Summary
This Y Combinator conversation tackles one of AI’s most fundamental open problems: sample efficiency, or why humans can master a new game in a handful of tries while state-of-the-art models often need tens of thousands of training examples. The discussion frames intelligence not as accumulated skill but as the rate of skill acquisition — a framing attributed to François Chollet — and uses ARC-AGI puzzles as a concrete benchmark where humans intuitively excel but frontier models still largely fail.
The theoretical centerpiece is world models: internal simulations of how the world behaves that allow an agent to plan without collecting new environment samples. The conversation uses Newton’s laws of motion as an example of a perfect world model — NASA can intercept an asteroid years in advance without needing a million training runs — then works through how Monte Carlo Tree Search (MCTS) approximates this in domains like Go. Specific numbers ground the discussion: AlphaZero’s MCTS runs 800 simulations of 30-step rollouts per move, requiring roughly 24,000 model invocations per single action, a cost that scales catastrophically as action spaces grow.
The episode explains why MCTS breaks down at scale, why model predictive control offers a complementary path, and what it would actually mean to build a world model capable of human-like generalization. For engineers and researchers thinking about the next architectural leap beyond transformer scaling, this is a clear-eyed technical map of the terrain.
📺 Source: Y Combinator · Published July 17, 2026
🏷️ Format: Deep Dive
