The WWDC two weeks ago was probably the biggest reversal in more than a decade.
I remember that on the day of the press conference, Apple's stock price closed down almost 2%.
The direction of the financial market has directly guided the news media and technology bloggers who are rushing to the first launch, if you now go to the major platforms to search for the first content of the day, then there is a high probability that the content you see is singing about Apple.
But just one day later, Apple's stock price rebounded strongly by more than 7.2%, the biggest gain since November 10, 2022.
I don't know if it's because Wall Street is coming back to the scene or what.
At this time, when you look at the relevant content, Qingyishui has become Apple's AI revolution, leading the trend of science and technology.
It's not the media duplicity, it's the world that changes fast.
So what is Apple's magic power to make a 180-degree turn in the attitude of the capital market in just two days?
For the first time, Apple Intelligence has given the AI industry a practical way to implement it commercially.
I remember in the tweet that day, I said, "I don't think Apple is outdated in the AI era", "Brother Guozi may not understand innovation, but it must know how to make money." ”
Today we're going to talk about it.
Unique combination of hard and soft advantages
The main reason why the AI industry has always had a mediocre business prospect is because its industry leaders OpenAI, Google and other companies have never been able to strike a balance between revenue and expenditure, and the benefits from paying users are far less than the cost they pay for Nvidia chips.
It can be said that AI has been popular for almost two years, and the biggest beneficiary is not OpenAI, which has set off a wave, but Nvidia has made a lot of money by peddling its own artificial intelligence chips.
You're selling shovels on the edge of a gold mine, that's fine.
As a result, I asked, who had found the gold?
- Not a single one.
That's a big problem.
From this point of view, Apple's decision to use its own M2Ultra chip to train large models, although it may not be as good as NVIDIA in absolute performance, but from the perspective of cost and energy efficiency, Apple may be the only manufacturer that hopes to achieve a large model business without losing money.
In addition to relying on its unique hardware advantages on the cost side, Apple has also relied on raising the hardware threshold to pass on the cost of end-side AI to consumers.
At present, Apple's device-side large model is about 3 billion parameters, and according to Apple's own statement, its implementation effect can be comparable to that of an open-source model with 7 billion parameters.
According to the introduction of the official website, Apple has used the grouped query attention mechanism and the LoRA framework in the model framework this time, compressing the process of inference (that is, there are fewer arithmetic problems to be done in the process of querying answers), effectively reducing the memory occupation, so that the 3 billion parameter model can run normally with only 8GB of memory.
But even so, the whole series of iPhones can only be supported by the 15 Pro and 15 Pro Max, which directly cut the iPhone before the 14th generation.
As we all know, Apple has always maintained the rhythm of only Apple can do in the storage part of the product, and the price of upgrading the memory capacity is about the same as buying the same weight of gold, and after the advent of the AI era, the harvest of hardware and capacity may be further intensified.
That is to say, before the release of the iPhone 16, if you want to experience Apple AI, sorry, you have to start from 7999 for mobile phones and 4799 for tablets.
In addition, not only memory, but also storage capacity is at risk of being full.
Originally128The capacity of this mobile phone software is becoming more and more inflated, and this system integrates a large model with 30 parameters, and it feels like the future iPhone will have to be at least 512 to be enough, which is equivalent to spending at least 3000 yuan every time you change the iPhone.
It's time to test the adaptability of Apple fans.
Balance the third-party ecosystem and control the developers
In addition to hardware issues, software issues are also plaguing the AI industry as a whole.
At present, the ecology of our mobile phones is basically application-centric, but AI is precisely de-application.
I'll give you a simple example, let's say I plan to eat beef noodles at noon today, I need to open Meituan or Ele.me first, find the store I want to eat, choose a good flavor, add coriander, etc., and then place an order and wait for delivery.
And what is the age of AI? You just need to say to the AI, "I want to eat xx beef noodles, add more green onions, coriander, not chili" and then that's it, you don't need to open the takeaway software at all, all the operations AI will help you do it in the background.
Sounds pretty convenient, right, but this is almost an epic blow to third-party app manufacturers, and all business models built according to the current use logic will collapse.
One of the most impactful is advertising.
Now how endless the built-in advertisements of domestic software are, everyone should be deeply aware of it.
What pop-up ads, open screen ads, shake ads, and even the smart island where ads are explicitly prohibited in Apple's design specifications, domestic APPs can add them to you.
I basically don't go to JD.com to buy things now、Tmall,When I have shopping needs, I find other software to jump in from the opening screen and shake the ad.,The speed is faster than I can directly open JD.comAPP。
You can imagine that when you no longer need to go through the difficulties of using the functions of these software, you can directly say the idea that AI will do it for you, and no one will see the ads, then naturally no one will vote, which will fly away for developers.
Moreover, the impact of advertising is only for some small apps, such as WeChat, Alipay, Meituan, Didi and other super apps, and the impact of the decrease in open rate will be even more tragic.
This is a screenshot of the Chinese and English versions of Didi, and you can see if you compare it, how many "diversified businesses" that these large domestic manufacturers have stuffed into their software that has nothing to do with the main functions.
Those who do taxis need to do online lending, those who do takeaway games need to do games, and those who make maps need to do small videos.
If the software open rate decreases, and the "diversified business" that has been added with great effort has even lost its entrance, how can it make money if it is directly withdrawn from the bottom of the kettle?
There are also WeChat and Alipay, these two are even more powerful, I can't wait to directly integrate all the apps in your phone, and all kinds of small programs don't want users to get out of their software.
What iOS, MIUI, see me Tencent OS, Ali UI teach you how to make a mobile phone.
If AI turns WeChat back into a chat tool, do you think Tencent is in a hurry?
Finally, let's talk about the user experience, Apple Intelligence's biggest advantage over other AI phones currently launched is that it is system-level.
It can call more data, and then use a large model on the device or cloud to provide users with services that better meet their needs.
But the problem is that this model may work abroad, because data such as email, iMessage, etc. are in the hands of Apple, but they are different in China.
The data of major domestic software is basically in a state of separation, WeChat data is in Tencent, and Taobao's data is in Ali.
So how can you Apple understand users?
Meituan is the one who knows what I usually eat the most, and Tencent is the one who I usually like to chat with the most.
You Apple can't even get this data, so how can you say that you really understand my needs?
Besides, Tencent has its own mixed yuan, Ali has its own general meaning, and they all have large models, so why do users have to use you?
Therefore, it is very likely that in the future, software such as WeChat and Taobao will also have their own end-side models.
Because for these super apps, these models actually have their own business logic behind them.
For example, if you also send a WeChat message with a video link to a friend, the one sent by calling Mixed Yuan is Tencent Video, and the one sent by calling Apple Model may be Douyin Video.
In order to maintain the whole system, I am afraid that not many manufacturers are really willing to access Apple's Apple Intelligence.
But if it is judged that Apple Intelligence cannot be promoted, it is too underestimated Apple's control over its own ecology.
First of all, Apple has basically completely opened up its model to developers this time, and for small and medium-sized developers, the reduction in advertising business may certainly affect some of their revenue, but is it possible that they have no revenue in the first place? For small and medium-sized developers who have no traffic and no large model development capabilities, this is an opportunity to subvert the original large manufacturers with the help of Apple's resources!
App giants don't do it, do they? You don't do it, some people do.
Secondly, for Siri's acquisition of third-party functions, Apple also has a design. iOS18 opens the entrance of third-party apps directly in the control center and lock screen this time, but the price is that you have to agree that your software can directly communicate with Siri.
In other words, if manufacturers want to add the control center as a global entrance to their own software in iOS18, they must adapt to Siri, so that AI can directly call your functions without opening your program.
Again, if you don't access, some people will access.
As for the final data fragmentation problem, on the one hand, it depends on communication, but presumably the more major App manufacturers, the more difficult this communication is.
I really can't communicate, and another way is to read the content of the screen directly.
苹果曾经披露过一篇技术论文，叫《ReALM: Reference Resolution As Language Modeling》，讲的就是如何把屏幕上读取的东西都转化成文字，然后给到大语言模型去理解。
In short, as long as it is in the Apple ecosystem, it can always get user behavior data.
And on the other hand, if each big factory APP comes with its own model, then Apple will not suffer, after all, if it eventually evolves into at least dozens of GB per software, then the iPhone must not be sold at 32G memory, 1TB memory?
Therefore, let's not underestimate the power of Brother Guozi to do whatever he wants in his own closed ecology, as long as there is a new deterministic technical route, as long as it occupies the commanding heights of innovating the user experience, it can always adjust the interest pattern faster than the open ecology and embrace the new.
There are no eternal friends, only eternal interests.
Resolve legal and regulatory issues
All of the above we are talking about the implementation at the commercial level.
But another equally critical issue for the AI industry is how to address the different regulations around the world.
The simplest OpenAI at your Apple press conference, we Chinese users can't use it at all, how can you Apple solve this?
Although there has been no official news about this until now, from the perspective of the structure of the entire Apple Intelligence, Apple should have thought of a way to deal with this problem a long time ago.
Let's briefly review the structure of Apple Intelligence, which is composed of three parts: the device-side self-developed model, the cloud self-developed model, and OpenAI's GPT. Like some small tasks, the difficulty is not high, Apple also said at the press conference, if it can be solved on the end side, it will be solved on the end side.
It really can't be solved on the end side, and then it will be transmitted to the cloud, and Apple's own larger-scale cloud model will help deal with it.
And because Apple entered the large model late, the performance of the cloud model is not enough to solve the problem.
Hey, Apple has helped you find the most famous GPT in the industry to help deal with it.
You will find that in this structure, GPT is actually a bit redundant, because if the performance of Apple's cloud model catches up, GPT is actually not needed to process it at all.
So in fact, GPT is not very important to Apple Intelligence, it can only be regarded as a transition and supplement at this stage.
GPT之于Apple Intelligence，有点像搜索引擎之于Safari。
When we use Safari, we can change various search engines such as Baidu, Bing, Google, Sogou and so on at any time according to our needs.
In the same way, in the future of Apple Intelligence, we can also change various models according to actual needs.
In particular, Apple, as a multinational company that has been operating for many years, has considerable experience in how to adapt its products to the national conditions of various countries, so it is very likely that Apple Intelligence will use GPT for products sold in the United States in the future, and the products sold in Europe and China will use the corresponding regional model.
Looking at the AI wave in the past two years, although it seems that the entire industry is making great strides towards artificial intelligence, there are actually only a handful of products that have actually gained a certain amount of user retention after a period of release.
More than one of my friends has asked me before, isn't AI a bubble? If not, then why don't I feel the difference it makes in my life at all?
In the past, the AI industry either focused on improving model capabilities, or focused on the development of to B applications, which either had a bad effect on consumer applications, or could only be popular for a while.
And when AI finally penetrates into mobile phones, computers, tablets and other consumer electronics products that we are most commonly exposed to in our daily life, do we feel that AI has come this time?
It should be admitted that Apple is no longer the most sensitive company to technological development, but its ability to promote the implementation of consumer electronics-related technologies is still unparalleled.
Apple hasn't caught up in this wave of generative AI, but it's too early to say that Apple has fallen behind in the AI space.