Lex Fridman PodcastRichard Craib: WallStreetBets, Numerai, and the Future of Stock Trading | Lex Fridman Podcast #159
CHAPTERS
- 0:00 – 2:30
Numerai’s core idea: a crowdsourced hedge fund powered by ML + aligned incentives
Lex sets up the conversation by framing Numerai as a hedge fund that crowdsources machine learning models from a global community. The key differentiator is incentive alignment: participants keep their model code private while contributing predictions, with performance tied to staking.
- •Numerai as a crowdsourced hedge fund where AI models (not humans) drive trading
- •Obfuscated/anonymized data shared freely to attract diverse modelers
- •Incentives aligned: contributors retain IP; only predictions are submitted
- •Lex’s broader theme: distributed systems, freedom, and emergent coordination
- 2:30 – 6:25
WallStreetBets and the GameStop saga: how anonymous crowds coordinated
Richard and Lex unpack why the GameStop episode was such an extraordinary feat of decentralized coordination. They discuss credibility mechanisms on Reddit (karma, history, screenshots) and how a swarm of individuals can produce large-scale market impact.
- •Why coordination among anonymous users is surprising and non-obvious
- •Credibility signals: karma, past calls, and (possibly fakeable) screenshots
- •Emergent behavior: individuals join a collective trade as momentum builds
- •The role of internet-native culture in enabling rapid coordination
- 6:25 – 16:41
Short squeezes explained: why GameStop was a ‘clever’ positive-sum attack
They break down the mechanics of shorting and short squeezes, and why targeting heavily shorted stocks made this event different from typical hype-driven buying. Richard argues the squeeze dynamic can create a positive-sum opportunity for the crowd relative to the already-short position holders.
- •Definitions: shorting, infinite loss potential, and forced covering
- •Why GameStop: extreme short interest made it structurally vulnerable
- •How forced buying by shorts amplifies upward price moves
- •Richard’s framing: squeezing existing shorts can be ‘positive sum’ for the group
- 16:41 – 18:47
‘Evil shorting’ vs ‘chill shorting’: ethics, risk, and market function
Lex challenges shorting as feeling morally dubious, and Richard offers a distinction between predatory short campaigns and diversified, risk-managed short exposure. They connect shorts/longs to leverage, portfolio construction, and even accelerating technological change (e.g., long Tesla, short oil).
- •Lex’s moral intuition: shorting can feel like rooting for failure/manipulation
- •Richard’s taxonomy: coordinated/predatory vs diversified/risk-managed shorting
- •Numerai’s approach: many small positions; shorts used to balance market exposure
- •Argument that long/short portfolios can fund innovation and shift capital allocation
- 18:47 – 22:45
Robinhood trading restrictions, collateral cascades, and systemic risk
The conversation turns to Robinhood limiting trading and what might have happened otherwise. Richard describes how rising meme-stock shorts increase collateral requirements, forcing hedge funds to deleverage—potentially causing broad market distortions (selling good longs to cover bad shorts).
- •Counterfactual: how much worse could squeezes have gotten without restrictions?
- •Collateral requirements can spike quickly, affecting brokers and funds
- •Cascading deleveraging: cover shorts + sell longs to raise collateral
- •Meme stocks (AMC, BB, Nokia): not nostalgia—targeting high short interest
- 22:45 – 24:20
Do hedge funds matter? Liquidity, ‘evil billions,’ and the case for AI hedge funds
Lex asks whether hedge funds do good for the world or should be ‘destroyed.’ Richard argues markets and venture outcomes rely on liquidity and that removing hedge funds from history would be harmful, while still criticizing media-driven funds that try to ‘kill’ companies.
- •Hedge funds as liquidity providers enabling public market exits
- •Venture capital ecosystem depends on liquid public markets
- •Distinguishing necessary market functions from predatory behavior
- •Richard’s provocative idea: maybe society could ‘get away with’ mostly AI hedge funds
- 24:20 – 31:13
Vlad (Robinhood CEO), Elon’s interview style, and the value of being ‘real’
They reflect on Vlad Tenev’s appearance with Elon Musk and how professionalism can read as evasiveness during crises. The discussion broadens into CEO PR, risk minimization, and how authenticity (and mistakes) can build trust compared to polished non-answers.
- •Richard’s impression: Vlad sounded overly ‘professional’/corporate for the moment
- •Elon’s first-principles pressure: “spill the beans” vs PR language
- •PR teams often optimize for minimizing risk, reducing authenticity
- •Parallel to Zuckerberg/large platforms: realness vs guarded messaging
- 31:13 – 33:24
What Numerai is: giving away obfuscated hedge-fund-grade data to the world
Richard explains Numerai’s central mechanism: distribute high-quality but obfuscated data so anyone can model it without knowing feature meanings. This bypasses the huge data-budget advantage traditional funds have, opening participation to global ML talent.
- •Numerai shares data (unusual for hedge funds) by obfuscating features
- •Underlying data comes from expensive vendors; sharing raw is contractually impossible
- •Participants can still find predictive patterns without semantic feature labels
- •Focus: stock-specific factors over long histories (15–20 years)
- 33:24 – 43:00
Modeling approach: cross-sectional equity, one-month horizon, and practical ML winners
They clarify the prediction task: ranking stocks in a portfolio context over roughly a month horizon (held for months), not high-frequency trading. Richard notes tree-based models like XGBoost work well, and highlights a finance-specific twist: reducing exposure to dominant features to manage risk and stability.
- •Cross-sectional global equity: relative ranking more than single-stock narratives
- •Time horizon: ~1 month prediction; holdings ~3–4 months
- •Not HFT: no order book/tick-level focus
- •XGBoost/tree models as strong baselines; risk/exposure neutralization is key
- 43:00 – 50:00
How to participate: submit predictions (not code), automate via Numerai Compute, stake with NMR
Lex asks for a practical onboarding path, and Richard outlines the workflow: sign up, download data, build a baseline model, and submit predictions while keeping code private. Staking with Numeraire (NMR) signals confidence, improves weighting, and punishes poor performance by burning stake rather than paying Numerai.
- •Get started: numerai.ai, download dataset, run example Python baseline
- •Submit only predictions; models remain private (contributors keep IP)
- •Automation: Numerai Compute-like setup for periodic prediction submissions
- •Staking with NMR: optional at first; later used as ‘skin in the game’ and weighting
- 50:00 – 53:40
Numerai vs Kaggle: the stock market as the true out-of-sample test
They compare Numerai’s competition to typical ML contests where the test set is known and static. Numerai’s out-of-sample evaluation is the live market—non-stationary, adversarial, and relentlessly efficient—making it a uniquely difficult and long-running learning problem.
- •Kaggle: toy problems + fixed held-out test; Numerai: live market feedback
- •Non-stationarity: distributions shift across eras; no guarantees of generalization
- •The ‘hardest data science problem’ claim tied to real-money efficiency pressure
- •Long-term iteration: users submit weekly for years to survive regime shifts
- 53:40 – 1:01:08
Incentive engineering: staking as a solution to trust, bots, and coordination
Richard explains early failures: multi-accounting and random models chasing luck. Staking created credible commitment and enabled robust aggregation, and he notes Numerai open-sourced parts of this mechanism (Erasure), imagining how staking could reshape online discourse and even WSB-style coordination.
- •Early Numerai issues: bots/multi-accounts + luck-seeking submissions
- •Staking enables credible signals of belief; reduces incentive to spam accounts
- •Stake-weighted ensembles improve performance and align incentives
- •Erasure protocol and the broader idea: staking could change internet coordination
- 1:01:08 – 1:11:48
Numerai Signals and the ‘manage all the money’ master plan: more data, more orthogonal signals
Richard introduces Numerai Signals, where users contribute predictions built from their own datasets (alternative data) rather than Numerai’s core dataset. He describes the mission to ‘manage all the money’ by scaling both data and talent, including rapidly expanding dataset size and rewarding uncorrelated signals (e.g., WSB sentiment).
- •Signals: contributors use any dataset to rank stocks; Numerai imports external signals
- •Reward structure emphasizes orthogonality to core signals (uncorrelated alpha)
- •Example: a scraped WallStreetBets sentiment signal integrated into Numerai
- •Ambition: 10× dataset growth per year for a decade; aggregate global modeling talent
- 1:11:48 – 1:18:22
Markets, money, and crypto: what finance enables and why privacy matters
They zoom out to philosophical questions about money and the role of finance in “bringing the future forward” via loans and capital formation. Richard reflects on early crypto skepticism (no cash flows), then Ethereum’s smart contract vision, and raises the practical privacy problem of transparent blockchains (with interest in Zcash).
- •Finance as a mechanism to create future possibilities (loans, investment, innovation)
- •Bitcoin valuation skepticism vs Ethereum’s programmable smart contracts
- •Transparency downside: public ledgers enable de-anonymization and surveillance
- •Privacy coins (e.g., Zcash) as a different technical invention enabling new ‘games’
- 1:18:22 – 1:22:52
Memetics and Dogecoin: how jokes move markets and culture
Lex and Richard explore how memes—amplified by figures like Elon—can become real economic forces. They discuss humor as a high-velocity transmission channel for ideas and speculation, and the tension between playful levity and the risk of a world where nothing is taken seriously.
- •Dogecoin as a case study in memetic momentum and reflexivity
- •Elon as catalyst; the internet as multiplier for jokes into coordinated buying
- •Humor as a propagation mechanism that can influence real capital flows
- •Cultural concern: memes can both relieve pressure and undermine seriousness
- 1:22:52 – 1:49:55
Advice: building startups, learning deeply, and confronting mortality
Richard advises founders to play the “real game” with sincere, bold ambitions—mediocrity is riskier than aiming big in venture contexts. He cautions young people not to start companies casually, praises the learning leverage of good jobs, and closes with reflections on COVID, isolation, and how mortality can reset priorities and ambition.
- •Startup advice: communicate true ambition; don’t tone it down to appear safe
- •Don’t start a company unless it can be your life’s work; avoid ‘quick flip’ mentality
- •Value of jobs as paid education; build expertise at intersections (ML + quant + crypto)
- •Meaning-of-life reflections via illness: isolation, fear, and renewed motivation
