What are experts saying on what computer vision and visual AI looks like in the next 2-5 years?

Experts anticipate the next two to five years in computer vision will be defined by a push beyond generative content toward a deeper understanding of the physical world. Dr. Fei-Fei Li identifies "world modeling" and spatial intelligence as the next major phase of AI development, a sentiment echoed by startups now focusing on these areas [4, 14]. This involves moving beyond text to richer, multimodal experiences, with transformative improvements expected in AI's ability to use images and video as inputs and outputs for productivity and education within the **next 12 months** [5, 12]. The viral success of image and video generation models has already demonstrated that multimodal capabilities are a primary driver for mainstream user engagement, shifting the competitive landscape beyond pure language model performance . Technologies like Neural Radiance Fields (NeRFs) are considered a significant revolution in enabling 3D computer vision, which is foundational for creating these sophisticated spatial models .

Despite recent breakthroughs in generative video, a tension exists regarding the maturity of current visual AI capabilities. While some models have achieved massive viral reach, Oriol Vinyals of Google DeepMind believes the industry **has not yet seen the equivalent of the "GPT moment"** for video and images, suggesting a more profound, paradigm-shifting advance is still to come [6, 17]. Nonetheless, the pace of improvement is rapid, with major labs like OpenAI expected to deliver substantially greater capability improvements over the next year . This progress fuels ambitious goals, such as generating entire movies from AI models by the end of 2026 . The evolution will also see a shift in consumer applications, moving from a focus on productivity toward using AI to facilitate deeper human connection and social relationships [11, 16, 18].

Underpinning this rapid evolution are dramatic improvements in the efficiency and cost of AI. Over the next five years, experts project a combined **100x improvement** from hardware becoming 10x more efficient and software adding another 10x in efficiency gains . This will be necessary to meet the forecasted 1000x increase in demand for AI inference over the same period . The real-world impact of these cost reductions is stark; for example, the expense of monitoring all CCTV cameras in the United States with AI is predicted to drop 10x annually, from $30 billion to just $300 million in two years . These economic drivers are critical for making advanced, large-scale computer vision applications commercially viable.

These advancements are expected to unlock novel applications, particularly in scientific research and healthcare. The convergence of AI reasoning and robotics is paving the way for autonomous labs capable of "self-driving science," which could dramatically accelerate R&D in fields like life sciences and materials science [8, 23]. However, there is some disagreement on the timeline; while near-term human-AI collaboration in labs is expected, the vision of fully autonomous, closed-loop systems is considered by some to be a longer-term goal rather than a near-term reality . Looking further ahead, in the medical field, retinal prosthetics are on a path to achieve near-native acuity comparable to 20/20 vision within the next decade, demonstrating the long-term potential of visual AI to directly augment human capabilities [1, 2, 3].