Amaç Herdağdelen revises METR time horizon lag analysis
The author argues that METR's eight-month gap between 50% and 80% AI time horizons is largely driven by task transition spread, where tasks of the same duration are solved months apart rather than transitioning slowly as a group. By fitting a hierarchical Bayesian item response model, the author finds that individual tasks transition in a median time of about four months. The analysis also explores implications for labor displacement and shows that pooling tasks with constant hazard rates can falsely suggest falling hazard rates. Open article →