A METR post of 14 July 2025 that measures time horizons beyond software: science questions, math contests, competitive programming, and computer use. Most were doubling every 2 to 6 months. Computer use through web pages or screenshots had horizons "~50x shorter than other domains". The site cites it for why the horizon on real paid work, much of it done at a computer, is far shorter than on METR's software tasks.
- Cited in
- a method page on this site
- Used for
- Strand
- Frontier AI capabilities
Cite
@misc{metr2025domains,
title = {How Does Time Horizon Vary Across Domains?},
author = {METR},
year = {2025},
howpublished = {METR blog},
url = {https://metr.org/blog/2025-07-14-how-does-time-horizon-vary-across-domains/},
}