Scott analyzes whether AI forecasters will plateau near human levels or keep improving, arguing that even non-miraculous progress could yield 4-12 percentage point improvements in prediction market accuracy. Longer summary

Scott examines whether AI forecasters will plateau near current human performance or continue improving far beyond it. He analyzes a 2010 study by Daniel Reeves showing prediction markets only marginally outperformed simple statistical models, but argues this doesn't necessarily limit future progress. The post reframes those results to show they actually represent meaningful gains in difficult-to-predict domains like sports, then applies this reasoning to geopolitical forecasting. Using three different 'anchors' - comparing AI improvements to past technological gains, human-to-superhuman forecaster improvements, and chess AI progress - Scott estimates that maximally capable AI forecasters could improve prediction market accuracy by 4-12 percentage points over current levels, which he argues would be genuinely valuable even if not miraculous. Shorter summary