EquiLibre's AI Has Never Had a Losing Month. That's the Scary Part.
A poker-crushing AI now trades billions a day. The perfect record is the tell, not the proof.
Three researchers left Google DeepMind, built the first AI to beat professional players at no-limit poker, and pointed the same machinery at the stock market. Their company, Prague-based EquiLibre Technologies, just closed a Series A at a valuation north of $500 million, led by Creandum, the firm’s largest single investment ever. The round amount is undisclosed. Early backers include Richard Sutton, who more or less wrote the book on reinforcement learning.
Here’s what EquiLibre does. It trains reinforcement-learning agents, AI that learns by trial, reward, and repetition, and turns them loose on live markets. The company says its agents trade billions of dollars a day on instruments like the S&P 500 and Nasdaq, through a partnership with a quant firm, and that they haven’t had a negative month since going live in early 2025.
That last claim is the one everybody’s repeating. It’s also the one that should make you slow down, because the company hasn’t disclosed the time period, the method, or the drawdown behind it, per Tech Funding News.
The mental model here is Ergodicity.
When most people hear “never a losing month,” they average it in their head and file it under low risk. Ergodicity is why that instinct misfires. A system is ergodic when the average across many players equals the average for one player over time. Markets aren’t that. If a thousand agents each clear 5% a month but one blows up, the group average still looks spectacular. The one that blew up is done. It doesn’t get to keep compounding.
That’s gambler’s ruin. Your average return can be gorgeous and completely irrelevant, because you only walk one path through time, and a single deep drawdown ends the walk. A monthly win rate tells you almost nothing about that. Max drawdown, behavior in a market regime the model has never seen, leverage under stress: that’s what decides whether “never had a losing month” survives contact with the first month that’s genuinely different.
None of this means EquiLibre isn’t excellent. Beating the most liquid markets on earth is absurdly hard, and this team might be the one to do it. But I spent years at Robinhood building derivatives and prediction markets, and the first thing you learn on that desk is that the P&L people brag about and the risk they’re actually carrying are two different numbers. It’s the same discipline we design around at /mkt in regulated markets: the question is never the good day, it’s the worst one you’ve modeled.
So here’s my take.
Don’t ask EquiLibre how often it wins. Ask what the worst month looked like, over what window, and what it would take to break it. A perfect record isn’t proof of safety. Sometimes it’s just proof the bad month hasn’t shown up yet.
If this was useful, share it with someone who builds things. And if you want the full toolkit of 50 mental models, you can grab my book, Mental Models: How to Think, Act, and Win, on Amazon right now.
If you want the mental models behind breakdowns like this, my book, Mental Models: How to Think, Act, and Win, is on Amazon now.
Informational and educational only. Not investment advice, and not an offer or solicitation of any security. Figures on volume, growth, and licenses are as reported by the company and haven't been independently verified; Cyclops is private and didn't disclose a valuation. /mkt is referenced solely as an operating example of building in regulated markets, not as an investment opportunity.
