Apple prepares another AI model to run on iPhone

Apple prepares another AI model to run on iPhone


Apple researchers presented the ReaLM model, an artificial intelligence model capable of running locally on devices without requiring much hardware power

Researchers from Apple has developed an artificial intelligence model called ReALM, which can run locally on devices and absorb the context of conversations to answer questions. According to internal tests, the technology would be comparable to LLM OpenAI’s GPT-4 to perform some quick tasks. The novelty is added to the MM1, another artificial intelligence model developed by Maçã and revealed in March this year.



The results were published in a scientific article signed by seven researchers linked to Apple. The name ReALM stands for Reference Resolution As Language Modeling and focuses on the power of artificial intelligence in analyzing diverse data, such as data contained on screens that cannot easily be reduced to text.

The AI ​​engine is divided into three main entities:

  • On screen: information displayed on the device screen;
  • Conversational: Tasks based on what the user asked the AI ​​during a conversation. The data can arise through the initiative of the information requester or through a response from the AI ​​assistant.
  • Background: Information about activities that occur in the background of the device and are not part of the home screen or conversation, such as a request to stop an alarm or playing music in an application.

This way, the model would be able to interpret natural language requests: for example, the article uses an address list and says that the user can ask to “call the second to last option” or “call the one at a given street” . Despite not being mentioned in the study, the Siri assistant appears to be the biggest beneficiary of the new development, as the AI ​​engine could be added to the system.

Comparison with GPT-4

Tests conducted with GPT-3.5 (used in the free version of ChatGPT) and GPT-4 (the latest model from OpenAI, available on ChatGPT Plus) reveal that even the lightest versions of ReaLM outperform version 3.5 and can compare with GPT-4, outperforming the technology in specific contexts, such as those specific to the Actions domain.

Therefore, the researchers conclude that ReaLM would be an option for a system “that can exist on the device without compromising performance.” It’s another sign that the technology could be used with Siri on iOS.




Apple's new AI model could power Siri (Image: Play/Apple)

AI models for mobile phones face a common performance challenge: some can perform some tasks locally on the device, but more complex tasks still require a connection to a server via the cloud.

Concentrating everything on the device is better to guarantee the privacy of conversations, but there are performance problems: the premise is that Apple’s model solves this problem.

Apple and artificial intelligence

After a quiet 2023, Apple is expected to enter the generative AI race in 2024. The expectation is that the iOS18the next generation of operating system of the iPhone, it has a strong focus on artificial intelligence and a Siri revamped to function like ChatGPT or Gemini.

The news on artificial intelligence for iPhone and other products from the Cupertino giant should be announced during the WWDC conference, scheduled for June 10th. It is worth mentioning that the CEO Tim Cook promised that the The company’s artificial intelligence will ‘break new ground’ and also that Apple negotiates with companies like Google, OpenAI AND Baidu for possible third-party AI integration into iOS.

You can consult the original article on ReaLM (in English) at arxiv.org.

Trends on Canaltech:

Source: Terra

You may also like