Because ai is following a predicable trend of exponentially increasing task length @50% probability: https://x.com/METR_Evals/status/1902384481111322929

If we are the top of an s-curve, the recent samples on that curve would be below the trend line, not above it.