant model with humans tweaking the edges. Goodrow does not usually think of it as the algorithm at all. It is a bunch of systems on different services: code, machine learning, heuristics — married to the behavior of all the people who come every day. If nobody showed up tomorrow, it would not work. When outsiders say the algorithm did this or that, it is sometimes just viewers doing that, and the machinery keeping track and reacting in finer-grained situations.

The evolutionary path looks like most production ML. You start with a simple heuristic: recommend the most popular videos. You collect data, make fewer mistakes, and chip the heuristic away in favor of something that can learn from past situations. Diversity’s first version was almost comic: not more than three videos in a row from the same channel. It worked. Who needs six? Over time they made it more fine-grained, replacing the global rule with something that can react to individuals. Nearly every change ships only after a large A/B experiment, running from a week to months, measuring hundreds of variables with confidence intervals. Fewer dismissals, more five-star survey answers, are the kinds of movements that count as an improved viewer experience. User studies sit alongside. They also watch impact on channels before launch.

How do you make a viral video? Goodrow does not know. They tried to detect virality forming — first and second derivatives of view count, whether traffic was coming from Twitter rather than search — and never got very good at it. Manufacturing one is even further out of reach, unless you get your kids to walk into a live news shot. After the fact they can see where it was recommended, who watched, where it was found. People do expect that if something they would like is going viral, YouTube will show it to them. Meeting that expectation is the point, not chasing virality for its own sake.

Derek Muller’s *Why Are 96 Million Black Balls on This Reservoir?* hit about 30 million views in a few days and was how Fridman discovered Veritasium — a sudden excited-nerd thumbnail in a diet of MIT lectures. From inside, Goodrow says, the system did what any recommender would do. Show it to some people. They like it. Find people a bit like them. They like it. Expand the circle. Keep going until the feedback says you have gone too far. Copying the black-ball topic would not make your video canonical; Muller’s might have been the one. Clipping, though, he will endorse without mystery. An NBA upload can be found as “Warriors versus Rockets” or as “Steph Curry.” A highlight of the shot is another door. You never know how people will search for what you made, so cut clips, add titles, and make finding it as easy as possible.

Fridman dreams of a system that auto-clips the exciting bits of a soccer match *and* of a forty-topic conversation. Goodrow’s team has tried. It depends on understanding the video, which is still crude. Goals in a soccer match might be findable; who scored may still need a human. Coherent topics in a conversation transcript are not something they will be good at right away. Google Search already jumps you to the dishwasher-filter segment of a long video, but that path, for now, is based on creator annotations. Automatic versions are in progress. A lot of YouTube systems start almost entirely on audience behavior; content understanding arrives later as refinement.

The hard specific of video, versus language, is volume and precision. You are classifying into something like a million classes whose distinctions are tiny: a jersey number, a chest logo, whether that was a goal, whether the moment was interesting. Fridman cites Yann LeCun’s bet that self-supervised next-frame prediction on enough video yields physics and common sense. YouTube already has systems that watch all the videos, Goodrow says, but they are supervised classifiers looking for specific things. He is not sure next-frame prediction gets you to understanding. That is roughly what video codecs do: predict the next frame and fix the residuals. There are so many frames that a tiny per-frame error can lead to wildly different videos. The dream compression is a sentence — two people talking, about this — and that is the open-ended hope. The world is more complicated than it gets credit for. They have been trying to summarize videos in text for at least eight years. If eight years ago was 0% solved and done is 100%, they are maybe less than a quarter of the way.

Ten, twenty, thirty years out, Goodrow thinks YouTube is evolving to take the place of TV. He grew up in the 1970s watching a tremendous amount of television; people told his mother it would wrap his brain and she should kill the set. For his own family, YouTube is a better version: on demand, tailored to what each kid wants, and able to surface things television would never have scheduled. That is the direction he sees from watching his house, people using it the way he used TV as a child.

What still excites him in search and discovery is introducing people to more of what is already there. He had worked at YouTube eight years before he learned, only last year, that he could watch USC football games from the 1970s — content that was on the platform when he arrived. Seven years to notice. The opportunity is to let more people know what YouTube actually holds, while stripping the negative consequences of openness, and maybe, eventually, getting every watch to feel like the best video someone has ever seen. Virtual worlds are not something he thinks about much; Google does, and he would want YouTube present if VR ever takes off. He has not seen it take off yet. Fridman closes with Proust: the real voyage is not new landscapes, but new eyes.