Preseen

Over the past year I've been building out Preseen. We’re focused on building the best AI forecasting agent for decision makers. Every decision relies on a forecast. Too often decisions are poorly made because the underlying forecast is bad. Humans are constrained. But if you can have a system that can forecast anything as well as the best humans, then human decision making will be improved. AIs are reaching the point where they are better than some of the best human forecasters. This is going to change decision making. Preseen will help drive that change.

AI superforecasters are going to be much better than human ones. Excited about what @VminVsky is building with Preseen.

Every decision is a forecast. Better forecasts mean better decisions. @VminVsky and Theo are building the best forecasting agents at @AskPreseen. I spent my career building systems, so I want to be precise about why this is a hard technical problem and not a wrapper. A language model out of the box is a poor forecaster. It reasons from priors frozen at training, inherits the biases of whatever it retrieves, and will happily give you a confident number with nothing underneath it. Getting from there to a calibrated probability means solving several problems at once. You need fresh, structured data about the state of the world, which is an ingestion and normalization problem across thousands of heterogeneous sources, not a search query. You need to force independence between lines of reasoning so errors stay uncorrelated, which Preseen does by running separate agents against primary records and reconciling their estimates the way you would aggregate a panel of analysts. And you need calibration, which is its own discipline: a system that says 80% has to be right about eight times in ten, measured across hundreds of resolved questions. Forecasting is one of the rare AI problems with delayed ground truth. Every question eventually resolves, so every forecast becomes a labeled training example. Preseen has been accumulating that feedback since last fall, on top of a data layer Venia spent months building: crawlers and normalization pipelines across country APIs, satellite imagery, procurement records, and stranger signals like diplomatic flight patterns and international school enrollments. Last December that machinery flagged unusual Russian flight activity before it showed up anywhere humans could see it. A general-purpose model gets none of this for free. The loop has to be built, and it gets stronger with every resolved question. The Preseen bot won the FutureEval tournament last fall, first out of 164. It came third out of 1,283 in the Metaculus Cup this spring, ahead of some of the best human forecasters alive. Then in Q2 it won Market Pulse outright, the first time a bot has ever beaten a field of humans to win a Metaculus tournament, and it won by a wide margin. Theo turned $35 into roughly $1.9 million on Kalshi over seven months. The trend line is very clear that Preseen will soon beat even the best human forecasters. Any decision that matters will need AI forecasting at the table. Everyone wants to know the future. I am very proud to lead Preseen’s seed round at @spc. They’re hiring.

"The future is already here -- it’s just not evenly distributed." -- William Gibson