AI models require vastly more data than humans to learn the same tasks, suggesting current training approaches have hit a fundamental efficiency barrier.

What happened: The article argues that frontier AI models are trained on between 10s to 100s of trillions of tokens—close to a million-fold more than the ~200 million tokens a human sees from birth to adulthood. Even after pretraining, AI models require hundreds of human expert examples per task (generated by specialists across fields like legal, management consulting, and document conversion), whereas humans learn new skills with far fewer examples once educated.

Why it matters: Data—not hardware, model size, or algorithmic tricks—appears to be the primary driver of AI progress. The article notes that the data industry producing expert labels and reinforcement learning (RL) environments is earning billions annually and soon "deca-billions." This suggests that as AI labs pursue white-collar automation and AI research automation, they remain heavily dependent on scaling data collection rather than solving the underlying sample-efficiency problem humans solve.

What to watch: Scaling current model parameters by an order of magnitude would reduce data requirements by only a factor of ~10, according to the Chinchilla scaling laws cited. This implies humans are thousands to millions of times more sample-efficient than current AI models and may operate on a fundamentally different learning curve—a gap that raw parameter growth alone cannot close.