ChatGPT-4 may be the best thing that could happen to Apple

In the current AI hype, Apple is the only player that seems strangely reticent, while Google, Meta, and Microsoft are beating the drum for their AI-boosted products weekly. And yet, Apple may be the winner of AI’s breakthrough based on Large Language Models (LLMs). Here’s what the Playbook could look like.

Apple has redefined the personal computer thrice: with the Apple II, it invented the category in the first place (1977). The user-friendly internet in the gestalt of the WWW was developed on Steve Jobs’s NeXT computers in the early 1990s, and in 2007 with the iPhone, the PC became an everyday companion. Apple evolved into the most valuable company in the world.

With LLMs, Apple could reinvent the idea of the PC for the fourth time and truly make it a personal computer. But not as a Copilot equipped with generic world knowledge like Microsoft. Apple has the chance to create a digital alter ego of every human being. The core here is a personal LLM that is trained by all of its user’s interactions with an Apple device; everything you

• write and read (emails, chats, search, social, texts)
• say and hear (phone & video calls), and
• experience (on the screen, photos, videos, microphone).

Even if this reads quite spooky in the first (OK, and second) moment, the personal benefit of such a trained omniscient companion would undoubtedly be extremely high. It would know its user extremely well (maybe better than the user knows himself — but that’s another discussion) and would be perfect assistance. The main argument against such a trained LLM is Privacy. Who would want to store their entire life (with all their intimate secrets) in the cloud and leave it to third parties for monetization or leaks?

But that is precisely Apple’s unfair competitive advantage. Only Apple could integrate large language models into its ecosystem like this without compromising privacy:

1. With Apple Silicon (MX processors), Apple brings everything it needs to implement LLMs so that most processing occurs on the personal device. Hence, no training data must be transferred to the cloud (neural engine, GPUs, frameworks).
2. Apple would use federated learning techniques to share model updates without exposing raw user data for model training and improvement. By aggregating these updates from many devices, the LLM could be continuously refined while protecting user privacy.
3. Unlike the other Big Techs, Apple has put privacy consistency at the center of its communications and doing in recent years. Apple has an almost unassailable trust advantage: it’s the privacy brand today.
4. Apple can integrate its AI services not only into existing services (Messages, Mail, Apps) but also across the entire hardware (iPhone, iPad, Mac, AirPods, Watch, AppleTV, Glasses (?)) in a privacy-compliant manner.
5. Apple does not need monetization in its business model of personal data like Google, Meta, and meanwhile also Amazon.

To connect Apple’s ecosystem with this new egosystem, the question remains whether on-device processing of LLMs is possible. Don’t models like GPT-X with billions of parameters require massive cloud infrastructures for this? Yes and no. There are practicable ways to solve this challenge:

1. Model compression: techniques such as pruning, quantization, and knowledge distillation can be used to reduce the size of LLMs without significant performance degradation. These compressed models are better suited for on-device processing, although they may not be as powerful as their larger counterparts.
2. FlashAttention: is a new IO-aware exact attention algorithm that speeds up the training of Transformers and reduces its memory footprint — without any approximation. Transformer for actions approaches (like adept.ai) follow this path.
3. Model partitioning: dividing the LLM into smaller sub-models that work together in a distributed manner can enable parallel processing and reduce the overall computational load on the device.
4. Prioritization of AI components in Apple’s Silicon Roadmap (neural engine, GPUs, RAM).
5. Adaptive models: developing models that adapt their complexity to the available resources can help balance performance and resource usage. For example, if a device has limited computing power, the model could switch to a simpler version to ensure smooth operation.

The recent leak of Meta’s LLaMa model showed that LLMs might perform very well on local devices. Model partitioning or hybrid approaches could make personally trained LLMs in conjunction with generic cloud-based LLMs an unbeatable tandem.

Apple’s advantage is obvious: the lock-in via personal LLMs on Apple devices would go over the entire lifetime of the user and diffuse into all conceivable hardware. Maybe, Apple will become the first 10-trillion-dollar company after all.

Disclaimer: I assume Apple (like the rest of the world) was surprised by the success of OpenAI’s ChatGPT. And, of course, I have no — zero! — an idea about what is being worked on in Apple Park and the Apple offices scattered all over Cupertino. So everything written down here is my ChatGPT4-like hallucination.