In the past two days, the foreign media The Information released a report on the cooperation between Apple and Baidu on AI, which caused an uproar. Let’s briefly summarize a few key information points in foreign media reports that triggered heated discussions among the outside world:
First, as a supplement to the AI capabilities in Apple’s iPhone, iPad, Mac and other series of products, Apple’s partner in overseas markets is OpenAI, and its partner in China is currently basically determined to be Baidu;
Second, a person directly involved in optimizing the large language model of both parties revealed that Baidu’s Wenxin large model has encountered difficulties in understanding prompt words and accurately responding to common scenarios raised by iPhone users;
Third, Baidu hopes to be able to save and analyze iPhone user data to further optimize the AI model, but Apple is unwilling to do so. It does not want to use user data for AI training, and there are differences between the two parties.
Fourth, in its cooperation with OpenAI, Apple does not need to pay it. In its cooperation with Baidu, Apple must pay the latter to use Wenxin Large Model 4.0, and also agree to bear the cost to retrain and adjust the model.
At present, the outside debate around the above information points is generally divided into two voices.
One voice is questioning Baidu: Can Baidu work? It seems that it cannot meet Apple's needs. Why can OpenAI do it but Baidu can't?
Another voice is of course questioning Apple: Is Apple good? Apple Intelligence has received a lot of negative reviews in other markets, and it even looks like a GPT in a shell?
Objectively speaking, both voices do make sense. After all, Zhao Mingdu of Honor next door used AI to treat more than 2,000 viewers to drink coffee at the press conference, and Apple’s AI has not yet landed in China.
Apple and Baidu do need to withstand such doubts - they are still the most high-profile mobile phone companies on the planet, and almost all manufacturers must benchmark against them; one has been doing AI for ten years, and has always been at the forefront of technology and applications when the wave of large models hits.
So, where is the real problem? Is it Baidu or Apple?
1. Why is it so difficult to implement Apple Intelligence?
Before discussing this issue, let's first take a look at how AI capabilities can be implemented on smart terminal products such as mobile phones, tablets, and computers.
The most critical one is the large model. Simply put, based on the four basic capabilities of artificial intelligence in understanding, generation, logic, and memory, the large language model can better and more clearly understand the instructions issued by humans and give responses.
For example, when you need to use AI capabilities on your mobile phone, such as AI photo retouching, AI painting, or wake up a smart assistant to complete certain operations (such as asking about the weather, asking about nearby delicious and fun places), the mobile phone needs to call the ability of large models.
But there's a problem.
If you simply put a large model into a mobile phone and run it, it will be limited by factors such as computing power and running memory. Qualcomm Group Vice President Ma Dejia has previously said that running a 10 billion parameter model directly on a mobile phone will take up at least 20GB of memory. If all large models are placed in the cloud, there will be security and privacy risks.
In order to solve this problem, the current mainstream approach in the industry is to run the cloud-side large model and the device-side large model at the same time.
The AppleIntelligence large model has three calling levels:
The first layer is the language model AFM-on-device with about 3 billion parameters that runs on the Apple mobile phone, that is, the device side; the other layer is the large-scale server language model AFM-server that runs on private cloud computing; the third layer is the third-party model, such as Baidu's Wenxin large model and OpenAI's ChatGPT.
Picture/Shenwan Hongyuan Securities
Guosen Securities previously pointed out in a research report that in the cooperation between Apple and third-party partners, the latter plays more of a supporting role. For example, it may only need to handle about 10% of users' response needs.
At present, most domestic mobile phone manufacturers are also following this "cloud + terminal" path.
OPPO's self-developed Andes GPT model includes both the cloud and the terminal. The Andes model is also divided into three layers, including the Titan model with 180 billion parameters, the Turbo model with 70 billion parameters, and the Tiny model with 7 billion parameters.
Among them, the Titan model and Turbo model are deployed in the cloud, and the Tiny model is deployed on the terminal.
Vivo is basically the same. The self-developed Blue Heart large model matrix also includes large parameter models and small parameter models. The former is placed in the cloud and the latter is placed on the terminal.
However, although the "cloud + terminal" large-scale model operation idea is the same, in actual implementation, Apple and other mobile phone manufacturers are different.
One of the biggest differences is that the application development ecology is different.
For example, Qualcomm and MediaTek currently dominate the system-level chip market for most Android phones. Since this year, from MediaTek’s Dimensity 9400 to Qualcomm’s Snapdragon 8 Extreme Edition, they are emphasizing AI performance. More importantly, both of them are accelerating the construction of an AI ecosystem.
Qualcomm launched a new AIHub in February this year, which is the developer platform. So far, it has access to 137 models, covering almost all current mainstream large model manufacturers' products. Developers can call the models they need to use on this platform, as well as the target hardware to be deployed, tune the models, and then integrate them into applications.
The same goes for MediaTek. At the first Dimensity Developer Conference in May this year, it invited partners such as Alibaba Cloud, Baichuan Intelligent, Transsion, Zero-One, OPPO, Honor, vivo, and Xiaomi to launch the "Dimensity AI Pioneer Program" to support developers in developing generative AI applications for terminal vendors equipped with Dimensity chips.
Apple's iOS ecosystem is different. It has been closed from the beginning.
In the long run, this means Apple has more imagination in AI capabilities. IDC China research manager Guo Tianxiang previously analyzed Caixin.com that Apple’s closed system has strong control capabilities for application developers. In addition to the system’s own applications that can use AI capabilities, third-party applications must also comply with Apple’s development standards to achieve true cross-application and system-level AI functions.
"Apple's AI function is currently closest to the concept of 'AI agent', and it is difficult for Android manufacturers to achieve this level." Guo Tianxiang said at the time.
A former Apple employee and iOS App developer once told Silicon Valley 101 that he even felt that Apple was going to "dominate the world." "For AI, Apple actually allows you to use your own data to train your model on its GPU, and then deploy this model to your App for use. In fact, it basically covers the entire development and inference from front-end to back-end to large models."
But a potential challenge for Apple is that whether its own large models are useful or not has not been successfully verified by the market - and although Apple has been saying that it will integrate other large models, its actions so far are relatively conservative, and it has basically only reached cooperation with two companies, OpenAI and Baidu.
Moreover, after Apple Intelligence was launched overseas, user feedback was not ideal. Many users complained that Apple Intelligence even seemed to be a shell of GPT.
For example, one netizen said that he could hardly find anything that could allow Apple Intelligence to be calculated on a large model on the device or in the cloud. All his requests were redirected to GPT chat.
From this point of view, the cooperation between Apple and Baidu is actually more difficult.
To a certain extent, this also means that Apple's reliance on Baidu's Wenxin model may be higher than expected - The Information's report also mentioned that Baidu's insufficient estimate of the depth of cooperation between the two parties may also be a big reason for the unsmoothness between the two parties.
But Baidu may indeed have encountered challenges. The foreign media report mentioned that Baidu’s Wenxin large model encountered difficulties in understanding prompts and responding accurately.
As for whether these problems can be solved in the future and whether Apple can successfully receive the olive branch, it is necessary to test Baidu's achievements in the past ten years.
2. Time is tight and tasks are heavy. Apple’s anxiety is written on its face
"In the first five minutes of Steve Jobs interviewing me, I decided to abandon all caution and logic and join Apple." In 1998, it only took Jobs five minutes to convince Cook.
Later, in the book "The Biography of Tim Cook", author Leander Carney mentioned this story and wrote:
They are the best partners.
But another detail about this story is that Jobs himself said, "Cook really doesn't understand products."
Judging from the span of the past ten years, Apple under Cook's leadership is not unsuccessful. Apple once became the first stock in the history of the US stock market to exceed US$2 trillion. In the past ten years, it has remained the most successful mobile phone company on the planet. With a long-term 20% market share, it has taken away 80% or more of the industry's profits.
It's just that the huge commercial success has not eliminated people's doubts about Apple, that is, the lack or even lack of innovation capabilities.
If Apple could ignore these in the past, it can no longer do so.
In the past few years, all Android mobile phone manufacturers have been striving to go high-end. Now if you want to buy the latest flagship mobile phone product from any mobile phone manufacturer, you must have a budget of about 8,000 yuan.
From the past when users had "no choice" in the high-end market to now having "too many choices", Apple should certainly be anxious, and it has already written its anxiety on its face.
Apple, which has never cut prices, has begun to adopt a price reduction and promotion strategy this year. From January this year to 618 in the middle of the year, and then to Double Eleven some time ago, Apple has launched corresponding preferential measures.
Another interesting perspective is that scalpers can hardly make money from the iPhone. When the iPhone 16 series was first sold this year, it was common on social media for scalpers to make two to three hundred or even less per phone.
IDC data shows that in the first half of this year, China’s smartphone market shipped more than 140 million units, a year-on-year increase of 7.7%.
Picture/IDC
Among them, driven by vivo, Huawei, Xiaomi and other manufacturers, the Android market grew by 11.1% year-on-year. In contrast, the iOS market shipments fell by 3.1% year-on-year, and its market share also fell out of the top five in the Chinese market.
In this regard, Canalys research analyst Zhong Xiaolei said: "Local manufacturers have demonstrated market leadership and occupied the top five seats in the mainland Chinese market for the first time in history. In recent years, Chinese manufacturers have firmly implemented high-end strategies and deeply collaborated with local supply chain partners, which has gradually achieved results in both software and hardware."
However, in the third quarter, thanks to the release of the iPhone 16 series, Apple returned to the top five positions, with a market share of 15.6%.
But for Apple, instability is, to some extent, a crisis.
This is why, when facing the new direction of change of AI, Apple is under greater pressure than it has ever encountered before - this may be a new turning point for it, but it may also be the beginning of decline.
It needs to urgently prove to users and the market that it is still a dancing elephant.
Although before this, from building cars to Vision Pro to folding screens, Apple's elephant has been clumsy every time it tries to turn around - it has given up building cars, Vision Pro may be discontinued, and folding screens are still stuck on patents...
Apple has no choice.
Text|Ting Chao Ti