A new sensor developed by researchers in China has the potential to significantly cut the energy needed for visual imaging by processing light directly at the point of detection. In a study published on August 19 in the journal Nature Sensors, scientists described a novel 2D chip called "LightTok," which can convert raw light into data tokens that artificial intelligence (AI) models can use without needing to move the data elsewhere. This design eliminates several energy-heavy steps that traditional sensor systems typically rely on. The concept behind LightTok, according to Miao Feng, director of Nanjing University's Institute of Brain-Inspired Intelligence, was to integrate the creation of data tokens directly into the sensor. This allows the chip to generate tokens that AI models can process as soon as light is detected. These tokens contain all the essential image information, bypassing the need for extensive data transfer and processing. In conventional visual systems, light signals go through multiple steps. A sensor captures the light, which is then converted into digital pixels by an analog-to-digital converter. This data is stored temporarily before being sent to a separate chip, where the image is divided into sections, like tiles in a grid. Each section is then transformed into a token for AI processing. One widely referenced study found that the analog-to-digital converter alone accounts for about 66% of an image sensor’s energy use. Moving processing to the cloud can further increase energy consumption. LightTok simplifies this process by collapsing five steps into one, integrating sensing, memory, and computation within the same pixel. The researchers achieved this using an array of single-layer molybdenum disulfide floating-gate phototransistors, a technology that can detect light, store that information, and perform calculations. Liang Shi-Jun, a physics professor at Nanjing University, described the chip to Xinhua, a Chinese state-run news agency, as a design that physically eliminates unnecessary data movement, the main cause of energy waste. "Light comes in, tokens come out," he said, explaining the name "LightTok." In tests, LightTok achieved an 87.3% accuracy rate in image recognition, while using only 10% of the energy required by traditional methods. However, its current maximum resolution is limited to a 32 by 32 grid of photosensitive pixels, which is much lower than the quality offered by modern smartphone cameras or drones. Miao Feng noted that the technology could be scaled using the complementary metal-oxide-semiconductor (CMOS) manufacturing process, which is used to make chips in smartphones, laptops, and drones. If successfully scaled, LightTok could revolutionize remote sensing technologies. For example, drones scanning disaster zones or remote areas might be able to operate longer due to reduced energy use in visual processing. Kumar Sokka, CEO of Acre Security, a company focused on real-world sensing for critical infrastructure, called the research a "small-scale demonstration" but emphasized its potential importance for physical AI. He noted that while the innovation is promising, it is not a complete solution to the energy challenges of converting sensor data into usable forms for AI models.