On June 10th, local time, Apple's annual Worldwide Developers Conference (WWDC) was held. This time, Apple finally brought the long-rumored latest artificial intelligence (AI) strategy. As the missing piece in Apple's current growth puzzle, its AI progress has been widely watched by the market.
As the outside world widely speculates on how Apple will reach a cooperation with OpenAI, Apple's attitude towards OpenAI is very subtle. As the "hottest" in the AI technology circle, OpenAI and ChatGPT were "ignored" at this conference.
A reporter from the Daily Economic News noticed that although OpenAI CEO Sam Altman appeared at Apple's headquarters that day, he did not make an official appearance. This was undoubtedly in stark contrast to Microsoft's 2024Build conference last month. At that time, OpenAI became one of the undisputed protagonists, and Altman was also invited to the stage to make a high-profile speech.
At the same time, Apple Intelligence was pushed to the center of the stage and was mentioned about 60 times at the WWDC conference. Apple emphasized the capabilities of Apple Intelligence and local AI models from beginning to end, and only mentioned artificial intelligence 3 times.
Altman's low profile and the high profile of "Apple intelligence" have raised questions from the outside world: What exactly is "Apple intelligence"? And why has this star in the field of AI been reduced to a "supporting role" on the Apple stage?
Apple Smart Analysis: Three-layer Technical Architecture
Let’s answer the first question first: What is Apple Smart?
Earlier reports said that Apple's smart phone is a shelled GPT-4o. However, a blog post released by Apple shows that the basic generative model behind Apple's smart phone is a model developed by Apple itself.
According to reports, Apple's smart assistant is designed to be a highly personalized and privacy-protecting smart assistant, deeply integrated into various operating systems such as iOS, macOS and visionOS, providing a wide range of functions from major upgrades to Siri to writing and communication tools. The technology behind it is divided into three layers.
The first layer is Apple's local model, which includes a language model and a diffusion model. The language model is a small model with only 3B parameters, which can be used on small terminals such as iPhones and provide good performance. The model has been fine-tuned for a range of tasks, such as writing and perfecting text, determining the priority and summary of notifications, creating interesting images for conversations with family and friends, and taking in-app actions to simplify cross-app interactions. There is an adapter on top of these fine-tuned model weights that can select the most appropriate weights based on the task.
Local model fine-tuning process Image source: Apple official website
The second layer is a language model based on a larger server, and the model function is available through private cloud computing. According to Apple's official website, private cloud computing is for advanced functions that require the use of larger basic models to reason about complex data.
For each data inference request, the user's device will establish an end-to-end encrypted connection with the private cloud computing cluster. Only the selected cluster can decrypt the request data, and the data will not be retained after the response is returned, and Apple cannot access it. The company claims that this is the most advanced security architecture ever deployed for large-scale cloud AI computing.
It is reported that all of Apple's basic models are trained based on its AXLearn framework, an open source project released by the company in 2023.
Apple basic model training process Image source: Apple official website
The third layer is third-party large language models. For example, as part of Apple's intelligence, Apple's Siri and writing tools integrate OpenAI's ChatGPT. This ChatGPT integration powered by GPT-4o will be launched later this year and users can access it for free without creating an account.
In addition, Apple also mentioned in its official introduction that it will support more models to access Siri in the future. According to reports, Apple and Google are in negotiations, which may be good news for Google Gemini.
Apple's ambition: starting its own AI journey
Since foreign media broke the news that Apple has joined hands with OpenAI, how ChatGPT will be integrated into the iPhone has been widely anticipated. However, judging from the press conference, Apple's attitude towards OpenAI is actually very subtle.
Foreign media technology reporter Gurman forwarded a photo of Altman at Apple headquarters. Image source: X platform
First of all, Altman has no presence. Although he was photographed at Apple headquarters, he did not make an official appearance at the press conference, and later only expressed his pleasure to cooperate with Apple on the X platform, in contrast to his high-profile appearance at Microsoft's 2024 Build conference.
Image source: X Platform
Secondly, the "explosive" scenes of ChatGPT connecting to Siri that everyone expected did not appear. ChatGPT based on GPT-4o was only introduced as a supplementary function of Apple's intelligence. Apple has always emphasized the capabilities of Apple's intelligence and local AI models, and ChatGPT has only become a "supplementary function."
After the press conference, the rumor that ChatGPT might be used to replace Siri was refuted. Judging from the fact that Apple's stock price fell by nearly 2% after the press conference, the capital market seemed to be caught off guard and was not satisfied with Apple's intelligence.
However, a day later, Apple's stock price soared. As of the close of June 11 local time, Apple rose by more than 7%, the largest increase since November 10, 2022, breaking the record closing high set on December 14, 2023, and leading the S&P component stocks for the first time since November 30, 2021. The market seems to have reacted: Apple has greater ambitions.
Image source: Google Finance
Some analysts believe that ChatGPT is only a stopgap measure, and Apple may have many considerations.
The first is the issue of privacy. If AI is to maximize its capabilities on hardware devices, it must obtain all the information on the device. This kind of important data must be kept locally, which is why Apple emphasizes local models. At the press conference, Apple repeatedly reiterated its privacy protection measures and launched a new privacy security mode. The company said that Apple devices will ask users for permission before sending user data to OpenAI servers.
Secondly, Apple does not want to be held hostage by a single model and become a data pipeline. According to foreign media reports, in a meeting after the keynote speech, Apple's senior vice president of software engineering Craig Federighi hinted that Google's Gemini artificial intelligence model may be added in the future. When discussing Apple's vision for third-party AI integration, Federighi emphasized the company's commitment to providing users with more choices.
For Apple, Cook wants to build its own AI ecosystem. In addition to launching Apple intelligence based on local models, Apple also emphasized the capabilities of its own chips - Apple intelligence will enable private cloud computing mode on dedicated servers using Apple chips.
This has sparked heated discussions among "tech enthusiasts" on foreign technology forums. One netizen posted that "Apple's intelligent private cloud will use Apple chips as power, which may be the most important announcement. Nvidia is almost the only supplier of AI hardware, and they limit the ability of any company to adopt AI. Apple's decision to use its own chips shows that they will be able to better accelerate their investment in AI because they will have more AI chips available. This also indicates that Apple will continue to actively innovate hardware in order to compete with Nvidia."
Image source: Macrumors, a well-known technology forum
This article comes from the WeChat public account "Daily Economic News" (ID: nbdnews) , author: Wen Qiao, and is authorized to be published by 36Kr.
