A NEW, CHALLENGING AGI TEST STUMPS MOST AI MODELS

Summary from the TLDR newsletter

The ARC-AGI tests consist of puzzle-like problems where an AI has to identify visual patterns from a collection of different-colored squares and generate the correct “answer” grid - reasoning AI models like OpenAI's o1-pro and DeepSeek's R1 score between 1% and 1.3%.