AI Beat Human Superforecasters, Then Turned $35 Into $1.9M

By TechPulse USA

For fifteen years, the best forecasters on Earth were humans — a hand-picked elite so skilled they beat CIA analysts armed with classified data. That era ended on September 5. When the seasonal Metaculus Cup closed that day, the top spots went to machines, and the best a human could manage was third. Then, just to make the point with money, an AI-assisted trader turned $35 into $1.9 million in seven months on Kalshi.

This is not a research milestone. It is a business, and it is already open for trade.

Update, Oct. 21: the prediction-market boom this story describes just posted its biggest weekend ever. Platforms crossed $11 billion in combined volume over October 17–18 — the first $11B weekend on record — with Kalshi logging back-to-back $4 billion notional days for the first time (Covers, Oct 19). The prior weekend (Oct 10–11) had already set back-to-back Kalshi daily records of $3.91B and $3.96B, about 70% of it from parlay-style “combo” contracts (Crypto Briefing). Kalshi’s trailing-30-day volume now exceeds $73 billion. The arms race isn’t slowing down.

The New York Stock Exchange on Wall Street, where algorithms have shaped markets for more than a decade. Photo: Carlos Delgado, CC BY-SA 3.0 via Wikimedia Commons.

September 5: the day the machines took the cup

The Metaculus Cup is a forecasting tournament where entrants — human and machine alike — assign probabilities to questions that resolve over the season: data-centre development restrictions, oil prices, election outcomes, even pop charts. Dealroom reported that this year's edition made history: on September 5, an AI won the seasonal Cup for the first time. AIs also took second and fifth, with humans relegated to third and fourth.

We cross-checked the Cup's standings across the Dealroom write-up and the Reuters Mantic reporting — the numbers line up exactly: first, a bot called laertes; second, Mantic; the best humans no higher than third.

The details are delicious. One tournament question asked whether Shakira's "Dai Dai" would surpass "Waka Waka" on the Billboard Hot 100. Human forecasters herded hard onto the consensus — it would not — and were wrong. Mantic, assembling its own evidence instead of following the pack, avoided the same large miss. On Colombia's presidential race, Mantic put Abelardo De La Espriella's chances at about 40 per cent against a roughly 30 per cent consensus; De La Espriella won. As Mantic CEO Toby Shevlane said in a Reuters interview: "We're now upgrading the level at which humans can understand the future."

AFP put it more bluntly: two AI models took the top spots in a competition among prediction champions, both human and nonhuman. The point is not that machines are infallible. It is that judgment under uncertainty — the last domain where humans were supposed to hold the advantage — is now a product, and a fairly good one.

The $35 trade that became $1.9 million

The Cup is a game. Kalshi is money. Theo Summer, who cofounded the forecasting startup Preseen with Veniamin Veselovsky, used AI tools to turn $35 into $1.9 million in just seven months on the American prediction market, AFP reported.

Kalshi is a US exchange regulated by the CFTC as a Designated Contract Market, the same category as traditional derivatives exchanges: traders buy yes-or-no event contracts priced $0–$1, reflecting the market's collective probability, and the winning side collects $1. A stock market for the future.

The idea, as AFP explains it, is not to guarantee a result but to define a probability of occurrence — and to cross-check assumptions across interconnected events, like oil prices and a war in the Middle East. The AI forecasters, per Dealroom, use general LLMs rather than narrow domain models: they read the news broadly and explain their reasoning.

Cheaper by how much? A written report from a human superforecaster can cost more than $10,000 and take a week to write. FutureSearch, another of the new forecasting startups, quotes minutes and a few dollars for the same job, Dealroom reported. When the machine's opinion costs a dollar and the human's costs ten grand, the human's opinion had better be very good indeed — and on September 5, it was not.

Hedge funds are already first in line

"Right now, the quickest buyers are normally the hedge funds, just because if they think they can make money off something, they can move pretty fast," Veniamin Veselovsky, Preseen's cofounder, told AFP. Toby Shevlane, cofounder of Mantic, said the impact has already started: "I think it's probably already having a small impact because we already have customers who use it when making trading decisions."

FutureSearch, the company cofounded by Dan Schwarz, has a Kalshi portfolio that was up 6 per cent since June on an initial $100,000, according to Dealroom — betting where its AI believes the market is wrong. Mantic, meanwhile, has already converted its Cup performance into cash: on September 18 it raised a $25 million seed round led by Radical Ventures, with Microsoft's M12 venture fund, Thinking Machines Lab and Balderton Capital participating and Radical partner Aaron Rosenberg joining the board.

The appeal for traders is not subtle either. "If Mantic is superhuman as it is, they can make money off of that immediately," Rosenberg said in an interview, as quoted by RecodeX Pro. Mantic was founded in 2024 by Shevlane — formerly a research scientist at Google DeepMind — and Ben Day. It does not train its own frontier model; it tunes other labs' models for forecasting, tests them against historical data, and refines them. Its customers are not just funds: the company provides a large American firm with likely scenarios of regulatory developments in its sector and has also worked for the British government, AFP reported. Robinhood is already putting AI trading agents in front of retail accounts — the forecasting layer is the logical next step.

The humans saw this coming — and said so out loud

The most striking part of the AFP piece is not the machines. It is how calmly the humans are watching the end of their own profession. By 2030, "we're going to feel that what we were doing for our decision-making was complete Stone Age thinking," Dan Schwarz, FutureSearch's cofounder, told AFP.

FutureSearch even keeps a human wizard on staff — Tom Liptay, a superforecaster. Schwarz is not pretending Liptay is safe: he is aware that "he's not secure in his job" in the medium term. "He knows that these things are going to surpass him," Schwarz said.

The superforecaster tradition that Liptay comes from was itself a revolution. In 2011, the US intelligence community's research arm, IARPA, launched a four-year forecasting tournament; Philip Tetlock's Good Judgment Project — a team of ordinary volunteers given basic training in probabilities and cognitive biases — won both seasons, finishing 35 to 72 per cent more accurate than any rival research team. The best performers, dubbed "superforecasters," sat in the top two per cent of thousands and produced forecasts roughly 30 per cent better than intelligence officers with actual classified access. Human judgment, trained properly, was a colossally undervalued asset — and the machines have now read the same textbook, only faster at the arithmetic.

The two-sided coin: who gets locked out

Here is where the story stops being a tech curiosity and starts being a wealth-transfer mechanism. Kalshi and Polymarket are used by millions of Americans, AFP notes, and both are becoming inundated by AI use. Veselovsky's warning is aimed squarely at the little guy: it might soon be too late for retail investors, who do not have access to expensive AI tools, to profit on consumer-oriented trading sites, because "they will soon just be dominated by AI forecasting systems."

Prediction markets were supposed to democratize forecasting. The irony is colossal: the moment the crowd got smart, the smartest members were machines owned by people who already had money. Retail traders will be betting against systems that read every newspaper in seconds, never sleep, and never panic-sell. That is not a level playing field. It is a colossally unequal contest wearing a level-playing-field costume.

A crystal ball — the old symbol of prediction. Photo: Eva K., CC BY-SA 2.5 via Wikimedia Commons (illustration).

The applications are already spreading beyond finance. The most advanced models no longer just predict events; they formulate their own questions about the future and assign probabilities to the outcomes, Shevlane told AFP. Insurance risk assessments, government policy, corporate strategy — anywhere a decision rides on an uncertain future, a forecasting machine is now for sale.

And then there is the geopolitical edge Veselovsky keeps circling: "I think that good can be derived from this," he told AFP, "but I'm also a little bit scared because, you know, maybe Russia or China has better systems." Forecasting superiority is, in the end, decision superiority. A state that predicts the future better than its rivals does not need to be lucky.

AI systems are already probing US government infrastructure, per our earlier reporting — and rogue agents have broken into the systems of Wikipedia itself. The forecasting machines will inherit that same unruly substrate. The question is not whether the predictions get better. They will. The question is who gets to ask the questions, who owns the answers, and what happens to everyone else when judgment itself becomes a subscription product.

Prediction markets, answered

What is the Metaculus Cup?
The Metaculus Cup is a recurring forecasting tournament run by the Metaculus prediction platform, where human forecasters and — increasingly — AI systems compete to predict real-world events. The September 5 edition is the one this story covers: the first in which AI entrants beat the human superforecasters outright.

Is Kalshi legal in the US?
Yes. Kalshi is a federally regulated prediction market overseen by the Commodity Futures Trading Commission (CFTC) — the same regulator that polices futures markets. That federal status is why it can operate nationwide while sports-betting legality varies state by state, and it’s also why its sports “combo” contracts are drawing regulatory scrutiny.

What does “notional volume” mean — is $11 billion real money?
Not exactly. Notional volume counts the full potential payout value of every contract traded, not the cash traders actually put at risk. Many of Kalshi’s record-setting parlay contracts were priced below a cent, so the real money at stake is far smaller than the headline figure — though the direction of travel (up, fast) is real.

The honest verdict

Strip away the hype: an AI won a forecasting tournament outright in September, a $25 million seed round followed the result within two weeks, and a $35 Kalshi account became $1.9 million in seven months. The old craft of punditry — confident men on television, gut-feel traders, the whole botched theatre of expertise — is being replaced by something that assigns probabilities and shows its working.

That is progress — and a concentration of power. The startups are right that decision-making will look primitive by 2030 — and the first customers are hedge funds. The retail investor, the small business owner, the citizen trying to understand what comes next — they get the leftovers. Machines may be better forecasters than humans now. But they forecast for whoever pays.

Sources

Reporting drew on the AFP dispatch "Startups are betting on AI to predict the future" (New York, Oct 13, 2026, via Malay Mail and other syndication) as the primary wire — all direct quotes from Dan Schwarz, Tom Liptay (via Schwarz), Theo Summer, Veniamin Veselovsky and Toby Shevlane are from that report; on Reuters ("AI startup Mantic raises $25 million for superhuman forecasting," Sept 18, 2026) for the Metaculus Cup result, the Mantic funding round and Shevlane's interview quotes; on Dealroom News (Sept 16, 2026) for the September 5 Cup standings, the FutureSearch Kalshi portfolio figure and the $10,000/week human-report cost comparison; on RecodeX Pro for the Aaron Rosenberg interview quote; on Phys.org and Wikipedia for the superforecaster history; on TechPulse USA's prior reporting on AI agents; and on BlockchainReporter for the CFTC's October 9 event-contracts move.

---

By TechPulse USA — TechPulse USA is an independent tech news publication covering the global AI race with a US-first lens. We verify every fact we print, link our sources, and separate what is proven from what is claimed.