Apple ( ) continued its slow and steady approach to integrating AI into devices such as the iPhone, Mac and Apple Watch on Monday, announcing a host of new features at WWDC. The company also launched the Foundation Models Framework, which allows developers to write code using Apple’s AI models.
Live Translation was one of the more buzzworthy AI announcements made at the event. This feature translates FaceTime and phone calls from one language into another in real-time. Apple also demonstrated Workout Buddy, a voice assistant powered by AI that provides words of encouragement and useful information during exercise. In a video demo, Workout Buddy told the jogging woman that this was her second run of the week. Apple announced an update to Visual Intelligence – a tool which uses AI to interpret images taken by a device’s cameras. The new version also uses screenshots to identify products or summarize webpages. Apple showed off updates to Genmoji, and Image Playground – two tools that create stylized images using AI. It also demonstrated how AI can be used to automate tasks, create text, summarize emails and edit photos.
These incremental announcements did not do much to dispel the idea that Apple is still playing catch-up in AI. The company has not yet developed a model that can compete with the best offerings from OpenAI, Meta or Google and still sends some challenging queries to ChatGPT.
According to some analysts, Apple’s incremental approach to AI is justified.
According to Paolo Pescatore of PP Foresight, “the jury is still out” on whether users will gravitate towards a phone that has AI-driven features. Pescatore says that Apple must strike a balance between bringing something new and not annoying its core users. It comes down to whether AI is driving revenue increases.
Francisco Jeronimo is an analyst at IDC and says that Apple making its AI models available to developers is crucial because of the company’s large reach with coders. Jeronimo stated that it “brings Apple close to the kind AI tools that competitors like OpenAI, Google and Meta have offered for some time.” Apple’s AI models are not the most powerful, but they run on personal devices, which means that they don’t require a network connection. They also do not incur the fees associated with models from OpenAI or others. Private Cloud Compute, a service offered by the company to developers, allows them to use cloud models while keeping their data private. Apple will need to make bigger leaps in the future with its AI use, as its competitors are exploring the possibilities of the technology.
Google and OpenAI both showed off futuristic AI assistants that can see the world and talk in real-time through a device’s camera. OpenAI announced last month that it would be acquiring a company founded by the legendary Apple designer Jony Ive in order to develop AI-infused hardware.
Apple is still behind in building advanced AI but it is publishing AI research at an increasing rate. A paper published a few days prior to WWDC highlights significant flaws with today’s advanced AI models. This is a convenient conclusion if you’re still getting up to date.
This paper shows that the latest OpenAI models and others, who use a simulated reasoning to solve complex problems, fail when they reach a certain complexity. Apple researchers tested various models on increasingly complex versions of the Tower of Hanoi and found they were successful up to a certain point, but then failed dramatically. Subbarao Kambhampati of Arizona State University, who has published similar research on the limits of reasoning model, says Apple’s study reinforces the idea simulated reasoning approaches need to be improved to tackle a broader range of problems. Kambhampati says that reasoning models are “very useful, but they have important limits.”
But despite the fact that the research suggests a more cautious approach towards AI, Kambhampati doesn’t believe Apple is complacent. “If you know the inner workings of Apple, they are still very enthusiastic about LLMs,” Says Kambhampati.
