Apple AI Amplification! Claims its device-side model outperforms GPT-4

According to media reports, in a recent paper, Apple's research team claimed that they have proposed a model ReALM that can run on the device side, and this model can surpass GPT-4 in some aspects.

The parametric quantities of ReALM are 80M, 250M, 1B and 3B, all of which are very small and suitable for running on the device side of cell phones and tablets.

ReALM investigates mainly, the process of getting AI to recognize the referential relationships between various entities (e.g., names of people, places, organizations, etc.) mentioned in a text.

The paper categorizes entities into three types:

On-screen Entities (On-screen Entities): refers to what is currently displayed on the user's screen.

Conversational Entities: This refers to the content associated with a conversation. For example, if the user says "call mom", then the contact information of mom is a conversational entity.

Background Entities: Entities that may not be directly related to what the user is doing or what is displayed on the screen, such as the music that is playing or the alarm that is about to go off.

The paper says that while large language models have proven to be extremely capable on a wide range of tasks, their potential has not been fully utilized when used to solve the problem of referring to non-conversational entities (e.g., screen entities, background entities).

ReALM, on the other hand, is a novel approach, and by comparing its performance to GPT-3.5 and GPT-4, it was shown that the smallest model performs comparably to GPT-4, while the larger model significantly outperforms GPT-4.

This research is expected to be used to improve the Siri assistant on Apple devices, helping Siri better understand and process context in user queries.