Apple has reportedly developed an internal service ( APPLEGPT) similar to ChatGPT for employees to test new features, summarize text, and answer questions. Recent research suggests that Large Language Models (LLMs) may run on Apple devices, addressing the challenge of limited DRAM capacity.
Image credit : Apple
Apple’s progress in AI: unveiling an internal service and potential “AppleGPT”
In a strategic move, Apple has reportedly developed an internal service reminiscent of ChatGPT, which will assist its employees in a range of tasks from testing new features to summarizing text and answering questions based on accumulated knowledge. Has been prepared to help.
Table of Contents
July speculations
In July, tech analyst Mark Gurman hinted at Apple’s venture into building its own AI models, the centerpiece of which was a new framework called Ajax.
Ajax, with its versatile capabilities, could potentially pave the way for applications like ChatGPT, informally known as “AppleGPT”.
Recent Research Insights
Recent hints from an Apple research paper titled “LLM in a Flash: Efficient Large Language Model Inference with Limited Memory” suggest the potential integration of large language models (LLM) on Apple devices, including the iPhone and iPad.
Addressing memory constraints
The research paper, discovered by VentureBeat, highlights a key challenge – deploying LLM on devices with limited DRAM capacity.
LLM, which is known for its wide parameters, faces challenges when running on devices with limited DRAM.
Also read : OnePlus 12 and OnePlus 12R to Take Center Stage at ‘Smooth Beyond Belief’ Event in India on January 23
Innovative Solution: Adaptive Loading
Apple’s proposed solution involves on-device execution of LLM by storing model parameters in flash memory and retrieving them into DRAM as needed.
Keyvan Alizadeh, a key contributor and machine learning engineer at Apple, explained that the approach involves an estimation cost model aligned with flash memory characteristics.
Strategic Strategies Employed
The research team applied two main strategies – “windowing” and “row-column bundling.”
“Windowing” focuses on reusing active neurons to reduce data transfer, while “row-column bundling” increases the size of data segments read from flash memory.
Notable Enhancements
The implementation of these technologies saw a significant 4-5x increase in the Apple M1 Max system-on-chip (SoC).
Adaptive loading based on context theoretically opens the door to the execution of LLM on devices with limited memory such as iPhones and iPads.
In conclusion, Apple’s progress in AI, including the development of an internal service and the potential integration of AppleGPT, shows the tech giant’s commitment to advancing its capabilities in artificial intelligence and shaping the future of on-device language processing.
