“The performance of base LLMs on ARC-AGI V1 remained low despite a 50,000x increase in model scale, indicating that scaling pre-training alone was insufficient for fluid intelligence.”