Apple in talks with PrismML, which shrinks AI models up to 15x to run on iPhones (www.cnbc.com)

Apple is in talks with startup PrismML, which has developed a technology that can compress AI models by up to 15 times, making it feasible to run them directly on iPhones while significantly reducing memory usage. This breakthrough could enhance the speed and privacy of Apple's Siri by enabling more on-device processing, thereby eliminating the delays and privacy concerns associated with cloud-based AI. PrismML's technology, which simplifies model storage from 16 bits to fewer than four possible values, has already successfully compressed Alibaba's Qwen model from 54 GB to under 4 GB, making it operable on the latest iPhone models. The implications of this technology could be transformative for the AI and ML landscape. If effective, it would lower the demand for large memory chips and datacenter infrastructure, allowing previously cloud-dependent tasks to run locally on devices. However, analysts caution that while PrismML’s technology might reduce some computing demands, it won’t eliminate the need for powerful chips altogether. Rather, it could shift more processing power to personal devices, maintaining Apple's privacy-centric approach while potentially sparking new applications and usage patterns in AI. The testing of this technology's real-world performance will be critical to validating its claimed efficiencies and effectiveness.