A new artificial intelligence (AI) model by OpenAI, known as o3, recently achieved human-level results on the ARC-AGI benchmark, a test measuring "general intelligence". It scored 85%, significantly outperforming previous AI bests and comparable to the average human score. Creating artificial general intelligence (AGI) is a primary goal for major AI research labs, and this result suggests progress towards that aim. The ARC-AGI test assesses an AI's "sample efficiency"—its ability to adapt to new situations with minimal data. Existing AI, like GPT-4, requires extensive data to perform tasks, struggling with less common scenarios due to limited examples. For AI to handle varied, unpredictable jobs, it must generalize from few data points—a key element of intelligence. OpenAI's o3 succeeded by mastering grid square patterns, solving puzzles with limited examples much like human IQ tests. Although the specifics of o3's functionality are unclear, its adaptability is evident. It identifies "weakest" rules that cover new situations with minimal assumptions, enabling greater adaptability.

This process resembles Google's AlphaGo AI, which used "chains of thought" to solve tasks. Each chain represents a potential solution, evaluated using a heuristic, or guiding rule, to choose the most suitable one. Despite the promising test results, it's uncertain if o3 truly advances AGI closer to human-like intelligence. Its success might not indicate inherent improvement over previous models but could be due to specialized training for ARC-AGI. OpenAI has not fully disclosed details about o3, so its true potential remains speculative. Understanding o3 will require thorough evaluation and could reveal its capability to rival human adaptability. If so, it might revolutionize economies and technology, ushering in new considerations for AGI governance. If not, while still impressive, it would leave daily life largely unchanged.