Core Viewpoint - The current AI models face a significant technical dilemma known as the "split-brain" problem, where the quality of answers varies drastically based on how questions are phrased, indicating a lack of generalization ability in handling tasks outside their training materials [2][3]. Group 1: Training Challenges - The "split-brain" issue often emerges during the later stages of model training, where models are fine-tuned with curated datasets to learn specific domain knowledge or improve conversational style [6]. - Fixing errors in AI models can lead to new problems, akin to a "whack-a-mole" game, where addressing one issue may inadvertently create another [6]. - The complexity of model training highlights the need for appropriate data combinations, which is why AI developers invest heavily in domain experts to generate training data [7]. Group 2: Limitations of AI Models - Current AI models do not possess a true understanding of how the world operates, which is a fundamental limitation compared to human cognition [3][7]. - This lack of understanding implies that models struggle with generalization and cannot effectively handle tasks outside their training scope, raising concerns for investors who expect breakthroughs in fields like medicine and mathematics [8].