Dwarkesh Patel: 'The Data Black Hole at the Center of AI'

Dwarkesh Patel·medium signal

Patel argues AI progress is overwhelmingly data-driven rather than architecture-driven: frontier models train on tens to hundreds of trillions of tokens versus the ~200M a human sees from birth to adulthood — a roughly million-fold sample-efficiency gap. The thesis reframes 'scaling' as fundamentally a data problem and casts architecture tweaks as second-order. Published June 19, 2026.