The Key Thing Human Brains Have That AI Is Trying To Learn
World models are the most promising path to closing AI's sample efficiency gap with humans.
“the two major problems that we have left to solve is intelligence per watt and intelligence per sample.”
A Y Combinator discussion frames sample efficiency—how quickly models learn from limited data—as one of AI's defining unsolved problems, proposing world models as the leading research direction. The conversation references ARC-AGI as a concrete benchmark where frontier models still fail despite training on the entire internet, while humans solve the puzzles intuitively. The framing of intelligence as 'rate of skill acquisition per sample' (crediting Chollet) positions this as foundational to AGI research rather than incremental capability work.