Every domain, and where it stands.

Would be measured on OSWorld success rate, from OSWorld leaderboard. Needs a dated OSWorld / WebArena series. This is the gap most likely to set an AGI floor.

Reported elsewhere: Horizons 40–100× shorter than software, improving at a similar rate.

Would be measured on RLBench success rate, from RLBench. Needs a dated RLBench horizon series with human baselines.

Reported elsewhere: Improving far slower than the software cluster.

Would be measured on miles between disengagements, from California DMV disengagement reports. Needs a defensible series; the public tracker is community telemetry, not a controlled benchmark.

Reported elsewhere: About 0.6 doublings per year, roughly five times slower than software.

Clinical work carries over weeks and the outcome is often unobservable at the horizon of the task, so "a doctor takes N hours" is not a well-posed baseline for most of it. Existing medical benchmarks score answer accuracy on vignettes, which is a different quantity from how long a case a system can carry unattended.

Billable-hour records exist in enormous volume but are confidential and are not task-decomposed, so the one dataset that would give human baselines cheaply is the one nobody can publish. Public legal benchmarks test retrieval and classification, not how long a matter a system can run.

The work whose value is easiest to measure is also the work whose results are least likely to be shared, and success is confounded with market luck over exactly the horizons that would need measuring. Timed professional baselines exist inside firms and are not published.

Video length is used as a proxy for task length, and METR flag that the correlation with difficulty is weak. A two-hour film is not a two-hour task. The instrument is wrong here, not merely missing.