Apple’s AI Future Pt. 2

As predicted in my previous blog post, Apple is now sharpening its focus on AI, emphasizing privacy and on-device processing.

A few months ago, Apple held its highly anticipated 2024 WWDC. The highlight? The introduction of Apple Intelligence, a personal intelligence system poised to redefine user interaction across the Apple ecosystem.

Apple’s approach to AI is unique. Unlike tech giants like Google and OpenAI, Apple doesn’t have a frontier model. Instead, they’ve perfected the application of specialized, smaller models tailored to specific tasks. These private models operate on-device, ensuring user data remains secure while delivering high performance.

At the heart of Apple Intelligence are two primary models: a ~3 billion parameter on-device language model and a larger server-based model available through Private Cloud Compute. These models handle tasks ranging from text summarization to image creation, adapting dynamically to user needs through a system of adapters. Adapters are small neural network modules that fine-tune the models for specific tasks without altering the base model's parameters.

Private Cloud Compute

Beyond on-device models, Apple is launching Private Cloud Compute to balance on-device processing with the computational heft of server-based models. This setup ensures privacy while providing the computational capacity for more demanding tasks. The infrastructure allows these server models to run on dedicated Apple silicon servers.

Benchmarks show that Apple’s on-device models outperform several larger models like Phi-3-mini and Mistral-7B in tasks such as summarization and natural language understanding. The server models are benchmarked to compete against GPT-3.5-Turbo, offering a balance of power and efficiency. However, GPT-3.5-Turbo is already losing relevance.

Personalized User Experience

Apple Intelligence introduces a semantic search feature across all devices. This feature leverages natural language processing to deliver relevant search results, transforming how users interact with their devices. By understanding the context and intent behind queries, Apple’s semantic search personalizes the user experience, making devices more intuitive and responsive to individual needs.

This creates a strong moat against model-producing companies like OpenAI that don’t have access to the context of the user's data.

Privacy Concerns

With access to user data context, privacy becomes a concern. Apple’s approach ensures that user data is processed on-device or through encrypted channels in Private Cloud Compute. This system ensures that user data is never exposed or retained, addressing fears of mass surveillance and security breaches. Privacy isn’t just a feature for Apple; it’s a cornerstone of their AI development strategy.

Strategic Integration

Apple's strategy has always been about perfecting applications rather than pioneering them. Their integration of ChatGPT into iOS, iPadOS, and macOS exemplifies this approach. By partnering with OpenAI, Apple enables users to access powerful AI capabilities without leaving their native apps. This move mirrors Apple’s deal with Google, where Google pays $15 billion annually to be the default search engine on iOS devices.

This integration allows Apple to leverage the best AI models without developing a frontier model themselves. It positions Apple as an integrator platform, potentially paving the way for a specialized AI marketplace within the Apple ecosystem. This strategy could foster a competitive environment where AI companies strive for the best integration, ultimately benefiting Apple’s vast user base.

On-Device Actions Models

The most exciting development from WWDC is Apple’s on-device actions models. These models represent the first steps towards smart agents capable of performing tasks on behalf of users. Imagine a future where your phone doesn’t just respond to commands but anticipates your needs, much like the AI assistant in the movie "Her." This vision transforms the smartphone from a tool into a personal companion, seamlessly integrating into every aspect of your daily life. For example, the on-device model with context knowledge could plan a birthday party, setting reservations, sending calendar invites, and messaging attendees.

Some predictions

• Apple is uniquely positioned to provide context-aware, life-integrated models, making AI more like a personal assistant than a prompt-heavy tool. Expect Apple to focus heavily on on-device processing to emphasize their privacy advantage. This not only enhances privacy but also reduces costs by running models on client devices instead of servers.
• Look for upgrades to more pro-tier clients with heavy-duty on-device GPUs to unlock more AI functionalities.
• AI Agents are gaining traction among startups, and Apple, with its deep OS integration, is uniquely positioned to build an action AI that can learn and execute tasks on your behalf. This could be VERY disruptive, but that’s a topic for another blog post.