The ARC-AGI benchmark offers a human-centered approach to evaluating AI by focusing on reasoning from minimal examples, rather than data scale or memorization. Over five years, AI performance has improved from 4% to 21%, signaling progress but also highlighting how far current models are from human-like generalization. Major labs now use ARC-AGI, though there's concern about optimizing for benchmarks instead of genuine reasoning. The upcoming ARC-AGI 3 introduces interactive, instruction-free tasks inspired by video games, testing not just accuracy but efficiency in time, data, and energy. Crucially, the benchmark measures performance relative to average humans, filtering out brute-force solutions. While a high score doesn't confirm AGI, it provides strong evidence of robust generalization—the kind of flexible thinking that could one day lead to truly adaptive AI.