Hacker News front pageJul 6, 2026, 23:45 UTC+8

AI Superforecasters Are Here: Frontier Models Turn $35 Into $2M on Prediction Markets

The AI Superforecasters Are Here

Scott Alexander reports from a prediction market conference where AI forecasters are now making real money. Preseen turned $35 into $2M on Kalshi in 7 months; FutureSearch claims a 25% market-neutral edge over stocks and similar margins on Kalshi and Polymarket. He tested FutureSearch on the probability of US respiratory infections halving by 2040: the AI deployed 3 sub-agents, read 16 sites, cost $8 and 5 minutes, and returned 7% with 212 sources. Preseen gave 8.8% on the same question. A human superforecaster was also consulted, but the post does not disclose her number. These systems wrap frontier models like Claude or ChatGPT in a scaffold that runs long research chains with sub-agents and tools—slower and pricier than a normal chat, but capable of placing bets.

Why it matters: Scott Alexander reports concrete returns from AI forecasters with named startups and numbers. Held at 78 because it's a personal blog relay without reproducible methodology—strong signal, not yet a paper or product release.

Read the original ↗Export Markdown

Take

Scott Alexander came back from a prediction market conference with two striking numbers: Preseen turned $35 into $2 million on Kalshi over seven months, and FutureSearch claims a 25% market-neutral edge over stocks. He tested FutureSearch on one question—the probability of US respiratory infections halving by 2040. The AI spun up 3 sub-agents, read 16 sites, cost $8 and 5 minutes, and returned 7% with 212 sources. Preseen gave 8.8% on the same question.

The approach wraps frontier models like Claude or ChatGPT in a scaffold that runs long research chains with sub-agents and tools. Slower and pricier than a normal chat, but capable of placing real bets.

I'd discount this on two fronts. First, $35 to $2M is an extreme return with no strategy details, drawdown info, or luck adjustment disclosed—don't read it as a replicable track record. Second, FutureSearch's 25% edge is self-reported, with no independent audit or long public history. But the direction isn't surprising: prediction markets are structured judgment tasks, and AI is more consistent than humans at information retrieval and probability calibration. The real signal is that these systems are moving from papers and benchmarks into real money flows, with a working paid loop.

Source: Hacker News front page · astralcodexten.com
Published: Jul 6, 2026, 23:45 UTC+8 · Jul 6, 2026