A consortium of researchers led by Google DeepMind has published a white paper proposing Visual General Intelligence (VGI) as an alternative pathway to artificial general intelligence — one grounded in visual experience rather than language.
Published on arXiv on August 26, the paper challenges the dominant language-centric view of AGI development. While large language models like GPT have demonstrated remarkable transfer capabilities through autoregressive training on web-scale text, the authors argue that vision may provide a more fundamental and physically grounded foundation for intelligence.
"What capabilities and forms of intelligence can emerge from visual modalities such as images, videos, and geometry?" the paper asks. The authors, led by Hirokatsu Kataoka of Google DeepMind and including researchers from Oxford, CMU, and other institutions, propose that visual intelligence could develop toward general intelligence through scaling generative video models — similar to how GPT scaled language modeling.
The paper outlines several core principles for VGI: recovering the underlying physical structure of scenes (geometry, materials, pose, and dynamics), lifelong learning without catastrophic forgetting, active perception where agents choose where to look and how to interact, and new evaluation metrics focused on physical consistency, creativity, and continual adaptation rather than static accuracy benchmarks.
A key argument is that complex visual systems evolved hundreds of millions of years before human language, suggesting vision serves as a primary driver for understanding environment, causality, and spatial relationships. The authors propose that autoregressive visual prediction — predicting the next frame in a video sequence — could lead to emergent reasoning capabilities grounded in physical reality, analogous to how token prediction in LLMs produced unexpected abilities.
The paper also acknowledges significant challenges, particularly the enormous computational requirements for training generative video models at scale, and the absence of discrete symbolic units in vision compared to the finite vocabulary of language.
