Instead of abstract 'AGI' labels, track how long a system can reliably pursue a single task end‑to‑end (its task‑time horizon) and watch that horizon extend. The post cites current limits and extrapolates to about one‑week reliability by 2030–31 and one‑year reliability by 2034, after which broad substitution risks rise.
— A simple, dated yardstick helps policymakers, investors, and regulators calibrate timelines and thresholds for AI oversight and economic planning.
Arnold Kling
2025.10.04
80% relevant
Julian Schrittweiser projects models will 'work for full days (8 hours)' by mid‑2026 and match human experts across industries by end‑2026—an explicit task‑time horizon forecast.
msmash
2025.09.29
90% relevant
Claude Sonnet 4.5 autonomously worked for 30 hours to complete an end‑to‑end software project, a clear extension of the 'task‑time horizon' metric (from prior 7‑hour runs) that the idea proposes for tracking real AI progress.
Arnold Kling
2025.09.24
100% relevant
J. Zachary Mazlish: 'monitoring task‑time’s is, for now, the best way of monitoring AI progress,' projecting one‑week/80% reliability by 2030–31 and one‑year by 2034.
← Back to All Ideas