Richard Craib on WallStreetBets, Numerai and the Future of AI Stock Trading

High-Level Overview

Lex Fridman talks with Richard Craib, founder of Numerai, a hedge fund that gives away obfuscated financial data to thousands of data scientists and trades on the combined predictions of their machine learning models. Recorded in early 2021 at the height of the GameStop saga, the conversation opens with WallStreetBets, short squeezes and Robinhood's trading restrictions, then moves into how Numerai works, why financial data is such a strange machine learning problem, how staking cryptocurrency aligns incentives, and Richard's ambition for Numerai to manage all the money in the world. It closes with thoughts on money, crypto, Dogecoin, startup advice, books and a brush with mortality.

The tone is enthusiastic, curious and occasionally playful, with Lex framing events philosophically (distributed power versus centralized elites, freedom versus chaos) and Richard grounding them in the mechanics of markets and incentive design. The conversation moves from news-of-the-moment finance to technical machine learning and then out to big-picture questions about money, ambition and meaning.

The Discussion Explores:

• How WallStreetBets coordinated anonymously to squeeze GameStop, and why it was a positive-sum trade
• The difference between "evil shorting" and "chill shorting," and why hedge funds still matter
• Robinhood, Vlad Tenev's conversation with Elon Musk, and the value of CEOs being real
• How Numerai works: free obfuscated data, predictions instead of code, and stake-weighted ensembles
• Why financial data is counterintuitive for machine learning and why Numerai calls it the hardest data science problem
• Staking as a solution to anonymity and cheap talk, possibly even for Reddit
• Numerai Signals, the plan to manage the world's money, and freeing talent from redundant hedge funds
• Money, crypto, Dogecoin, startup boldness, book recommendations and Richard's bout with COVID

Deep Dive into Main Ideas

The GameStop Saga as Decentralized Coordination

Richard calls WallStreetBets "my favorite story ever" because he loves decentralized coordination. He explains that anonymous users can't make contracts, but they can build credibility through past upvoted investment ideas, karma and screenshots of their positions, and that was enough for a flock of traders to move in one direction. What made it unusual was that they bought a bad stock precisely because it was bad: GameStop had been shorted more than its own float, so if enough people bought it, hedge funds would be forced to buy back shares at very high prices.

He explains that short sellers face unlimited potential losses because there is no ceiling on how high a stock can go. A normal pump, like a bubble in watermelons, is zero-sum, but squeezing someone who has already shorted is positive-sum for the buyers as a group. Lex, in his intro and throughout, frames the episode as regular people coordinating to take on misbehaving centralized power, while worrying aloud about crowds being manipulated by charismatic leaders or bots. Richard notes the difference from Occupy Wall Street: WallStreetBets was "a little bit more nihilistic, a little bit more yolo," and hedge funds were genuinely scared of it.

Evil Shorting, Chill Shorting and Why Hedge Funds Matter

Richard distinguishes two kinds of shorting. The evil kind, which he attributes to Melvin Capital, means building a huge position, recruiting friends to short too, and using the press to bring a company to its knees. The chill kind is what quant funds like Numerai do: hundreds of small positions, none more than 1% of the fund. Numerai had about 250 shorts, and Richard says it doesn't even want them to fall, only to rise less than its longs; shorting lets the fund stay market-neutral and borrow to go long more heavily, for example shorting oil companies while going long Tesla to "make the future come forward faster."

On Robinhood, Richard wants to see the counterfactual history. He explains that when shorts surge, prime brokers raise collateral requirements. Numerai had a short in BlackBerry that doubled in a day. If commonly held hedge fund shorts had kept doubling, funds would have had to sell their good stocks to cover, causing cascading deleveraging. He defends hedge funds broadly: "I don't think you want to rerun American economic history without hedge funds," because they provide the liquidity that lets venture-backed companies go public. He and Lex also discuss Vlad Tenev's appearance with Elon Musk on Clubhouse, agreeing that Elon's unprofessional directness is refreshing and that CEOs who talk like PR departments, as they suggest Mark Zuckerberg has come to, make things look worse.

How Numerai Works

"So Numerai is the first hedge fund that gives away all of its data," Richard says. The data is bought from vendors whose contracts forbid sharing, and would cost around a million dollars a year to buy yourself, so Numerai obfuscates it: users download what looks like a million rows of numbers between zero and one with unlabeled columns. Underneath are stock-specific factors such as price-to-earnings ratio and one-year momentum across 15 to 20 years of history. The task is cross-sectional global equity ranking with roughly a one-month prediction horizon; the fund holds stocks for three or four months, which Richard considers closer to investing than high-frequency plumbing or arbitrage.

Users never upload their models. "We want you to keep your baby Frankenstein at home and take care of it," he jokes: they submit only predictions, so their code stays their intellectual property, and many automate submissions from a small cloud server. Numerai combines predictions weighted by how much of its cryptocurrency, NMR, each user stakes on their model. The stake-weighted ensemble beats any single model, and even small edges matter: going from 10% to 12% returns in a market-neutral fund pays for all the fees. Getting started means going to numer.ai, downloading the data and running the example Python script, which builds an XGBoost model. Users can participate without staking; if they stake and do badly, their NMR is destroyed rather than paid to Numerai, so the company isn't the house profiting from their losses.

Why Financial Data Is So Hard

Richard says tree models like XGBoost work nicely out of the box, but the counterintuitive lesson of financial data is that you should reduce your model's exposure to the features it likes most. Even if tech was the best sector for a decade, you don't put all your money into tech. He contrasts this with stop-sign detection, where you would never reduce reliance on the color red. Users who load up on one or two features can shoot up the leaderboard by luck, just as a stock-picker can double their money on one stock, but the best users aim to stay near the top for a long time.

Numerai calls it "the hardest data science problem on the planet" because the out-of-sample test is the live stock market, the time series is non-stationary, and billions of dollars are competing to remove any inefficiency. Lex compares Numerai's contribution to ImageNet: the value lies in making a hard problem accessible, with clear standards for judging performance. Richard recalls that the very first winner, told he'd won $300, replied that he'd spent far more on compute, and that some users have submitted every week for four years.

Staking as a Fix for Anonymity

Richard explains that Numerai initially didn't work at all: people made multiple accounts and random models hoping to get lucky. Staking solved this by giving users a negative as well as a positive incentive, eliminating the point of bot accounts and showing which models their creators truly believed in. He says he sleeps easy because users now stake more on their models than he has in his own hedge fund. Lex wonders whether staking could improve internet discourse, since anonymity makes it too easy to be nasty. Richard notes that Numerai has open-sourced its staking protocol, Erasure, and imagines WallStreetBets users backing a promise not to sell with an irreversible smart contract, which he thinks would have let them squeeze hedge funds day after day.

Asked whether AI-run trading is dangerous, Richard argues a model trained on Numerai's data could never even formulate an idea like squeezing GameStop, whereas a slightly more powerful WallStreetBets could be quite dangerous. He recounts a famous hedge fund manager telling him there are profitable trades his firm chooses not to make because they would mess up the market, and doubts a Reddit army would show that restraint.

Numerai Signals and Managing the World's Money

Numerai Signals lets users submit rankings of about 5,000 stocks built from any data they can find, with a roughly seven-day horizon, and rewards only signals uncorrelated with Numerai's core signal. One user, whom Richard believes is in India, built a sentiment signal scraped from WallStreetBets and open-sourced the notebook. Numerai's stated mission is to manage all the money in the world, which Richard says requires all the data and all the talent; he plans to grow the data Numerai gives out tenfold every year for at least a decade.

His deeper argument is about waste: every hedge fund buys the same data and hires the same PhDs, and a huge share of American PhDs end up at hedge funds "and we're not even on Mars." He describes Numerai as "freeing up people who work at hedge funds to go work for Elon." One of the top users works at NASA's Jet Propulsion Laboratory on a Europa mission and runs his Numerai model automatically in his spare time.

Gambling, Money and Crypto

On whether trading is gambling, Richard says gambling math applies: bankroll management matters as much as skill, and you want many independent bets rather than correlated ones. Prices should reflect companies' underlying value. He says he has no respect for Long-Term Capital Management's roughly 30-times leverage, which anyone at a gas station would call a bad idea, and notes Numerai uses no derivatives beyond shorting. Asked what money is, he says finance lets you bring a new future into being, as a student loan lets someone get a degree they couldn't otherwise afford.

Richard initially dismissed Bitcoin because it had no future cash flows, but was won over by the Ethereum white paper and joined its crowdsale. He's no longer investing in crypto, likes Zcash's privacy technology, and notes that anyone can watch his NMR moving on the blockchain. On Dogecoin, he and Lex agree jokes spread faster than anything else, sparked by Elon Musk's tweets, and Lex worries about a future where nothing is taken seriously while defending sincerity.

Startups, Books and Mortality

Richard's startup advice is to be honest about how big your ambition is rather than toning it down to seem low-risk, because a mediocre company is the real risk: "play for keeps play to win go big." Lex agrees, lamenting cautious autonomous vehicle companies focused on fundraising. Richard recalls pitching First Round Capital while Renaissance Technologies co-founder Howard Morgan was in the room, and playing in Jim Simons' charity poker tournament, where he knocked out previous champion Peter Muller and heard Renaissance CEO Peter Brown doubt AI would matter much in finance on the same day a poker AI was beating humans.

He recommends Zero to One, which he was reading while starting Numerai, and names David Deutsch's The Beginning of Infinity as maybe his favorite book for its rigorous optimism that anything not forbidden by physics can be solved. His advice for young people: don't start a company "unless you prepare to make it your life's work," and don't underrate a job that pays you to learn; he thinks two more years as a quant would have accelerated Numerai. On meaning, he describes a recent hard bout of COVID in South Africa as a reminder of mortality that was ultimately liberating. Lex 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."

Conclusion

This episode uses the GameStop frenzy as a doorway into a bigger idea: that well-designed incentives can let large numbers of anonymous people coordinate productively. WallStreetBets did it chaotically through karma and screenshots; Numerai does it deliberately through free obfuscated data, private models and staked cryptocurrency. Richard Craib makes the case that hedge funds are useful, that the kind of shorting matters, that financial machine learning rewards humility about your favorite features, and that a single crowdsourced AI fund could free talent for more important work. Along the way he offers clear-eyed advice on ambition, bankroll-style risk management and learning while you earn.

This is a summary of "#159 – Richard Craib: WallStreetBets, Numerai, and the Future of Stock Trading" from Lex Fridman Podcast, giving you the GIST so you can decide if you want the REST. Listen to the full episode.