Unexpectedly, Apple has emerged from the AI race’s shadows with two ground-breaking research papers that allude to a game-changing technology that has the potential to completely transform the iPhone experience. Although some would have thought Apple was lagging behind in the field of newly developed generative AI, recent reports paint a different picture.
The crux of this revelation lies in Apple’s pursuit of a technology capable of executing powerful artificial intelligence (AI) models directly on iPhones and other Apple devices, without sacrificing speed or performance. The innovation at play involves utilizing flash memory, the very storage medium for apps and photos, to run large language models (LLMs) responsible for comprehending and generating natural language.
LLMs serve as the backbone for AI chatbots like Apple’s Siri, offering human-like interactions and valuable information. Yet, the memory demands of LLMs, requiring substantial gigabytes of RAM, present a challenge for devices like iPhones with limited RAM capacity. Apple’s solution involves a novel technique outlined in their research paper titled “LLM in a flash: Efficient Large Language Model Inference with Limited Memory.”
Key to this innovation is the concept of windowing, a recycling technique that minimizes the data transfer between flash memory and RAM by reusing already processed data. Complementing this is ‘Row-Column Bundling,’ a reading technique that enhances data retrieval speed from flash memory. Apple researchers claim their approach can run LLMs up to twice the size of available RAM, offering impressive speed-ups on both CPUs and GPUs.
This discovery has far-reaching implications that go well beyond simple technological improvements. With Apple’s new strategy, future iPhones may be able to run sophisticated AI models locally, opening up a world of possibilities that do not depend on cloud servers. This might boost Siri’s powers, enabling more intelligent responses, better speech recognition, and real-time language translation.
In parallel, Apple’s foray into generative AI doesn’t stop here. Another research paper introduces HUGS (Human Gaussian Splats), a method to generate animated 3D avatars from short monocular videos. This opens doors to applications like virtual try-on, telepresence, and synthetic media, hinting at a future where users can create novel 3D scenes using their iPhone camera.
Beyond these research papers, Apple’s larger strategy involves integrating generative AI into Siri and other apps. Reports suggest Apple is developing its own formidable generative AI model named “Ajax,” designed to rival the most advanced models globally. With a projected timeline for generative AI features on iPhones and iPads around late 2024, Apple’s grand vision of a smarter, more interactive ecosystem seems poised to become a reality, ushering in a new era for AI on Apple devices.