A 21-Year-Old at a Party Called GameStop. He Was Right.

Richard Craib, founder of the crowdsourced hedge fund Numerai, tells Lex Fridman that the GameStop saga is his favorite story. What he liked was not the memes so much as the machinery underneath: anonymous WallStreetBets users hyping one another, posting proof they had bought and were holding, and then watching that decentralized coordination move real money. Fridman is struck by the same thing from the other side — a large number of ordinary people coordinating against centralized power, especially when that power seemed to be misbehaving.

Craib learned about GameStop in the most ordinary way. A few days before this conversation, at a party, a 21-year-old told him the stock would go up 100% in a day. Craib found it strange that someone barely old enough to drink was teaching him how to trade in a social setting. The next day he checked. The kid was right, and made about 140% in a single session.

That kid, Craib says, is now a little wealthier and already hunting the next setup, and there are thousands like him. The decentralized entity is going to get bigger. People who sit with concentrated power in government and finance, he thinks, are probably a little scared, and that is dangerous and also a little good. It certainly makes a manager think twice about stuffing a huge short into one or two names.

Coordinating strangers is hard because coordination needs credible claims, and Reddit usernames come with no contracts, no identity, and no legal leash. Craib still sees small things that help. If you posted a good idea in the past, that post still sits on Reddit with its upvotes, and it lends credibility to the next pitch. Users also post screenshots of their trades. Screenshots can be faked, but a lot of karma plus a visible track record is enough for others to decide that this person probably did put a million dollars in, and that the trade is easy to follow.

You integrate that information, decide something is happening, and jump onto a little boat. Another person jumps on, then another, and suddenly a huge number of people are moving in the same direction, "like a flock of birds." What was strange this time, Craib says, is that it was not "let's all buy Tesla because we love Tesla." They were buying a stock because it was bad. That is a little too galaxy-brain for a decentralized community, and he is still impressed they found it.

The reason they liked GameStop was short interest — Craib believes it had been shorted more than its own float. If enough people bought the bad stock, they could squeeze hedge funds into capitulating and buying at very high prices. Shorting is a bet the price will fall; a short squeeze is what happens when the price rises so far that the short cannot hold. There is a floor at zero and no ceiling, so a double overnight puts brutal stress on the manager, whose only exit is to buy the stock they still think is going down.

If a crowd pools money to blow a bubble in watermelons, the price goes up and comes down and the game is zero-sum. If someone is already short the stock and you can force a squeeze, Craib argues, it is actually positive-sum. Some Redditors will make a lot, some will lose a lot, but the group as a whole takes money from the shorts. That, he says, is why the trade was clever. The Reddit side was not doing anything exotic — they were mostly just buying the stock, maybe some options. The exotic book was on the hedge-fund side.

Fridman admits that shorting has always felt, from outside, like the market equivalent of haters: not only a bet against a company but leverage and press used to knock it down. Craib will not go that far. Economically, shorting is how markets punish overvalued assets, and markets need that. What he cannot stand is the other half: putting down a huge position, getting buddies to join, making press, and trying to bring a company to its knees. Melvin Capital, in his telling, was doing "evil shorting." They were not even calling GameStop a fraud. They were calling it a bad company and trying to grind it into the ground.

"Chill shorting" looks like Numerai. The fund holds about 250 shorts, none more than 1% of the book. Craib does not know the names except for one that happened to be a meme stock. They do not even want those names to fall necessarily; they want them not to rise as fast as the longs. Shorting a little lets them borrow from banks and go longer in what they believe, because a long-short book has no net market exposure. When they were long Tesla they could do it with more money than they had. "We can short the oil companies and go long Tesla and make the future come forward faster," he says, "and I do think that's not a bad thing."

Robinhood, the brokerage that made the trade easy for retail, then centrally restricted buying in the names at the center of the squeeze. Craib is less interested in the gossip than in the counterfactual. He wants the alternate history in which the buy button stayed on. How much more damage would the momentum have done? Numerai got a taste anyway. They were not short GameStop; they were short BlackBerry, a "super bad company" the models always wanted lower, and it went up about 100% in a day. Meme stocks, in his definition, were names with memetic momentum on Reddit and, more importantly, high short interest: GameStop, AMC, BlackBerry, Nokia. Nostalgia for movie theaters was not the point.

You would not have needed many more days of that chaos to hurt hedge funds, and he thinks the risk was underrated. When shorts rally, collateral requirements jump. Numerai's prime broker had been asking for something like $40 per $100 of short exposure; the next day it wanted 100%. If that hits every commonly held hedge-fund short at once, you cover by buying the bad companies and you raise cash by selling the good ones. The names funds love fall, the names they hate rip, and the cascade feeds itself. A few more days of GameStop and AMC doubling would have meant more and more forced deleveraging.

That does not mean Craib wants hedge funds gone. WallStreetBets itself is a kind of distributed hedge fund; Numerai is a hedge fund; the category is broad. What he likes to tell people who hate the industry is that you do not want to rerun American economic history without it. Hedge funds make liquidity in stocks. If you love venture capitalists funding new technology, you also have to like the public market that lets those companies have a liquidity event. The country could maybe get by with an AI hedge fund and without the "evil billions" shops that work the media to kill companies. It still needs the function.

Vlad Tenev, Robinhood's CEO, went on Clubhouse with Elon Musk while the buy-button fight was still raw. Craib was excited to hear it, then froze in the first seconds. Tenev sounded like a 55-year-old talking to a 20-year-old. Musk was the 20-year-old. You could see, Craib says, why Citadel would get along with him: a respectable, professional attitude. Tenev even tried a jokey line, the way a 60-year-old Bank of America CEO would try to make a joke for the kids. Craib kept thinking, this guy is like 30. What is this?

What he likes about Musk's influence on American business is how anti-professional it is. Why say a hundred words about nothing? Musk cut in and told Tenev to spill the beans. Fridman argues that being real — going off the cuff, making mistakes, paying for them — is worth the cost, and that Tenev had a chance to be that and was not. Craib would rather a CEO make mistakes and be honest than be professionally never-wrong. He says Mark Zuckerberg had that quality when he was younger, until someone told him he could not say "lol" in an interview, and he became a distant person.

The hard hire, they agree, is the PR person who understands you can say "lol" on the record. Most communications staff exist to minimize the risk of the wrong sentence. That nervous energy compounds as teams grow. Fridman describes visiting YouTube: astonishing engineering, an astonishing system, and everyone scared, repeating that they just want to improve the health of the platform. Nobody would say they are sitting on a chaotic mess of billions of videos, hate, conspiracy theories some of which might be true, and they do not know how to handle it. The missing candor, he thinks, is what makes people smell a back room.

Numerai, Craib says, is the first hedge fund that gives away all of its data. That is roughly the last thing a hedge fund would do, because data is the edge. They do it by obfuscating. What you download for free looks like a million rows of numbers between zero and one; you have no idea what the columns mean. If you have done machine learning or regression, you still know you can find patterns without knowing the feature names. The underlying feeds come from vendors who would never allow a resale, under contracts that cost on the order of a million dollars a year. Fifteen or twenty years of history, stock-specific factors — price-to-earnings, one-year momentum, and many less common ones — get turned into a problem quants call cross-sectional global equity. You are not saying this name will go up. You are saying its relative place in feature space makes it not a bad holding in a portfolio.

The horizon is not high frequency. They ignore order books. Predictions look out about a month; they end up holding three or four months, which Craib thinks is closer to investing than to "plumbing" arb across exchanges. They could train an XGBoost model in-house and it would be quite good. Opening the data does better, and in stocks a little better is enough. The gap between 10% and 12% in an equity-market-neutral fund is enormous, because you are often charging 2% fees; two extra points is the whole fee. New libraries and papers get synthesized by the crowd. Weighting is not a beauty contest among one or two winners. Users stake cryptocurrency on the models they believe in, and Numerai averages the stake-weighted predictions. The ensemble is better than any one model, which is also how investing works.

You never upload the model. You keep the code — Craib calls it your baby Frankenstein — and send only predictions, so the whole fund is built from models Numerai has never seen. Some users run Numerai Compute, a small cloud image that answers a ping with a fresh vector. You do not have to stake. Many people try for months with nothing down, watching whether the model works on live data or is overfit. NMR, the Numeraire token, can be bought on Coinbase, Uniswap, Binance; you cannot buy it from Numerai, and if you do badly the stake is destroyed rather than paid to the house. Interests stay aligned. Getting started is unromantic: go to numer.ai, sign up, download the pack, run the example Python script, talk to the community, and expect to spend on compute. The first person who won got a personal email celebrating $300. They wrote back that they had spent way more on machines.

Asked what works, Craib will not crown transformers or bury neural nets. Tree models, XGBoost and its cousins, work nicely on this data out of the box; if you want to come 100th in the tournament, start there. The surprise, for people who know ordinary machine learning, is what you do next. In finance you do not want to be too exposed to any one risk. Tech may have been the best sector of the last decade; that does not mean you put all the money in tech. A trained model will tell you to. On Numerai, a lot of people figure out that you want to neutralize or minimize exposure to certain features — including the ones helping the model most.

That is backwards from computer vision. If an autoencoder decided that red and white were the stop-light features, reducing exposure to red would be insane. Financial data is full of patterns that appeared in the past and also full of patterns that have not appeared yet, so you cut the exposure that would overfit the last regime. Climbing the leaderboard is part art. Super-expose yourself to one or two features and you can rocket upward if you get lucky, with a lot of volatility, the way putting everything in one stock can make 100% and then not work for long. The best users try to stay high as long as possible rather than be first for a little bit.

Fridman compares it to Kaggle, then Craib draws the distinction that matters. Kaggle gives you a toy problem and an out-of-sample test with a known answer. On Numerai the out-of-sample is the live stock market; they do not know the answer either. Users have submitted every week for four years. Craib calls it the hardest data-science problem on the planet, and he knows it sounds like marketing. The claim, he says, stands up because billions of dollars are fighting over the same inefficiency, and nobody is leaving it inefficient on purpose. The deepest reason is that it is non-stationary time series. They regularize features and targets so patterns are findable, but there is no guarantee the next era's distribution will look like training, and even the training eras have the same dislocations. ImageNet made images into a problem by making them accessible and judged; Numerai is trying to do that for markets.

Craib keeps coming back to WallStreetBets because Numerai is trying to do a cleaner version of the same trick: coordinate anonymous people. At launch it did not work at all. People made multiple accounts, threw up random models, hoped to get lucky, and some of them did. Staking was the fix. If you put money behind a model, Numerai can weight the ones people actually believe in, and the incentive to farm bot accounts dies. Either you stake, in which case you should believe because you can lose the stake, or you do not, and the weight stays low. Positive and negative incentives together make the swarm usable. The stake sits on-chain, so in some sense users are trusting the blockchain more than they are trusting Numerai.

"Nowadays I sleep easy because I have less money in my own hedge fund than our users are staking on their models." Fridman immediately wants the mechanism on Reddit. Anonymity is freedom, and it is also too easy to be lousy in the comments. Numerai open-sourced the staking code as a protocol called Erasure. If Reddit wanted to, it could let people stake on claims. Imagine a WallStreetBets user saying they will not sell GameStop, attaching a smart contract that no one in the world, including them, can undo, with a million dollars locked against the sentence. It would not have to be millions. A huge number of small stakes would have been enough, Craib thinks, that the squeeze would have continued day after day, with every hedge fund collapsing. That is the power, and the danger.

Looking forward, could Numerai-style infrastructure — AI systems backed by humans — run the whole stock market? Some of the humans, or the systems, could be bad actors. Craib once met a famous manager who said they can see trades that would make money and mess things up, and they choose not to do them. The Reddit army, on the evidence, would not have that pause; the little voice of morality gets trampled when crowds start chanting. Numerai is oddly safer on that axis. You cannot model the obfuscated data and have the model invent "let's short-squeeze GameStop." The problem is not framed that way. Numerai can get very powerful without being dangerous in that particular way. WallStreetBets needs only a little more power, he says, to become pretty dangerous.

Numerai Signals is the other door. You still send predictions, but from your own data, not theirs. That is how the fund imports feeds it does not have, and you are rewarded only for the part orthogonal to the core signal — uncorrelated, which is why strange alternative data is valuable. Price-to-earnings is not going to look like a WallStreetBets thread. The week before this conversation, a user in India built a signal scraped from the subreddit. It is now one of thousands in the book. The structure is the same: a ranking over about 5,000 stocks, one for like, zero for dislike. They open-sourced a Colab notebook that turns natural language into a sentiment score and then a rank — this stock sucks, this stock is awesome. Signals look out about seven days; the core tournament is about thirty.

The mission statement is to manage all the world's money, which Craib treats as a first-principles claim: with good enough modeling and good enough data you should not lose, and the question is how much of each you need. Numerai is 10xing the data it gives out this year, and he thinks it can 10x every year for at least the next ten. Signals are the weirdness arriving through the back door while the core set covers the bulk. The industry, as currently built, is a waste of talent. One fund buys a data source and hires a PhD; the next fund buys the same source and hires another PhD; a third of American PhDs end up in hedge funds, and we are not even on Mars. Numerai, in that light, is about freeing people to go work for Elon.

One of the top users works at NASA's Jet Propulsion Laboratory. Craib visited. The office had Numerai posters and looked, to him, like an Apollo movie. The point is that the engineer did not quit. His mission is getting a spacecraft to Jupiter's moon — the Europa Clipper. Craib wants that intelligence in the stock market, because he thinks markets and hedge funds are good, and he does not want the man to resign. So the model runs on weekends and submits automatically, and he is one of the best users they have.

Is stock trading closer to gambling or to investing? Craib will not pretend there is no overlap. All trades are speculative; they have a horizon. Bankroll management from gambling applies, and small changes to how you size can beat improvements in skill. Numerai wants a lot of independent bets, because that is a higher Sharpe than a pile of correlated names in one sector. The point of the whole apparatus, he says, is that prices should reflect underlying value. A world in which a fund can look at the data and see that everything is super mispriced is super bad for society. Numerai does not use derivatives unless you count shorting as one. They put money into companies in small ways. They really did buy Tesla. The role in its success was tiny. He still thinks that matters.

Asked what money is, Craib does not have a clean answer. Crypto has warped his personal sense of it. You stop caring about the price and start caring about the product — the mechanisms, the little contracts and agreements between people that are bigger than cash changing hands at a 7-Eleven. What he can define is finance. When you can take out a loan, you can bring a whole new future into being. Without a student loan you might not get the degree; with the loan, a different world arrives. Rerun American economic history without loans, without derivatives, without money, and the system does not really work. Finance is the magic of bringing the future forward, and it is what lets bold builders take risks other people cannot.

Friends told him about Bitcoin around 2014. He was an equities person and bounced off it: no net present value, no cash flows, no dividends. Ethereum's white paper was the thing that felt visionary — constructions, especially smart contracts, that Bitcoin did not have. He is not looking to invest in cryptocurrencies anymore, and many of them are almost identical. The ones he still likes are the privacy coins. Zcash is a different kind of invention. On public chains, none of the transactions are private. He holds some Numeraire, and Etherscan will literally label the address "Numerai founder." They can see him move it and wonder why. Private transactions, he notes, also let you play different games, and with Zcash nobody will know which games.

Fridman asks, half-joking, why Dogecoin is going to win. Craib's serious answer is memetics. It is not only Musk tweeting. The tweet is a spark; the fuel is everybody on the internet spreading the joke, a WallStreetBets-style fire. Funny is extremely good at captivating discourse. Craib had not thought about the original Doge meme in a long time, then he thinks of it, remembers that it was really funny, and there is a small probability he buys some. Scale that across millions of people who have been told the same joke, and reminiscence becomes a persistent bid.

Fridman spins the thought into a dystopia in which centuries from now all human communication is humor, everyone high and laughing. He likes how it feels not to take things seriously for a minute, and he dislikes sentences that end in "lol," because if everything is a joke that way of thinking can destroy the world. Will memes save the world or destroy it? Both, they agree, are possible. Craib has always felt this about America: people tell jokes all the time and they are good at it, and if you pull someone aside for a sincere conversation it is hard to keep a straight face. Fridman will not stop posting the earnest lines — he is in love with the world — even when they draw mockery, because clichés sometimes capture basic truths. Musk's skill is the dance: joke, then snap back to a real problem, then joke again.

Craib's startup advice is to tell people what you really want to happen in the world. If you are raising, do not say you might one day make a million dollars when you believe something larger. You tone down optimism to look low-risk, and the actual risk is that the company becomes mediocre, because nobody wants to work at a mediocre company or invest in mediocrity. Play the real game. Ten people will tell you to care about a boring detail that will not matter in five years; push back. In fintech he keeps seeing lame ideas dressed in fake enthusiasm while bigger ideas sit untouched. Fridman uses autonomous vehicles as the example: the same stack can be a cautious fundraising story built on fear and decade-old tech, or a bet that rewrites cities, manufacturing, and how humans sit next to machines. The idea can be big and still die if you will not let yourself think it that big.

The investor process was a mystery to Fridman. Craib arrived in America, nearly ran out of money, met people quickly, and then sent email updates every three minutes so he would not let them down. They did not care. They can help with business development; they add little on product, especially when the product is years of market research. The useful meeting was Howard Morgan in the room — a 1980s co-founder of Renaissance Technologies with Jim Simons, later a partner at First Round Capital — while Craib was explaining that quants use mathematical models. Morgan said he had started Renaissance, he knew a bit. First Round did not want to invest in a hedge fund. Craib wrote a blog post arguing they were missing it: this is a tech company. They came in. Later, at Simons's charity poker night, a $10,000 buy-in Texas Hold'em game that the show *Billions* loosely depicts, he sat among Peter Brown, then Renaissance's CEO, Peter Muller of PDT, Robert Mercer, and a table of World Series bracelet winners. He knocked Muller out and got a small trophy; Muller had won the thing three times. Simons outlasted Craib. Rebuys were unlimited because it was charity, so the table went all-in immediately.

He felt out of place. The titans do not tell you how to build a hedge fund, the way a rocket engineer might talk about rockets, and they teased him as the Bitcoin kid. Peter Brown said he did not think AI would have a big role in finance — too noisy, unlike self-driving cars. Craib thought that a strange thing to hear from Renaissance's CEO, and funnier still because that same day, he recalls, the poker AI Libratus was starting to beat humans. Fridman hears those remarks as a prod: prove me wrong, and also, this will take a long time. Craib is building Numerai so that if a narrow intelligence ever wins all the money in the market, long before AGI, it is using their data.

He was reading Peter Thiel's *Zero to One* while inventing Numerai, and his favorite book is David Deutsch's *The Beginning of Infinity*, optimistic with the same rigor cynics use: anything that does not violate physics can be solved. Advice for a 20-year-old: do not start a company unless you are prepared to make it your life's work, and do not plan to sell it in a semester. He spent two and a half years as a quant; another two would have taught him more and accelerated Numerai. An air-conditioned room, free food, and a salary while you learn is an insane deal, as long as you do not buy the house, the car, and the story that you cannot leave. On his desk he kept papers on quant finance, optimization, and the Ethereum white paper — the intersection of crypto, quant, and machine learning that not many people had. Asked the meaning of life, he does not think it is the right question, only a useful palette cleanser, like looking up at the stars. The month before the taping he had a bad case of COVID plus a lung infection, about fifteen days in isolation, the usual tricks of coffee and willpower useless. Getting out was liberating. It shocked him into a new energy. You carry that little flag, he says, and go try to create something special. Fridman closes with Warren Buffett: games are won by players who focus on the playing field, not by those whose eyes are glued to the scoreboard.