Why does it matter?
Month-long autonomy would mean AI can take on real projects, not just short benchmark tasks.
Direct quote
Because our analysis puts low probability on the growth rate of time horizon on our tasks being much slower than one doubling every 8 months, the uncertainty in the extrapolated date of 1-month AI is fairly small (80% CI width about 2 years, central estimate mid-2029). If future progress instead follows the 2024-2025 trend, 1-month AI would arrive sooner, with half the probability in 2027 and even late 2026.
METR