After updating to iOS 17 beta, I was pleasantly surprised to find my dog in the photo album category, and the original “People & Places Humans” became “People, Pets & Places”. Not only my Schnauzer Potato, but also my friend’s Shiba Inu, Fries, who was fostered at my house a while ago, were all precisely in the category. This seemingly trivial function is actually the result of the evolution of AI machine learning, and there are many other similar functions in Apple’s ecosystem that quietly and effectively improve the user experience.
Apple hardly ever mentions the concept of AI, but it’s been integrated into every corner of the Apple ecosystem for a long time, and that’s especially evident in the new versions of Apple products that have come out since WWDC 2023. Apple Input Method, which many people use, got an AI boost in iOS 17, with auto-correction that predicts what the user wants to type and learns from the user’s typing habits to help the user type more efficiently.
Craig Federighi, Apple’s senior vice president of software engineering, said: When you want to type a Ducking word, the keyboard will also learn (your typing habits and predict what you want to type). The improved input experience is due to Apple’s optimization of the Transformer model on the device side, which many people don’t know is one of the key technologies supporting ChatGPT.
The new AirPods Pro’s machine-learning Adaptive Audio Mode automatically switches between noise-canceling and transparent modes when it recognizes a specific external sound, greatly reducing the need for users to manually switch between modes. There’s also iPadOS 17, which uses machine learning models to recognize fields in PDFs, so users can quickly fill in information with names, addresses, and emails from their address book. The much-anticipated journaling app Journal can utilize machine learning technology to intelligently record life moments based on a user’s recent activities, including photos, people, places, physical training, and more. It can also automatically add detailed information to entries such as photos, music, and audio recordings, making it easy to go back to them later.
The much-anticipated Vision Pro creates a digital body for the user, a feature that also uses machine learning technology – an advanced “encoder-decoder” neural network. watchOS 10’s smart overlay also uses AI to decide which information is more important to display at the moment. Even the animations in iOS 17 and iPadOS 17 use AI, with machine learning models that synthesize additional frames of animation to give the device a gorgeous, smooth slow-motion effect, which is why many beta testers have found the animations to be smoother and more elegant.
It can be said that Apple, which the public thought was a “poor student” of AI, is actually an AI “maniac”. It has always been Apple’s AI strategy to publicize machine learning by focusing on “local operation” and “privacy protection”. When ChatGPT was in the wind, Apple CEO Cook said: Artificial Intelligence (AI) has a lot of potential, although Apple has already applied machine learning and AI technology to some of its products, but it is relatively restrained and cautious in specific applications, and there are still a lot of problems to be solved in AI. When asked what he thought of ChatGPT, Apple CEO Cook began by saying that the big language model behind ChatGPT holds “great potential,” but he also expressed his concern that “bias, misinformation, and other issues could be worse in some cases,” and that AI needs to be regulated, and that the companies that develop and use it have a responsibility to monitor themselves.
OpenAI was recently hit with a class-action lawsuit over privacy concerns, with a California law firm claiming in a 157-page lawsuit that OpenAI secretly collected data to train its large-scale language models and accessed huge amounts of data through crawling the web, including from social media sites. Meanwhile OpenAI’s investor, Microsoft, is also named as a defendant, with the lawsuit claiming that tech giants like OpenAI have misused their extremely high technological capabilities in the pursuit of technological advancement without regard to the catastrophic risk to humanity.
That’s the problem with most generative AI products these days, which almost all grab large amounts of Internet content without authorization for algorithmic training, and star companies like Midjourney and Stability AI have been hit with copyright infringement lawsuits one after another. With Apple’s emphasis on privacy and security, it is clear that it will not adopt this controversial approach. Based on a relatively cautious attitude, AI-related functions in Apple’s ecosystem basically rely on local machine learning, and the amount of data required is very small, which, to a certain extent, avoids the sensitive issue of data on the cloud to train AI. This is different from the AI created by server clusters, supercomputers and massive data models represented by ChatGPT, and Apple’s AI appears to be low-key and alternative. However, this does not mean that Apple is not so positive about AI, Apple CEO Cook also made it clear that: we think that the wave of artificial intelligence is very grand, and we will continue to access AI in our products in a very thoughtful way. since the ChatGPT explosion, Apple has obviously increased the recruitment of AI-related positions, which is one of the best proof that Apple is positive about AI.
AI in Apple’s ecosystem doesn’t go to the cloud, so how does it do it? The answer is hidden in Apple’s self-developed chips. At WWDC 2023, the much-anticipated M2 Ultra chip made its official debut, which is manufactured using a second-generation 5-nanometer process and has up to 24 CPU cores, 76 GPU cores, and a 32-core neural engine that can perform 31.6 trillion operations per second. Apple says the M2 Ultra supports up to 192GB of unified memory and can train huge machine learning workloads, such as large Transformer models, in a single system. Again, Transformer is one of the key technologies supporting ChatGPT. Where this chip is powerful is in the memory. Even the most powerful discrete graphics cards today can’t handle the same large Transformer model because they don’t have enough memory, which the M2 Ultra can handle. This means that every device with Apple’s own chip is a device that can train AI locally, and the AI being trained is entirely for the individual user of that device. The benefit of this strategy is that it is possible to deliver AI capabilities to more users in a shorter period of time.
The arithmetic power needed for big models is extremely expensive at the moment, and star AI startup Inflection AI recently raised $1.3 billion, led by NVIDIA, yet Inflection AI turned around and spent $1.1 billion on 22,000 H100 chips from NVIDIA, and it’s amazing how big models are starting to become an asset-heavy industry like coal, oil, and steel. Extrapolating GPT-4’s cost of up to 12 cents per thousand cue words, if ChatGPT were to be deployed to Apple’s more than 2 billion devices around the world, the cost of running it would be astronomical, and it would also significantly raise the cost of endpoint devices.
If OpenAI is centralized, like a brain that everyone can call upon, Apple’s AI is decentralized, like countless cells in everyone’s body, and Apple has been suspected of “falling behind” more than once after the explosion of ChatGPT. Is this really the case? In fact, as early as 2016, Apple spent $200 million to acquire Turi (a startup dedicated to machine learning and artificial intelligence research) to lay out the field of artificial intelligence, gaining Turi’s expertise in developing machine learning tools and platforms.
In 2019, Apple again spent $200 million to acquire Xnor.AI, a company that provides low-power edge-based AI technology for Apple products. As a result, users often had access to AI-driven features such as Siri, album categorization, and search in Apple products in previous years, except that Apple was too ‘high-flying’ to make high-profile announcements like other companies. As an AI company, many people would agree that Apple’s AI technology is not as advanced as its own, but to say that Apple’s AI has no layout and has accomplished nothing would be an understatement.
When the AI represented by ChatGPT started from software, it was tricky to combine it with hardware, what Apple did was to combine AI at the hardware and software level, and also smoothly promoted it to the consumer group, and every user is using it. Although ChatGPT surpassed 100 million monthly users two months after its launch, like many big model products, it still has a lot of barriers and limitations, and the chatbot in the dialog box is not the ultimate interaction method of AI.
Aidan Gomez, CEO of AI company Cohere (and co-author of the famous paper Attention is all you need), said that the capabilities of current AI systems are ultimately limited because not all content is in text form.
Our current models are indeed literally ‘blind’ and that needs to change. At this stage, for the vast majority of people around the world, the iPhone may still be the first place where they feel the wave of AI 2 billion active Apple devices each year is a powerful processor AI training equipment, or users willingly pay money to purchase their own equipment to invisibly train AI, and at present only Apple can do it.
As the previous analysis of Ai Fan Er said, compared with ChatGPT and other big models, Apple AI is more focused on how to improve the end-user experience. This is typical of Apple’s product thinking. In 1997, Steve Jobs was publicly questioned by a programmer at a public conference about not understanding technology, and Jobs gave this answer: don’t get caught up in the so-called technology, you have to start from the customer experience, and then go back to the development of the technology, instead of starting from the technology, and then try to figure out where this technology can be used. I think that’s the right way to do things.