The Benchmark With No Instructions — ARC-AGI-3 (winning team!)

ARC-AGI-3 shifts the focus of AI research toward agency and dynamic goal acquisition, challenging systems to solve novel problems without relying on pre-existing, static knowledge. While LLMs demonstrate potential by leveraging language-based reasoning and code generation, they frequently fail to generalize across abstraction levels, often getting trapped in inefficient, incorrect hypotheses. The central tension lies in whether intelligence requires innate, Platonistic primitives or if it emerges from the statistical compression of fractured, path-dependent experiences. Current solutions rely heavily on "harnesses"—structured requirements and guidance—to steer LLMs, yet this raises fundamental questions about whether these models truly understand underlying domain constraints or merely simulate competence. Ultimately, the benchmark forces a confrontation with the "bitter lesson," testing whether scaling compute and data can replace the deep, structural abstractions characteristic of human intelligence.