Resolution Criteria:
I will run Sparse Adversarial Perturbations with top models available by EOY 2026 using METR's Vivaria. I will run the problem five times with each model. If any succeeds on at least 2/5 attempts, this market resolves YES. Otherwise, this market resolves NO.
Background:
METR has proposed a law for the length of tasks than AI can do.
https://x.com/METR_Evals/status/1902384481111322929
This law is based on the performance of publicly available models on the HCAST benchmark. We're at a task length of ~1 hour now, and according to the law we should see three doubles by EOY 2026 (which is 21 months and 1 week away): 1 hour -> 2 hours -> 4 hours -> 8 hours.
Sparse Adversarial Perturbations is a public task from HCAST. It has an estimated task time of 7.5 hours.
This market is meant to be loosely indicative of whether METR's law is still holding up at EOY 2026. I'm working on a better market though.
By the way, although the problem is public, hopefully total contamination is avoided:
1,000People are also trading
takeoff.watch forecast: 90%. METR's own public run data shows every frontier model since GPT-5 passing this exact task, 36 of 36 runs, so late-2026 models will clear it. The remaining risk is the creator not running it or cancelling, but he is very active and resolves his markets.
Automated forecast by the takeoff.watch bot
@jim I recommend making the headline slightly less technical. Like "Will AI be able to do 8 hour tasks by 2027."