Although it may not be surprising, the current applications of technology in Cupertino and the constraints of mobile hardware suggest that generative AI will not play a significant role in iOS in the immediate future.
Unlike many companies, Apple has refrained from embracing the recent surge in generative AI hype and has consciously steered clear of using terms like “AI” or “Artificial Intelligence” in its recent keynote presentations. Nevertheless, Apple has always valued machine learning, albeit discreetly, to enhance the overall customer experience.
An illustration of the technologies operating behind the scenes involves Apple’s utilization of AI for image processing. iDevices employ machine learning algorithms to identify and label subjects, perform optical character recognition, and incorporate links when capturing images.
However, this type of inconspicuous AI functionality may not suffice in 2024. Apple’s competitors are advocating for conceptual AI as an essential component for every device and application. According to a recent report by the Financial Times, Apple has been gradually acquiring AI companies and developing its own extensive terminology models to ensure its competitiveness.
Advantages of Apple’s Technology
Apple’s in-house silicon neural processing units (NPUs) drive its existing AI implementations. Since the introduction of the A11 system-on-chip in 2017, Apple has harnessed these accelerators, known as “Neural Engines,” to manage minor machine learning tasks, thereby freeing up the phone’s CPU and GPU for other functions.
The performance of Apple’s NPUs is particularly impressive. The A17 Pro in the iPhone 15 Pro can deliver 35 TOPS, which is twice as fast as its predecessor and approximately double the capability of certain PC-compatible NPUs from Intel and AMD.
Qualcomm’s latest chipset cards are comparable to Apple’s NPU performance. Qualcomm, like Apple, boasts extensive experience in NPUs for wireless devices, while AMD and Intel are relatively new players in this field.
Apple has highlighted its ability to run demanding games like Resident Evil 4 Remake and Assassin’s Creed Mirage, indicating that the platform’s computational power is not a limiting factor for handling larger AI models.
Moreover, Apple’s M-series silicon, utilized in both Mac and mobile devices, has demonstrated exceptional proficiency in executing AI inference workloads. Even a three-year-old M1 MacBook Air could efficiently run Llama 2 7B at 8-bit efficiency, with improved performance using a 4-bit normalized version, provided there is ample memory—less than 16GB may lead to performance issues. You can test this on your M1 Mac using Llama, a tool for running AI models.
In terms of storage, Apple may face challenges due to the storage requirements of AI models, typically needing about a gigabyte of storage for every billion guidelines when operating at 8-bit efficiency. To address this, Apple could develop smaller, quantized versions or reduce efficiency to Int-4, halving the storage needs.
Despite these considerations, it is anticipated that Apple’s upcoming devices may necessitate more storage or require more compact and focused AI model versions. Apple’s decision to design its own solutions rather than adopting existing models like Stable Diffusion or Llama 2 may be influenced by factors such as these.
Furthermore, recent reports suggest that Apple may have devised a solution to the RAM issue by utilizing flash memory to facilitate on-device movement of LLMs.
Embracing a Traditional AI Approach
When Apple eventually introduces AI features on its desktop and mobile platforms, it is expected to adopt a relatively conservative approach.
One potential area for improvement is enhancing Siri to operate in a manner that eliminates the need to interact with it as if addressing a child. Transforming insights into a format that Siri can better comprehend, possibly by leveraging a Language Model (LLM), could enhance its responsiveness to user queries.
By phrasing questions indirectly, users may reduce the likelihood of misinterpretation by Siri, leading to more accurate responses.
This approach offers several potential advantages. Firstly, Apple may be able to utilize a significantly smaller concept compared to models like Llama 2. Secondly, the issue of generating inaccurate responses with LLMs could be mitigated.
Apple has a track record of delaying the adoption of cutting-edge technologies, instead focusing on refining features until they provide tangible benefits. While generative AI has yet to demonstrate significant success, Apple’s strategic approach has positioned it as the leading smartphone vendor in 2024, contrasting with Microsoft’s efforts to leverage AI for Bing without substantial market share growth.
In conclusion, Apple’s cautious and methodical approach to integrating AI features may prove to be a successful strategy in the long run.
