The paper concludes that ReALM outperforms previous approaches and performs as well as the state-of-the-art language model, GPT-4, despite having fewer parameters. It also outperforms GPT-4 for domain-specific user utterances, making ReALM an ideal choice for a practical reference resolution system that can exist on-device without compromising on performance. This development could significantly enhance Siri's usefulness and intelligence in the upcoming iOS 18 and beyond.
Key takeaways:
- Apple's AI team has published a research paper describing a system where Siri can do more than just recognize images, potentially outperforming ChatGPT 4.0.
- The system, called ReALM, takes into account what's on the user's screen and what tasks are active, potentially making Siri smarter and more useful.
- Apple's model has shown large improvements over existing systems, with their smallest model achieving performance comparable to GPT-4, and larger models substantially outperforming it.
- The system can exist on-device without compromising on performance, which is a key factor for Apple. This development could be part of the next few years of platform development, starting with iOS 18 and WWDC 2024 on June 10.