Dwarkesh Patel argues that much recent AI progress comes from expanding and improving training data, including expert demonstrations and reinforcement learning, rather than major gains in sample efficiency.
He compares human and machine learning across language, robotics and driving, and responds to objections about evolution, sensory data and model scale. These comparisons are his argument and estimates, not established equivalences.
Patel considers whether data-intensive systems can still automate common work and discusses the harder question of automating AI research. He presents his predictions as a perspective rather than a settled outcome.
Watch on YouTube
