PrismML brings its tiny LLMs to Qualcomm-powered smart glasses

Your smart glasses can now answer questions about what you see in real time without relying on cloud processing.

PrismML, a startup founded by Caltech researchers and advised by UC Berkeley's Ion Stoica, has developed a compact language model specifically for smart glasses powered by Qualcomm's Snapdragon chips. At Qualcomm's Snapdragon Summit, the company demonstrated PrismML's 1-bit Bonsai LLM running locally on smart glasses built with the Snapdragon AR1 Gen 1 Platform. This version of the model is a 2-billion-parameter AI tuned for vision and language, enabling wearers to receive real-time information about their surroundings by asking questions. PrismML specializes in significantly shrinking larger AI models, achieving a 4x reduction in this case while maintaining nearly equivalent performance on standard benchmarks. The startup's overarching objective is to promote open-weight AI that can operate directly on devices, efficiently utilizing existing computing power and offering an alternative to proprietary AI solutions that demand substantial computational resources and raise privacy concerns. Although a model for Qualcomm's chips has been developed, no specific smart glasses incorporating PrismML have been announced for market release.

The integration of PrismML's compact LLMs into Qualcomm-powered smart glasses signifies a significant step towards on-device AI processing for augmented reality devices. This development directly addresses the growing demand for privacy and reduced latency in wearable technology, as running models locally eliminates the need to transmit sensitive visual and audio data to external servers. Such advancements could accelerate the adoption of smart glasses for a wider range of applications, from personal assistance to enterprise solutions, by enhancing their responsiveness and user experience.

This collaboration between PrismML and Qualcomm positions them to capitalize on the burgeoning market for AI-enabled wearables. By enabling sophisticated AI functions to run directly on the hardware, they are reducing reliance on cloud infrastructure, which can be costly and less reliable for real-time interactions. The focus on privacy and efficiency with open-weight models may also attract developers and businesses looking for customizable and secure AI solutions, potentially setting a new standard for AI capabilities in consumer electronics.