Ord reports two concrete comparisons. In a METR comparison, humans gained about four times as much score per doubling of resources as the early AI systems tested. In the UK AI Security Institute chart, based on capture-the-flag hacking tasks, the estimated token increase needed for a tenfold improvement in the 80% time horizon fell from roughly 10,000× for earlier models to about 20× for models through May 2026.
Mechanism: A sub-unit slope on a log-log plot means token budget increases translate into less-than-proportional growth in estimated human-time task horizon; Ord says the fitted exponent improves across model releases. Why now: The newer AISI data indicates the scaling relationship may have improved substantially within less than a year, making extrapolations from early models unreliable. First implication: Teams should measure task success and horizon against inference budgets directly rather than assume more tokens substitute proportionally for human time. Evidence boundary: The essay relies on limited task suites and indirect comparisons, including early models in the METR comparison and internal hacking tasks in the AISI chart through May 2026; it does not establish a general capability law. First watch condition: A like-for-like study of recent models and humans across the same tasks and time budgets could show whether the improving exponent continues toward parity or plateaus.