Apple in talks with PrismML on on device AI compression

Apple is in early discussions with PrismML, a startup developing technology that dramatically compresses large artificial intelligence models to run directly on iPhones, according to reports. The talks, which are described as preliminary, could influence how the company divides future AI work between its devices and cloud servers.

PrismML CEO Babak Hassibi confirmed that Apple is evaluating the startup's models, testing their speed, power consumption, and performance on devices. He characterized the discussions as progressing nicely but emphasized that they remain at a very early stage, with no partnership or licensing agreement announced. Apple declined to comment.

The technology in question centers on a compression technique that simplifies how an AI model stores its internal information. PrismML recently demonstrated this by shrinking a 27-billion-parameter open-source model from roughly 54 gigabytes to less than 4 gigabytes. The company claims this reduction allows the model to fit within the memory available to an app on a high-end iPhone, such as an iPhone 15 Pro. According to PrismML, the compressed versions use 10 to 15 times less memory, generate responses six to eight...