Apple scientists have created a smart computer system called ReALM (Reference Resolution as Language Modeling). It’s designed to make voice assistants like Siri much better at understanding and answering your questions and commands.
In a recent study (via VentureBeat), Apple explains a fresh approach to help big language models handle reference resolution better. This means figuring out unclear references to things on the screen and grasping the context of conversations and backgrounds. With this new method, ReALM could make interactions with devices feel more natural and easy to understand.
Apple has been having some difficulties making its voice assistant, Siri, work well with artificial intelligence. Lately, Siri has been trying out something new: describing pictures sent in Messages while using CarPlay or the announce notifications feature, as mentioned by 9to5Mac.
Despite challenges, Apple is still focused on improving AI. In a recent research paper, Apple’s AI experts talk about a system that does more than just recognize images. Surprisingly, they say one of their models works better than ChatGPT 4.0, a famous big language model.
The paper introduces ReALM (Reference Resolution As Language Modeling), a system made to improve a voice assistant’s abilities. ReALM looks at what’s on the user’s screen and what tasks they’re doing.
It sorts things into three groups: stuff on the screen, stuff talked about, and background stuff.
Apple thinks ReALM could really make Siri smarter and more helpful. Plus, they’re sure it’s fast and easy to use. When they compared it to OpenAI’s ChatGPT 3.5 and ChatGPT 4.0, Apple’s AI on devices did really well.
The paper shows big improvements compared to what’s already out there. Even Apple’s smallest model does as well as ChatGPT 4.0. And the bigger models do even better, beating ChatGPT 4.0 by a lot.
Apple wants their devices to work well without giving up quality. This matches their promise to keep users’ info safe. We might see more about Apple’s AI tech in the new iOS 18 and at WWDC 2024 on June 10.
Good