Crashing Waves vs. Rising Tides: AI Automation's Surprising Path
This presentation examines groundbreaking empirical research on how AI automation is actually progressing across thousands of real labor-market tasks. By analyzing over 11,000 text-based tasks evaluated by experienced workers against 40+ language models, the research reveals whether AI capabilities are advancing in sudden surges for specific tasks or rising broadly across the board—with profound implications for workforce adaptation and automation forecasting.Script
A decade of AI hype has promised us that automation will transform work overnight. But when researchers actually tested 40 language models on over 11,000 real workplace tasks and asked thousands of experienced workers to judge the results, they discovered something surprising about how this transformation is actually unfolding.
The authors frame two competing visions of how AI automation advances. The crashing waves model predicts sudden breakthroughs that dramatically transform narrow categories of work, while the rising tides model suggests capabilities improve steadily and evenly across all types of tasks. Which dynamic actually governs the automation of real labor-market work?
The data reveals a clear answer.
The core finding challenges conventional wisdom. When the researchers plotted AI success against task duration on a logarithmic scale, they found an extraordinarily shallow slope. A task that takes 10 times longer to complete drops AI performance by less than 8 percentage points. This flatness means that language models don't hit sudden capability walls as tasks grow more complex or time-consuming—they maintain relatively consistent performance across short and long assignments alike.
This visualization separates two distinct sources of improvement. Larger models, those exceeding 100 billion parameters, excel disproportionately on quick tasks, creating an outward rotation in the success curve. But when you compare models of the same size released at different times, you see something different: newer models improve uniformly across the entire duration spectrum. The rising tide lifts all boats, from 10-minute tasks to week-long assignments.
The rising tide pattern creates a dual reality. On one hand, the trajectory is rapid: current extrapolation suggests language models could handle the vast majority of text-based knowledge work at acceptable quality within five years, with capabilities doubling every few months. On the other hand, the mathematical structure of this progress means reaching true perfection will take far longer than optimists expect. The logistic curve flattens at the extremes, and real-world deployment faces friction that raw capability improvements cannot solve alone.
This research recalibrates how we should think about AI automation: not as a sudden tsunami that will crash over specific industries, but as a rising tide that steadily expands what machines can do across nearly all knowledge work. Visit EmergentMind.com to explore this paper in depth and create your own research video presentations.
