Inspiration: https://xcancel.com/ajeya_cotra/status/2029576934523556256#m
Will a model released this year (2026) be evaluated by METR to have a 50% time horizon of over 100 hours within 6 months of its release?
Resolves using their most up-to-date methodology at the time (currently the 1.1 time horizon). I will tend to try and resolve to the spirit of the question in edge cases, rather than its letter, but I'm happy to take questions and try to provide clarifications in advance.
1,000I'll take that bet. Every prediction I've made going on 15 years now has been correct about the year-by-year development of ML capabilities. Timeline I figured on five years ago, is right on pace, and hasn't changed, and has a remaining window of 18-28 months. What do you think? You have a huge sharpe ratio so I have to ask.