Craib, Richard
▶ Watch on YouTubeCore ideas
Craib frames the GameStop short squeeze as a rare "positive-sum" instance of decentralized coordination: rather than everyone simply piling into a stock they liked (which would be zero-sum, like a bubble), Wall Street Bets specifically targeted a heavily-shorted, low-quality stock, understanding that forcing hedge funds into a short squeeze creates genuine (if unevenly distributed) gains. He is fascinated by how the community solved a coordination problem despite anonymity — reputational proxies like Reddit karma, screenshots of trades, and post history substituting for the "credible contracts" that would normally be needed to trust a stranger's investment thesis — while also finding the anonymous, "nihilistic," YOLO character of Wall Street Bets more unsettling than the sincere idealism of a movement like Occupy Wall Street, precisely because its lack of stated ideology and its capacity for real, coordinated financial damage makes it harder to predict or reason with.
He distinguishes "evil shorting" (large, coordinated short positions paired with deliberate press campaigns to "bring a company to the ground," which he attributes to funds like Melvin Capital in the GameStop case) from what his own fund does: Numerai holds roughly 250 small short positions (never more than 1% of the fund each) purely to hedge market exposure and free up capital to go longer on stocks it believes in (citing Tesla as an example), arguing hedge funds overall are essential to a functioning economy — providing liquidity and enabling the IPO "liquidity events" venture capital depends on — even if a specific subset behaves destructively. He speculates the GameStop squeeze could have inflicted far more damage on aggregate hedge fund short positions had it continued longer, based on his own fund's experience of a prime broker suddenly demanding full (rather than partial) collateral on a shorted meme stock (Blackberry) after it doubled.
Numerai's own model is Craib's central subject: it is a hedge fund that gives away its underlying data for free, but in obfuscated, anonymized form (millions of numerical rows with no labeled meaning), so that participants worldwide — largely machine learning practitioners rather than traditional finance professionals — can build predictive models without knowing what the features represent. Participants never upload their model code (protecting their IP); they submit only predictions, which Numerai combines into an ensemble weighted by how much of Numerai's native cryptocurrency (NMR) each participant has "staked" behind their model — a mechanism that aligns incentives (bad models lose their stake; the fund only profits when good, staked models are rewarded, not when users lose) and that Craib credits as the direct solution to an early problem of users submitting random, unstaked models hoping to get lucky. He notes a related "Numerai Signals" product lets anyone convert arbitrary external data (his example: a natural-language sentiment score scraped from Wall Street Bets and converted into a stock ranking) into a data feed the fund can use, since only signals uncorrelated with Numerai's core data are valuable.
Craib argues Numerai's stock-prediction problem is uniquely difficult because it is non-stationary — patterns that worked in the past may not recur, and there is no guarantee that out-of-sample data resembles training data — and because it counterintuitively rewards models for reducing exposure to their single most predictive feature (to avoid overfitting to features that might stop working), unlike typical machine learning tasks such as computer vision. He credits tree-based models (like XGBoost) as reliable strong baselines, and frames the long-term vision as radical: Numerai's stated mission is to "manage all of the world's money," reasoning that consolidating hedge fund talent and data into one open, crowdsourced system would free redundant PhDs currently duplicating effort across competing funds (e.g., a NASA Jet Propulsion Lab scientist who contributes a winning Numerai model as a side project rather than quitting to work in finance).
Drawing an explicit parallel between Wall Street Bets and Numerai as two decentralized, "power of the people" systems, Craib highlights their key structural difference: Numerai solved the credibility/coordination problem formally through blockchain-based staking (open-sourced as a protocol called Erasure), while Wall Street Bets relies on informal reputation. He speculates that if staking mechanisms were applied to platforms like Reddit — forcing users to put "skin in the game" behind public claims — it could reduce bad-faith behavior online, though he considers it likely more dangerous for a Wall Street Bets-style human crowd to become more powerful than for an AI-driven system like Numerai to scale, since Numerai's models cannot even conceive of a scheme like "short squeeze GameStop" the way coordinated humans can. Craib also reflects on Vlad Tenev's Clubhouse conversation with Elon Musk as a case study in authentic, unscripted communication (contrasting Tenev's evasive, over-cautious tone with Musk's directness), broader reflections on money and finance as mechanisms for "bringing the future forward" (enabling loans, risk-taking, and ambitious projects that wouldn't otherwise exist), his early skepticism-turned-interest in cryptocurrency (drawn to Ethereum's smart-contract flexibility over Bitcoin, and to Zcash's privacy properties), and startup advice emphasizing genuine, undiluted ambition ("don't start a company unless you're prepared to make it your life's work") over hedged, low-risk-sounding pitches. He closes with a reflection on a recent, serious COVID-19 infection that left him isolated and confronting mortality directly, which he describes as ultimately energizing rather than purely frightening.
Excerpts
On why the GameStop trade was clever rather than a simple bubble:
“It's not zero-sum — if someone's already shorted the stock and you can make them short squeeze, it's actually a positive-sum game. So yes, some redditors will make a lot of money, some will lose a lot, but actually the whole group will make money, and that's really why it was such a clever thing for them to do.”
On how staking solved Numerai's coordination problem:
“Staking was our solution — could we make it so that we could trust, we could know which model people believed in the most, and we could weight models that had high stake more, and effectively coordinate this group of people... there's no incentive to creating bot accounts anymore either. I stake my accounts, in which case I should believe in them, because I could lose my stake.”
On the counterintuitive need to reduce exposure to a model's best feature:
“What a lot of people figure out is actually if you train a model on this kind of data, you want to somehow neutralize or minimize your exposure to certain features, which is unusual — because if you trained a stop-light detection model in computer vision, that's the last thing you want to do. Why would you reduce your exposure to the thing that's helping your model the most? That's this counter-intuitive thing you have to do with machine learning on financial data.”