Wikipedia Worked Once Writing Stopped Feeling Like Grad School

Jimmy Wales tells Lex Fridman he was watching the free-software movement: programmers collaborating under free licenses because a commons removes the need to negotiate a reward for every contribution. He thought the same pattern could extend beyond code to culture, and an encyclopedia seemed the obvious first try. People already share a sense of what an Eiffel Tower entry should contain—pictures, history, location, the architect—so they can chip in sources without first inventing the genre. He built Nupedia about two years before Wikipedia, on the theory that volunteers on the internet would be laughed at unless they were *more* academic than a traditional encyclopedia.

That theory produced a seven-stage review. One of the earliest entries to survive it had to be pulled a few days later, once a wider web audience spotted plagiarism and decided the piece was not even that good. The process built to guarantee quality had certified a lemon. Wales then tried to write a short biography of Robert Merton, whose Nobel-winning work on option pricing was the subject Wales had published on in academia. Drafts would go to the most prestigious finance professors they could find. It felt like "the bad part of grad school," intimidating rather than intellectually fun, and he abandoned it.

Jeremy Rosenfeld, an employee, showed him the wiki concept in December; Larry Sanger brought the same idea. They launched Wikipedia in January as a side project, afraid Nupedia's academic crowd would hate converting the whole thing overnight. In two weeks the wiki got more work done than almost two years of Nupedia. You could be first to type that Africa is a continent and hit save. Years later Wales did the same when an economist he thought was named Robert Alman won the Nobel Prize and English Wikipedia had no article: a stub, then other people jumped in. Collaboration that had felt like a thesis defense became a geeky hobby.

The barren Nupedia years were not wasted on talk. Clever people argued on email about neutrality, what an encyclopedia is, and then-fashionable XML ideas, including a shared datum such as New York City's census population that could update across languages. That idea is now Wikidata. The software they actually shipped with was UseModWiki, chosen because it was a single Perl script Wales could drop on a server. Content lived in flat text files. Site search was Unix `grep`. It was never going to scale, but it ran.

There were no real logins: you could call yourself Bob Smith, and so could the next person. Wales quickly wrote a passworded login so identities could be unique, the floor of any trust system. Early links used CamelCase—words smashed together—which looked ugly, so they switched to double square brackets. He had not considered that a standard German keyboard has no square bracket. German users learned alt codes or cut-and-pasted a bracket whenever they found one. German Wikipedia became a massive success anyway. Tool friction, in his telling, is less fatal than whether people want the work.

The philosophical fight was sharper. Someone uploaded the full text of *Hamlet*; the community said that is not an encyclopedia article, which forced a split. Original texts went to Wikisource. Wikipedia kept the summary. Wales's phrase for the goal is "the sum of all human knowledge," with *sum* meaning summary. Boundaries drift by culture. French encyclopedias often carried recipes; you would not find a chocolate-cake method in Britannica. A canonical recipe is almost a contradiction, he says. Someone had recently told him: ten Spaniards, twelve paella recipes.

Fridman recites the May 27, 2023 numbers: 6.66 million English articles, more than 4.3 billion words, 58 million pages if you count talk and other namespaces. Wales knows the figures and still finds them remarkable in a deeper sense. He remembers English and German Wikipedia passing 100,000, and a Polish visitor at a 2006 Wikimania in Boston showing a one-volume biographical encyclopedia of about 22,000 famous Poles, most of them then missing from English Wikipedia. Completeness is a mirage. The live question is where the limit is.

Wikipedians call that limit notability. Wales dislikes the word. Fridman is holding a classic clear Bic pen: the type can have an article, with history, ubiquity, and a company that also makes kayaks and, Wales jokes, anything long and plastic. The specific pen in Fridman's hand cannot, unless a great-grandmother gave it to him, because there is nothing to say. German Wikipedians used to illustrate the point with the rear nut on Uli Fuchs's bicycle wheel. You cannot have an article about literally everything.

Biographies of living people are where they try to be much more careful, because a wrong entry can be hurtful. Wales's mother, a schoolteacher later a pharmacist, has never been in the news except as his standard example of why she should not have a page. The danger zone has an acronym, BLP1E: a living person notable for only one event, often a crime victim. The serial killer may have an article; a private victim should not be turned into a biography nobody can actually write. Academics can have career-focused entries with no personal life; royalty is almost all family. He would rather talk about verifiability than notability. Telling a mother her child is not noteworthy is the wrong frame.

Fridman has only recently learned he has a Wikipedia page. His first thought was that he was not notable enough; his second was gratitude. Then the page became a place of attack. After interviewing Elon Musk and people on the far left and far right, he says, you put a toe in politics and the shark pulls you in. One campaign claimed he was never at MIT. He says he has been a research scientist there since 2015, at the LIDS laboratory, in a paid position, and that MIT has defended him. He also corrects a family geography: his father was born in Kiev, Ukraine, not in Chkalovsk, Russia; Lex himself was born in Chkalovsk in Tajikistan, a town now called Buston.

Wales notices that Fridman spent as much time on the MIT fight, which bears on his career, as on where his father was born, which almost no reader would clock and which matters to him. People with biographies, Wales says, are often as annoyed by a tiny error nobody else sees. That is why human dignity really matters, and why Wikipedians argue about undue weight. Years ago he found a short entry on a city-council member in a small Midwestern town—he thinks Peoria, Illinois—who had served for decades. A long paragraph detailed the son's DUI. Wales deleted it. The son's arrest was not the father's scandal.

A biography with a section titled Controversies is, in his view, poor practice. If an article has seven sections and that one is short, someone looking for a task will go hunt more controversies. Better to fold disputes into the narrative so they have to earn their context. Categories are equally sticky. Wales once objected to being listed under American atheists: the fact was true, but the category read as activists—Madalyn Murray O'Hair, Richard Dawkins—people who campaign on it. His argument carried. Mark Wahlberg, he says, should not sit in American criminals for a serious youthful conviction, but the entry should tell that story, because a rough youth and a turnaround is part of the life.

The morning of the taping, someone left a note on Wales's talk page with a dramatic headline about a racist hook on the front page. A Did You Know item quoted a sports figure comparing a well-known Black American football player to a Lamborghini, clearly a compliment. The note pointed to research that Black athletes are compared to inanimate objects more often, and said the writer was not deleting the item, only raising the question. The conversation that followed, Wales thought, was healthy: not a scandal headline, but a reason to be careful about analogies lifted from the media. He saw no need to weigh in. Refrigerator Perry was mentioned as a famous inanimate-object nickname that need not become a firing offense.

Wales has been playing with GPT-4 as a style tool: it is good at flagging weasel words and rewriting toward neutrality, though it can also bleach legitimate voice. The more useful experiment is to feed a Wikipedia article plus its sources and ask what in the text does not match the citations. The model does not have to be perfect. If it returns ten suggestions and seven are decent, that is already worth a good editor's time on obscure pages. Fridman notes that a hit piece about him included GPT-generated text; the journalist later apologized for an error the model had made. Training data, journalism, and Wikipedia now loop.

Wales has lived a smaller version of that loop. A web comic calls it citogenesis: an unsourced Wikipedia error is copied by a lazy journalist, then a helpful Wikipedian sources the magazine. *Biography* magazine, from the Biography TV channel, once profiled him as enjoying chess with friends. He liked the sound of that person and is not that person. The magazine said the detail came from Wikipedia; the history showed already-removed vandalism. He went public so it would not be reinserted as a reliable source. Years later he was invited to make the ceremonial first move at the world chess championship and still wonders whether that article was why.

Wikipedia's license is Creative Commons attribution-share-alike: the world may reuse the work, including commercially, but attribution is both a legal and a moral practice. Wales wants models to ground claims the way Wikipedia grounds them. GPT-4, he says, is trained to be amiable and does not really know or care about the truth. The morning of the interview he asked it for cases of white athletes compared to fast cars. It wrote a plausible essay, then three specific quotes with the right names and eras. He googled every one. None existed. He wants later models to say they do not know—or, as Wikipedia-style banners already do for disputed neutrality, to admit "87% bullshit."

A banner is, in a sense, a summary of the talk page. Wales had not thought of feeding the article and the talk page to a model and asking it to propose warnings. Fridman likes the idea for war and Israel-Palestine pages, where some passages are shared facts and others are seven-page fights, including over whether conflict is even the right title. Wales recalls a talk-page debate on whether an Israeli construction is a wall or a security fence, and a careful paragraph that explained why each side chooses its word. The page still needs one title, one lead photo, one set of categories. Those forced choices, not the body text, are where neutrality is hardest.

Asked whether Wikipedia has a left-leaning political bias, Wales says not broadly. Specific entries can be off; that is what the process is for. Extreme Twitter accusations usually collapse when he asks for an example. People with fringe views are often unhappy that Wikipedia will not treat those views as mainstream. A homeopath once accosted him outside TED in Vancouver over an article that calls homeopathy pseudoscience. Wales said he could not help: the page cites good sources, and those sources are unkind to the science. In 500 years a fringe thinker might look like a genius; today the encyclopedia still has to describe present evidence.

The biases that persist, he argues, are often obscure or apolitical. Japanese-anime articles tend to be very favorable because almost nobody except fans writes them. On something as politicized as mask efficacy in COVID, his last reading was closer to mixed evidence: masks seem to help, perhaps less than some authorities claimed. Both emotional camps may hate that. If a reader can say they have a view, understand other views, and see the complexity, society is slightly more mature. False neutrality is the other failure. The moon article should not weigh green cheese against rock. Flat-Earth theory belongs in a history-and-culture article, not as a some-say paragraph on Earth.

Fridman guesses the editor corps leans left, as Silicon Valley does, and asks whether that seeps in through small jabs—the right's Democrat Party versus Democratic Party—or through huge narrative fights such as Hunter Biden's laptop. Wales thinks parallel universes that cannot agree on basic facts feel worse than they used to, and he would lay a lot of recent blame on social-media algorithms that reward clever virality. His own most successful tweet, by a wide margin, was two words. A reporter told Elon Musk he should buy Wikipedia; Wales wrote "not for sale." It was a mic drop, not discourse, and he knows that is what the platforms pay for.

A fifteen-tweet thread weighing mask evidence will not go viral; a smackdown of a masked politician caught unmasked at dinner will. Wales's family parable is the crackpot uncle versus the sweet grandmother. At Christmas everyone rolls their eyes at Uncle Frank; he has three mates at the pub. Online, grandmother has 54 followers and the uncle has 714—not an influencer, just amplified. She posts and nothing happens. He posts, people pile on, and the system records engagement, page views, ads. Society, he says, gets more of what it rewards. He has told people at Facebook that if the product convinces the public they are breaking Western civilization, that is bad for business in the long run.

Mark Zuckerberg, Wales notes, has unusual control of the company, which in theory lets him sacrifice some near-term optimization. One concrete recommendation is a moratorium on all political advertising. Targeted ads broke the old constraint that a campaign needed a story 60 percent of people could live with. During Brexit, he cites ads to animal-rights activists promising Britain could pass real animal-protection laws outside Europe, and ads to fox-hunting advocates promising the hunt could return. Same referendum, opposite messages, each audience hearing what it wanted. Until that problem is solved, he would rather Facebook forgo the revenue.

His own experiment, WT.Social, was about to relaunch around trust ratings rather than session page views, on a membership model closer to HBO than to ad-funded broadcast. Advertising television, he says, pushes producers toward bland shows everyone can tolerate; a subscription lets *The Sopranos* and *Sex and the City* each refuse to be for everybody. He told Musk that Twitter Blue's interesting piece, however poorly rolled out, is revenue from people who might not pay for toxicity. Wikipedia's advantage is structural. There is no box that says what is on your mind. The box says this is an article about the moon: be neutral, cite sources. Ranting about Trump on the Trump talk page gets you told to stop.

Years before Musk, someone accused Wales of pedophilia on Twitter. He reported it, was told it did not violate the terms, emailed Jack Dorsey, and Dorsey said they would sort it out. Wales's point is that he can email Jack. A bullied teenager cannot. Years later the text claimed to be a ten-year-old accusing Wales of rape the week before. Legal and trust-and-safety leadership replied that the tweet did not violate the rules and invoked #MeToo. Wales's view: the reply should have been the police number. The tweets were later removed as spammy behavior. A Wikipedia fundraising experiment put a long-haired, tattooed developer on banners shown to about 80 million people a day; the man caught a wave of creepy abuse off-site. Wales gets the upside of fame. Ordinary people who go viral as a target, he says, is really quite tragic.

People assume Wikipedia's big fights are left versus right. Wales says no: it is "the party of the kind and thoughtful and the party of the jerks." Bring him a political opposite who is thoughtful and kind and they can have a real discussion. His example is abortion: a kind Catholic priest and a kind Planned Parenthood activist writing together. They will never agree on the conclusion. They can agree that Wikipedia will not take a side and will characterize the debate fairly. The ideological person often assumes that if others knew as much as they do, they would convert. That assumption is frequently wrong and still useful, because it makes people explain themselves.

Could a community that dislikes Donald Trump still write him fairly? Wales thinks the article is pretty good, and that self-awareness is the mechanism. He would not edit the Trump entry; he gets emotional and would fail as a Wikipedian. On mask efficacy he can catalog evidence without getting angry. Composition still matters because of blind spots, not because the project should be a battleground. Volunteers are, by a rough estimate, more than 80 percent male, college-educated, heavy on tech geeks, and that pattern is global. Asked why only white men edit Wikipedia, he jokes that the questioner has not met the Japanese Wikipedia community, who are also mostly geeky men, plus women.

People write what they know. As a parent he suspects early-childhood-development articles are weaker than they should be, because twenty-five-year-old childless tech geeks are neither interested nor qualified. A look at prizewinning novelists found men's articles longer and better. He knows hundreds of Wikipedians and does not think they are sexists who dismiss Maya Angelou. Hard science fiction is read mostly by men; other fiction more by women. Without women in the community, important women novelists get thinner coverage because nobody feels like writing them. He wants kind-and-thoughtful Catholic priests, conservatives, libertarians, and Marxists. People who arrive from Twitter with a chip on their shoulder, ready for battle, usually fail.

Community health, for him, is often about newcomers, not politics. One fundamental rule is ignore all rules, written that way partly to startle people. It means do not yell at someone for a malformed footnote; thank them and link the format guide. Newcomers show up with ideas discussed thousands of times. The ideal reply points to the old conversation. Interviewing Wikipedian of the Year Emily Temple-Wood for his book, he learned she had started as a vandal at thirteen. Someone wrote on her talk page: why are you doing this? We are trying to make an encyclopedia. She felt terrible, saw the point, and stayed. The timezone file Unix systems depend on, he adds, is unofficially official because one unpaid person and a small circle maintain it, and once something broke while he was on vacation. Openness plus a charity mission makes people raise a hand for a tiny, clever gift.

Wikipedia is a U.S. charity with no ads. Most money is small donations, often around $25, from millions of people who are still a tiny share of readers. In the early days the no-ads decision was as much aesthetic as strategic: Wales did not want banners on the site, and this was the era of punch-the-monkey ads. He has since thought harder about how a business model rewrites an organization's DNA over decades. Even a founder who believes fiercely in community control and neutrality would watch an ad-funded Wikipedia erode. If you read Nissan and see an ad for a new Nissan, you wonder whether they paid for the placement.

No ads means no incentive for listicle headlines about the ten funniest cat pictures, and no reason for an algorithm to walk a Queen Victoria reader toward Las Vegas hotels because those ads pay more. The first serious fundraiser was Christmas 2003. They had three servers; two crashed. Wales turned the surviving database machine into a front end as well. The site crawled. He hoped for $20,000 in a month and raised nearly $30,000 in two or three weeks. Later they tested messaging. A campaign close to his heart, about free encyclopedias in the languages of sub-Saharan Africa, underperformed. The pitch that works is fairness: you use it all the time, you should probably chip in.

People ask why Google, Facebook, and Microsoft do not just pay for it. Influence, he says, would creep in. Small donors are the best funding; major donors exist and are treated carefully. Fridman says his regretted minutes on Wikipedia are zero, unlike Twitter, and that he follows Depths of Wikipedia on Instagram. Wales gave its operator, Annie Rauwerda, a media-contributor-of-the-year award. On ChatGPT replacing Wikipedia as an interface, he is unbothered in spirit—the gift was always for the world—but he does worry that if Alexa reads the first two sentences without saying Wikipedia, people will stop knowing whom to donate to. The Wikimedia machine-learning lead hacked a weekend demo: ask which Wikipedia articles might answer a question, then answer only from those articles. Wales wants that as a search experience.

Fridman asks whether the co-founder fight is ironic, absurd, or important. Wales says unimportant. People are sometimes surprised to hear him say Larry Sanger does not get enough credit for early work, even though Wales does not think co-founder is the right title. The media love a falling-out story. Sanger has said nobody should trust Wikipedia because it assumes only one legitimate version of the truth on controversial questions. Wales disagrees outright: read a controversial entry and you will see a diligent effort to explain the relevant sides. Figuring out whether a minority view is pseudoscience, an emerging mainstream, or crackpottery is much of the work, case by case, with no magic rule.

Social media, Fridman notes, will screenshot one swing of that grappling and call it a conspiracy. Wales's exhibit is the recession definition. Headlines said Wikipedia, under Biden administration pressure or from lunatic leftism, changed the definition to make Biden look good. Wales's habit is to believe a hit for five seconds, then look. The article had always treated two quarters of negative growth as a traditional loose definition and then the extra nuance economists and agencies actually use. An editor moved that two-quarter line out of the lead. Trolls arrived; the page was semi-protected; people were sent to the talk page. From a Wikipedian's chair it was a routine structural debate.

The Twitter Files article was similar. Someone nominated it for deletion—anyone on Earth can do that—and even the nomination asked whether the material belonged on the Hunter Biden laptop page instead. Within hours an admin snowball-closed it: no chance of passing. Nobody proposed suppressing the information. Musk, Wales says, fell for the story of a cover-up. Fridman finds Wikipedia's Ukraine and Israel-Palestine pages balanced, then asks whether he is the leftist crackpot who cannot see it. Wales's reply is that you should always ask what you might have wrong, a little, not too much.

On governments, Wales is super hardcore. They have never bowed to pressure anywhere, and if a state threatens to block, knock yourself out: you will lose Wikipedia. The strategy works because governments know Wikipedia will not cave after two days to get back into a market. He is sympathetic when a company has staff on the ground who could be arrested. He checked before the interview: the Wikimedia Foundation had received zero requests from the FBI or other agencies to change content, and could not honor them anyway without detonating the community. He has met China's minister of propaganda, not to take orders but to explain how the site works. Public-health agencies, he thinks, lost trust because they did not trust the public: overblown claims to scare people into good behavior work briefly and then destroy default respect for boring government scientists. After high UK officials broke their own rules, his Dunkirk-spirit obedience became private judgment for his family, including sitting in a park with a book.

Fridman reads a two-upvote Reddit comment: a campaign to get the Hitler article to featured status, on the theory that voters preferred socially liberal topics, so Hitler had to be two or three times better. Today's featured political books, the commenter said, looked like anarchism and Karl Marx against 200-plus video games. Wales loves the video-game comparison because it matches the young, geeky, male volunteer base. He once woke to German journalists after German Wikipedia featured the swastika; it was not illegal in that educational context, and part of the point was the symbol's life in other cultures, but he joked to the community: warn him before you put a swastika on the front page. Hitler as today's featured article would need to be extremely good. Disproportionate video games, he thinks, are worth the community noticing.

Asked for a favorite article, he names Inherently Funny Words. Early on it was a dumping ground, then a deletion nominee, until someone argued it is a real concept in humor theory. Editors sourced each remaining word. His lingering disappointment is a picture that had to go for lack of a source: a milk cow with horns tied on, captioned that according to some, cow is an inherently funny word, hilarious to him because according to some sounds so much like Wikipedia. Depths of Wikipedia exists to surface that kind of obscurity. Fridman reads Octopolis and Octlantis, octopus settlements in Jervis Bay; humans assessed as a species of least concern in 2008; and the friendship paradox, documented by Scott Feld in 1991, that your friends have more friends than you do. Growing the Wikimedia Foundation, Wales adds, was harder than growing the site or the volunteer community.

In ten years he expects Wikipedia to still be an encyclopedia, not TikTok, with more AI tools and question-and-answer search over their own corpus. Most people will miss the important trend: Wikipedias in developing-world languages, including Swahili, doing well. Machine translation used to be useless for the community and then good only for economically important pairs such as English-Spanish; English-Zulu stayed terrible. New training methods, as he understands them, no longer need tightly aligned bilingual corpora, so a small-language editor might finally paste English and get Zulu good enough to share. At a hundred years the bet is financial: a cautious Foundation that builds reserves, plus a separate endowment that had just passed $100 million, with its own board, so a future CEO cannot blow it. If anyone is around in a hundred years, he believes Wikipedia will be too.

Large language models feel to him like the first web browser or the iPhone: a step change, with ugly low-hanging fruit such as spam that has read your LinkedIn. He predicts traditional news brands will matter more, because a plausible unknown site may be machine-written garbage. A fake photo of the Pope in an expensive puffer jacket made him admire the craft, not indict the Pope. Fifteen years ago the typical Twitter-photo lie was a mislabeled old battlefield; now a convincing still can be made in five minutes. He thinks his own credibility should equal a *New York Times* reporter's if he witnessed something, though nothing interesting happens around him. A deepfake of Zelenskyy saying something outrageous he simply does not believe. Fridman's long interviews, Wales says, carry a brand that makes this is my tape with Zelenskyy hard to dismiss as a fake.

For young people who want impact, Wales's first rule is not to make a cold calculation of what pays. If you are not interested, you probably will not be good. For anyone starting out he offers two pieces of advice that fight each other: persist through hard stretches, two servers dying on a holiday, and be prepared to pivot, the word he learned later, as he did when Nupedia was not working. The live question is which day you are in. Ask both: am I chickening out because of a bump, and am I being pigheaded about something that does not make sense? Knowing those are the choices is already useful.

MySpace, he used to say, was poised to be Facebook until Rupert Murdoch bought it and ads crushed the experience: three clicks and three ads to accept a friend request, versus Facebook's in-place accept. Then someone pointed out Murdoch paid about $800 million and the network stayed profitable through its decline. He made the money back and more. Failure depends on the scoreboard. Mental models of entrepreneurship built on Bill Gates and Mark Zuckerberg—one idea, straight to the moon, among the richest people alive—are extremely rare. If your standard is sadness at not being as rich as Elon Musk, almost everyone should be sad. When journalists ask how it feels not to be a billionaire, he asks how it feels to them. London is full of unknown bankers who make more than he ever will. He would not trade: the Chinese propaganda minister, a president who wants a meeting, a local Wikipedian offering a spare room.

Meaning, for Wales, has no external answer. It is internal: we decide. Yuri Milner's book arguing that the species' purpose is to survive off-planet does not, in Wales's view, motivate ordinary days; the distances are too large, and even a Musk-scale fortune dies on the ship. It did make him look back at Earth and think about million-year sustainability. He is excited about AI for cancer modeling and possibly climate, while bracketing the risks. Fridman adds that if civilization fails, preserving the richness of human knowledge would be a small consolation. That reminds Wales of Bergen, Norway, and Buekorps, a centuries-old neighborhood drumming contest written up in Norway's languages and then translated outward. Machine translation, he insists, goes both ways. Blocking Wikipedia, he told Chinese officials without success, also blocks a town of 50,000 from putting its festival into the place people actually look.