Is it possible to achieve general intelligence (AGI) with current deep learning architectures? - tlooto, The Most Powerful AcademicGPT

Artificial general intelligence (AGI) is commonly defined as the capability of a system to understand, learn and apply knowledge flexibly across domains at least as well as humans. In contrast, today’s deep learning models excel primarily at narrow tasks for which they are trained—speech recognition, image classification or game playing—yet often fail when asked to generalize beyond their training distribution [1]. 1. Achievements of current deep architectures Modern deep networks—particularly convolutional nets, recurrent nets and, more recently, transformer-based models—have demonstrated superhuman performance on benchmarks in vision, language and control. They learn rich internal representations via deep hierarchies of nonlinear features and scale effectively with data and compute, yielding emergent abilities such as few-shot and zero-shot learning in large language models [1][2]. For many practical tasks these models approximate highly complex functions $$f\theta: x\mapsto y$$ with