Published on January 6, 2026 6:26 PM GMTWe built OpenForecaster, an 8B model trained to make predictions on open-ended forecasting questions. It is competitive with much larger proprietary models in held-out testing. We train it on our OpenForesight dataset which has 52k forecasting questions created automatically from global news. The news -> forecasting question recipe is entirely automatic, making it easy to reproduce for new, or more data. Our training improves forecasting accuracy, calibration, and consistency of long-term predictions. Calibration from forecasting training generalizes to OOD benchmarks. Crucially, we open-source data, code, model and describe the full approach in our paper, which is summarized in the linked blog. We believe accelerating progress on AI forecasting is important, and open-sourcing achieves this better than existing organizations developing the same tech closed-source, for profit. AI forecasters could help us better understand potential impacts of decisions, and more broadly improve discourse by being easier to scale than human superforecasters, while having relatively lower side effects on harmful capabilities. But to get there, a lot of research needs to be done, which is why its important to engage AI (and forecasting) researchers across the board. Looking forward to any feedback!If you like our work, do share our X announcement.Discuss Read More
The OpenForecaster Project
Published on January 6, 2026 6:26 PM GMTWe built OpenForecaster, an 8B model trained to make predictions on open-ended forecasting questions. It is competitive with much larger proprietary models in held-out testing. We train it on our OpenForesight dataset which has 52k forecasting questions created automatically from global news. The news -> forecasting question recipe is entirely automatic, making it easy to reproduce for new, or more data. Our training improves forecasting accuracy, calibration, and consistency of long-term predictions. Calibration from forecasting training generalizes to OOD benchmarks. Crucially, we open-source data, code, model and describe the full approach in our paper, which is summarized in the linked blog. We believe accelerating progress on AI forecasting is important, and open-sourcing achieves this better than existing organizations developing the same tech closed-source, for profit. AI forecasters could help us better understand potential impacts of decisions, and more broadly improve discourse by being easier to scale than human superforecasters, while having relatively lower side effects on harmful capabilities. But to get there, a lot of research needs to be done, which is why its important to engage AI (and forecasting) researchers across the board. Looking forward to any feedback!If you like our work, do share our X announcement.Discuss Read More
Related Posts
Could space debris block access to outer space?
Published on December 17, 2025 3:59 PM GMTThis is a linkpost for forethought.org/research/space-debris-and-launch-denial.Nearly 15,000 tons…
Unless its governance changes, Anthropic is untrustworthy
Published on November 29, 2025 5:42 AM GMTAnthropic is untrustworthy.This post provides arguments, asks questions,…
Is the Constructible Universe All There Is?
Published on December 13, 2025 2:29 AM GMTFor more than 2000 years, geometers and mathematicians…