Artificial Intelligence (AI) is traversing a pivotal inflection point in its historical trajectory. Over the past decade, the field has been dominated by large-scale statistical deep learning, culminating in transformer-based large language models (LLMs) containing hundreds of billions of parameters and trained on vast corpora of human-generated data. These systems have demonstrated remarkable capabilities in natural language processing, multimodal understanding, and generative tasks, fundamentally reshaping human–computer interaction. However, the rapid expansion of scale has also exposed structural limitations: data inefficiency, weak causal reasoning, limited grounding in physical reality, and fragile generalization outside training distributions. As a consequence, the research frontier of artificial intelligence is now bifurcating into two complementary trajectories
Zen Revista (Sun,) studied this question.