Dwarkesh Patel reframes scaling as a data problem, not an architecture one.

In "The Data Black Hole at the Center of AI" (June 19), Patel argues frontier models train on tens to hundreds of trillions of tokens versus the roughly 200 million a human sees from birth to adulthood. That's a million-fold sample-efficiency gap. His thesis casts architecture tweaks as second-order and data as the real lever. If he's right, the competitive moat isn't your transformer variant, it's your data pipeline. That reframing is worth sitting with if you're betting a roadmap on architectural cleverness.