The only AI curve that matters (www.exponentialview.co)

This piece introduces the "99% step-length" — the number of sequential actions an AI can perform with at least 99% reliability before a human must intervene — as the single most important curve to watch for real autonomy. Today’s frontier systems reliably manage roughly 100 such steps; the authors estimate that, if current trends continue, step-lengths could exceed 10,000 by 2029 and grow another 3–10× a few years after that, enabling systems to operate unsupervised for weeks or even months. The write-up contrasts this benchmark with recent METR work that measures task-length doubling about every seven months but at 50%–80% success rates, arguing those lower thresholds understate the bar for operational trust. For the AI/ML community this reframes progress: sub‑99% reliability still demands continuous human oversight, while sustained 99%+ step-lengths would make genuinely persistent autonomous agents practical — and consequential. Technical implications include stronger demands on reliability engineering (robustness to distributional shift, silent failure modes, reward hacking), new evaluation standards that measure long-horizon cumulative risk, and urgent governance and monitoring strategies for systems that can act for weeks without intervention. The authors caution that many factors could derail the trend, but emphasize planning for a world where high‑step-length autonomy is plausible in the near term.