New Phone AI Tells Blind Users What's Around Them

About 1 billion people worldwide live with some form of vision loss. Today's AI helpers — the kind that describe what a camera sees — tend to be either huge and accurate but stuck in the cloud, or small and fast but vague. A vague description is worse than useless when you're crossing a street. Researchers at several universities set out to fix that gap with a model called Smol-VL-BLV.

The trick was in the training, not the size. The team used a small AI model with 500 million internal settings — tiny by today's standards — and taught it in two steps. First, a much larger AI acted as a coach, feeding it high-quality descriptions. Second, they rewarded the model every time its narration included the things that actually matter for getting around: words like "left" and "two metres ahead," plus warnings about obstacles. They also added a final polish stage so the model didn't forget how to describe ordinary scenes while learning to be a navigator.

The results were strong. Compared with the starting point, its score for describing spatial layout rose 19.3%, and its social-awareness score rose 14.8%. On reading signs and labels, it more than doubled its score, and its accuracy at answering questions about text in photos rose 44.2%. Crucially, it all fits on a phone: about 450 megabytes, running on the device with no network connection at all.

The catch is that this is a research paper, not an app you can download today. Its speed depends on how powerful your phone is, and an AI that guides someone through traffic has to be right essentially every time — a single confident mistake could be dangerous. Still, the direction is clear: helpful AI is getting small enough to live in your pocket instead of a data centre.