Apple's large language model team may be small, with 16 people, but it has a big budget: millions of dollars reportedly go into AI training every day.
This is reported by The Information, citing a person with direct knowledge of Apple's AI development. The so-called "Foundational Models" team is led by AI engineer John Giannandrea, who Apple poached from Google in the spring of 2018 and promoted to the board the following winter.
The team is reportedly focused on conversational AI, similar to ChatGPT. The previously leaked LLM "Ajax GPT" is said to be more powerful than OpenAI's GPT-3.5, with more than 200 billion parameters.
But that would still be significantly less than GPT-4's 1.8 trillion parameters, even if those parameters are distributed across 16 networked expert models. According to the report, Ajax GPT was developed for internal use only.
For end users, the benefits of LLM research will come in the form of more accurate voice commands for the iPhone. The Siri team plans to implement LLM-based voice commands that can be much more complex than the predefined voice commands.
For example, Apple is planning a voice feature that will let you send GIFs by voice command through Apple's Shortcut app, which can be used to program app sequences. This feature is expected to arrive with a new version of iOS next year.
Apple explores all areas of generative AI
In addition to the LLM team, other teams are working on computer vision, a core technology for Apple's Vision Pro, and generative AI for images, video, and 3D scenes. In the past, Apple researchers have demonstrated GAUDI, a generative AI for text-to-3D indoor scenes.
Apple has "at least two other relatively new teams" developing language and image models, according to The Information. The company is also allegedly working on multimodal models for text, images, and video. However, this is a long-term research project.
