Why I don't think AGI is imminent (dlants.me)

Recent discussions surrounding the imminent arrival of Artificial General Intelligence (AGI) have ignited significant debate within the AI/ML community, particularly in light of claims made by CEOs of leading AI firms like OpenAI and Anthropic. A contrasting perspective emphasizes the substantial cognitive gaps that transformer-based large language models (LLMs) must overcome to achieve human-level cognition. The article outlines foundational cognitive primitives—such as object permanence and spatial navigation—rooted in vertebrate evolution, which current AI architectures lack. For instance, while models trained on extensive data can track objects in videos, they struggle with underlying concepts like persistence and logical relationships, often leading to fragile performance in dynamic scenarios. The advancement of AI may hinge on developing models that not only observe but also interact within simulated environments, potentially enabling them to learn these essential cognitive primitives through embodied experience. Current projects like DeepMind's SIMA 2 and Dreamer 4 are exploring this realm but demonstrate that the interplay between embodied competence and language reasoning remains poorly understood. Benchmarks such as Stanford's ENACT reveal a stark contrast in performance between current models and human cognitive capabilities, underlining a long journey ahead. The piece calls attention to vital research avenues that could redefine the landscape of AI and move us closer to true AGI, although it suggests this endeavor may take decades to materialize.