"Image synthesis assisted by Grok Imagine Image, an AI partner within the Global Future Nexus ecosystem."
Vision is the oldest and most foundational of the senses. Advanced visual systems—camera-type eyes and compound eyes—date back to the Cambrian period roughly half a billion years ago, driving major changes in biological interactions long before language, writing, or any other symbolic system emerged. Over 80% of the information processed by the human brain is obtained through the eyes. Yet for all the remarkable progress in artificial intelligence, vision remains the modality where machines lag furthest behind. As researchers at a 2026 workshop on Visual General Intelligence observed, "it is not obvious that intelligence should be reduced to language". Sight may be more than a sensor; it may be the foundational intelligent function for capturing the structure of the world.
The Visual General Intelligence Paradigm
The most significant conceptual shift is the emergence of Visual General Intelligence (VGI) as a distinct pathway toward AGI. A white paper published in August 2026, drawing on contributions from researchers at OpenAI, Google DeepMind, Oxford, MIT, and Stanford, argues that intelligence emerging from visual experience and learning may provide a path to general intelligence that is fundamentally different from the language-centric approach that has dominated the field.
The paper's central insight is that vision is not merely another input modality. It is the substrate upon which many cognitive processes are built. As the authors note, in the human brain, visual and spatial reasoning abilities are far older and more foundational than language-based logical reasoning. To teach someone to navigate a maze, a verbal description confuses; a sketch makes it instantly clear. This suggests that a truly general intelligence must be grounded in visual understanding, not merely capable of processing images as a preprocessing step before text-based reasoning.
The Visual Reasoning Bottleneck
The current generation of vision-language models (VLMs) suffers from a fundamental architectural limitation. They must first convert visual input into language, then perform text-based reasoning. But many visual tasks cannot be precisely described in words. As Elorian AI, a startup founded by former Google DeepMind and Apple researchers, has argued, this creates a bottleneck: even the most advanced models like Gemini 3 Pro achieve only the level of a three-year-old child on BabyVision, a benchmark for basic visual reasoning.
The field is responding with new architectures. E-ViC (Embodied Visual Chain) moves reasoning beyond text and directly into the visual domain. By formulating visual operations like zooming and marking as executable primitives, it transforms perception from static prediction into an active verification process. Trained via agentic reinforcement learning, it autonomously discovers human-like "look-and-confirm" strategies driven solely by task-level rewards, consistently outperforming strong VLM baselines by an average of 10.1%.
The Neuromorphic Vision Revolution
At the hardware layer, a parallel revolution is underway. Traditional image sensors face significant challenges in energy efficiency and processing capability when storing, transferring, and processing massive amounts of visual data. The human visual system, by contrast, performs the entire process from light sensing to information processing with extremely low power consumption.
A major breakthrough came from Tsinghua University's development of NEOSTI, a neuromorphic electronic-opto spatial-temporal hybrid image sensor that integrates processing-pre-sensor in the optical domain, processing-in-sensor with nonlinear acquisition, and processing-near-sensor in the electronic domain. This enables parallel data computing while sensing, achieving competitive performance across several visual processing tasks.
Even more significantly, researchers at Tsinghua's Center for Brain-Inspired Computing have built a brain-inspired visual sensing-perception learning framework based on their "Tianmou Chip" (published as a Nature cover in 2024). Unlike traditional methods that rely on single-image end-to-end mapping, this framework uses different visual pathways to extract color, brightness, motion direction, and edge information—what the researchers call "visual primitives"—and learns to combine them into more complete, stable visual representations. The system can autonomously judge the reliability of information and compensate for interference or missing data, maintaining stable visual perception even under high-speed motion, strong light changes, and multiple interferences. In video instance segmentation, it achieves 115.1 frames per second while significantly improving detection accuracy.
Embodied Vision and World Models
The integration of vision with action is the frontier of embodied intelligence. Meta AI's research on embodied AI agents argues that world models are central to reasoning and planning, allowing agents to understand and predict their environment, understand user intentions, and perform complex tasks autonomously.
This has been operationalized in systems like Puffin-World, a unified multimodal world model that defines physics, geometry, and appearance as three native 3D world states. It connects gravity field perception, free-viewpoint spatial simulation, 3D world generation and reconstruction, and closed-loop exploration to predict the state of the world after a robot acts. Similarly, RynnVLA-002 unifies Vision-Language-Action and world models, enabling joint learning of environmental dynamics and action planning. It achieves 97.4% success rate on the LIBERO simulation benchmark without pretraining, and its integrated world model boosts real-world success rates by 50%.
At the edge, Om AI has released VLX, the world's first streaming multimodal model series designed for physical-world embodied scenarios. VLX-Flow provides real-time streaming perception, VLX-Seek enables precise localization, and VLX-Go translates perception into action decisions—forming a complete perception-grounding-action loop that can run on phones, drones, and robots.
Aligning Machine Vision with Human Vision
A crucial frontier is representational alignment—making artificial vision models process information more like the human brain. A 2026 study in Communications Biology introduced ReAlnet, a vision model aligned with human brain activity based on non-invasive EEG. The model achieves an average similarity improvement of approximately 3% and a maximum relative improvement ratio reaching up to 40% compared to traditional computer vision models. This alignment framework takes an important step toward bridging the gap between artificial and human vision.
The efficiency of these systems is also being democratized. KAIST researchers, in collaboration with MIT and Microsoft, developed "Upsample Anything," a universal technology that improves AI visual performance even with limited GPU memory. Presented at CVPR 2026, it received the CVPR Compute Gold Star Award for efficient use of computational resources.
Governance Implications
For Global Future Nexus, the development of visual AGI carries profound governance implications. A system that can see is a system that can surveil. The same visual capabilities that enable a robot to navigate a disaster zone, a drone to inspect infrastructure, or a medical AI to diagnose from images can be turned toward monitoring, profiling, and control.
The tension is captured by what researchers call the local justice–global externality paradox. High-value AI innovation clusters in wealthy economies depend on extraction of minerals, data, and labour from regions with weaker protections. Systems designed to enhance equity within one jurisdiction often rely on transboundary infrastructures that generate environmental degradation elsewhere.
The path forward requires a governance framework that treats visual intelligence not as a tool for domination, but as a capability for understanding. The "Visual General Intelligence" white paper calls for the computer vision community to clarify the principles it should pursue in the AGI era. That clarity must include not only technical benchmarks but ethical boundaries—ensuring that the sight we give to machines serves human flourishing, not surveillance.
The Path Forward
The dominant sense is awakening. The visual cortex of AGI is being built not in the clouds of language, but in the pixels and geometry of the physical world. As the VGI white paper concludes, vision may be the pathway to intelligence that captures the structure of reality itself. The question is not whether AGI will develop the capacity to see. The question is what it will choose to look at—and who will hold the lens.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
