A Call for Intrinsic Learning

Abstract

Current artificial intelligence systems predominantly rely on extrinsic learning mechanisms, with gradient descent and its variants serving as the primary means of model optimization. This approach treats learning as a distinct, external process separate from cognition. However, natural intelligent systems, such as the human brain, display intrinsic learning where learning and cognition are inseparable, integrated processes. We argue for a shift of focus toward intrinsic learning in AI systems, moving away from the heavy reliance on extrinsic optimization. We highlight the limitations of current AI methods, including their extreme sample inefficiency and dependence on vast amounts of human-generated data. By examining the shortcomings of current scaling approaches and proposing alternative pathways, we emphasize that genuine advancements in artificial general intelligence require systems that learn and adapt intrinsically. We encourage renewed attention to AI architectures that embed learning within the dynamics of the system itself, drawing inspiration from natural intelligence to foster more robust, efficient, and adaptive AI.