Apple appeared to lag behind in embracing generative AI technology initially, but recent advancements in contextual understanding research could potentially elevate Siri’s performance beyond that of ChatGPT.
During the rapid proliferation of ChatGPT and the surge of generative AI innovations from major tech players like Google, Microsoft, and Meta, Apple maintained a notably subdued presence. However, Apple’s researchers have introduced a novel model that holds the promise of enhancing Siri with generative AI capabilities, meeting the expectations of Apple enthusiasts.
The newly developed model, ReALM (Reference Resolution As Language Modeling), addresses the challenge faced by large language models (LLMs) in grasping contextual nuances in various scenarios such as on-screen displays, conversational contexts, and background references like applications or features running in the background. The ultimate objective is to achieve a seamless and intuitive voice assistant experience.
According to the researchers, human speech often includes ambiguous references like ‘they’ or ‘that,’ whose meanings are typically evident within the context to human listeners. Through training on diverse datasets, ReALM has demonstrated superior performance compared to GPT-3.5 and GPT-4, the underlying engines of ChatGPT, across all contextual evaluation metrics.
Enhancements Brought by ReALM to Siri:
-
On-screen Contextual Insights:
- ReALM was trained on “on-screen” data extracted from web content, encompassing details such as contact information. This training enables the model to interpret text within images (e.g., addresses and financial details), providing Siri with a better understanding of on-screen content crucial for Apple users seeking assistance.
-
Comprehensive Conversational and Background Comprehension:
- By leveraging datasets containing business listings, ReALM can decipher implicit references in conversations, such as instructions like “call the bottom one” regarding a list of nearby pharmacies displayed on the screen. Additionally, the model excels at recognizing “background entities,” encompassing elements like background music or active alarms on the device.
-
On-device Functionality:
- Notably, ReALM is designed to operate entirely on-device, a significant departure from conventional LLMs that heavily rely on cloud infrastructure due to high computational demands. This on-device approach aligns with Apple’s emphasis on privacy and could represent a substantial advancement in AI capabilities for Apple devices.
Apple has maintained a veil of secrecy around its AI strategies, with CEO Tim Cook hinting at a significant AI-related announcement expected later this year. The upcoming Apple Worldwide Developers Conference (WWDC) on June 10 has become a focal point for anticipating potential revelations in this domain.
