Apple and WeChat’s slower pace of AI development than public expectations is the combined impact of privacy protection principles and technical route choices, and also reflects the trade-offs and considerations between the two under the application-first strategy.
Many people are probably going to complain when they see the title –
Isn’t Apple in full swing pushing “Apple Intelligence”, isn’t WeChat quickly accessing DeepSeek and Yuanbao, how can it be said that they are slow?
If we think carefully, we can find that the rhythm of the two at the AI level is actually slower than most people think.
How can product managers do a good job in B-end digitalization?
All walks of life have taken advantage of the ride-hail of digital transformation and achieved the rapid development of the industry. Since B-end products are products that provide services for enterprises, how should enterprises ride the digital ride?
View details >
Stop talking nonsense, let’s break it open and talk –
01
Let’s look at Apple first –
At WWDC last year, Apple launched “Apple Intelligence” in a high-profile manner, trying to tell the public that it is not left behind in the AI wave, and that it is also ambitious to redefine AI “the Apple way”.
Indeed, Apple’s style has always been this way –
Don’t grab the first shot, but once you make a shot, you can often redefine the rules of the game and teach other similar products to be human.
But Cook’s high spirits did not last long, and he personally admitted at this year’s earnings conference that the AI version of Siri was postponed, and many of the cakes drawn a year ago became futures, and “launch later” became a high-frequency word in the official caliber.
And attentive viewers and technology media found that many features of the first version of Apple Intelligence after the launch were very tasteless –
Genmoji’s low-frequency demand is better than nothing, while the practicality of “writing suggestions” has been criticized as “gimmicks more than substance”, and the effect of image elimination is even visible to the naked eye compared to Meitu Xiuxiu.
A slightly more complicated question requires users to confirm “whether to access ChatGPT”, and only higher models such as iPhone 15 are supported, while China will continue to wait.
So both YouTube and Bilibili’s digital KOLs have expressed disappointment with Apple Intelligence to varying degrees.
What is incredible is that according to a report by The Imformation, Apple showed a very amazing “whether meeting time modification can catch up with children’s school activities” case that spans multiple apps such as email, calendar, and maps at WWDC 2024 was actually “edited”.
Several engineers said in interviews that they were surprised to see this demonstration at the launch event, because none of these features worked properly on the test equipment.
The above shows that Apple Intelligence repeats the release of Siri 12 years ago, that is, when it was demonstrated, it was a seller’s show, and it actually became a buyer’s show.
The trailer is better than the feature film, at least so far.
02
Let’s look at WeChat again –
WeChat’s embrace of AI seems to have a lot of actions, in summary, there are two levels –
The first level – the AI of a single function, such as AI reading of the official account, agent reply of the official account, WeChat input method supports raw pictures, WeChat reading supports AI summary, etc.
The second level – access to leading large model products, such as search access to DeepSeek, and add ingots to the address book.
Especially for the latter, various tiger and wolf praise words such as “subversion”, “revolution”, “dry overturn”, and “end” in the industry emerge one after another, but there is still a gap in the actual experience and expectations brought to users.
So, what is the reason?
The core is that WeChat essentially only opens an entrance for Yuanbao and DeepSeek, and does not really connect it with WeChat’s underlying product architecture.
In a certain sense, the role of the entrance is seriously overestimated, and if the entrance is useful, Tencent will add an entrance to WeChat when Tencent made a short video back then.
JD.com occupied the entrance of WeChat’s nine-square grid back then, and the result was completely incomparable with Pinduoduo, which really played at the bottom of WeChat.
Compared with adding an entrance to Weishi, what is the difference between the video account?
The core is that it opens up WeChat’s social system, and it has become an atomic component that is truly integrated, and now neither Yuanbao added to the address book nor the “AI search” DeepSeek added to the mouth has not been truly integrated with WeChat.
These seem to be peripheral battles, all small dots, not connected into a line, and not the main thread that incorporates WeChat’s core functions.
WeChat seems to have just posted a notice – “AI inside”, and this is very not “Zhang Xiaolong”.
My personal strong feeling is that it is difficult to capture the shadow of Zhang Xiaolong’s product philosophy from the above actions, so the problem is——
What causes Apple and WeChat to move slowly at the AI level?
03
“Privacy” is a very important answer.
Almost all tech companies will say, “We value user privacy,” and for many companies, this may just be part of the PR wording.
But Apple and WeChat are a little different – Apple takes privacy and data security as important core selling points, while WeChat, as an IM tool, determines that data security has become an important cornerstone of its survival.
This leads to a fundamental contradiction –
AI’s intelligence comes from data, which refers to both massive training data and sufficiently long and rich context.
The more data the AI of Yuexiu X needs, and stronger privacy protection means less data sharing.
This “tightrope” is extremely difficult for Apple and WeChat –
Apple’s solution is a “troika” –
First, the device-side model, the data does not leave the iPhone;
Second, private cloud deployment, invented a unique and extremely complex end-to-end encryption system (for details, please refer to the relevant video of “Lin Yi LYi”, the UP master of Bilibili), and the data was immediately deleted after the task was completed.
Third, if you can’t figure it out, directly ask the user to confirm and ask ChatGPT.
And this comes at a cost, and the performance requirements of the device-side model make models below the iPhone 15 Pro miss out, and at the same time lead to greater power consumption.
Even so, the degree of intelligence is still quite limited, and high-end tasks have to be transferred to a third party in the cloud, and the way of calling OpenAI also sacrifices the dominance of technology to a certain extent.
04
The same is true for WeChat, I don’t believe Zhang Xiaolong has not thought about combining AI and WeChat’s core social functions, but privacy and data security are the “sword of Damocles” hanging over WeChat’s head.
Let’s take a look at two paragraphs of WeChat’s “Service Agreement” and “Privacy Protection Guidelines” –
WeChat data collection description of mini programs
WeChat data collection description of chat history
It can be seen that WeChat can indeed collect a lot of behavioral data of users for analysis, such as mini programs, video accounts and other business-related data.
But there is one thing – “chat history” is highly sensitive anesthetic data, and it is absolutely impossible for WeChat to use it directly for training.
Students who understand the principles of large model training will know that no matter how end-to-end encryption the data transmission and storage process is, the step of tokenization before training will undoubtedly change back to the “language” of plaintext.
This is the internal training requirement of the “language model”, and WeChat is likely to be unable to bear the risk of this leakage.
Although the value of chat history itself as a training corpus for base models is extremely limited at this technical stage (chat data logic is simple and simple), it is an important part of the “context” of the application.
Without it, the AI products of the main WeChat thread are equivalent to using an eye.
Therefore, without cloud chat records, whether it is looking for end-side solutions or creating data for post-training, it will increase the difficulty of the main WeChat social thread to embrace intelligence.
There are also some views that WeChat can go further and develop more “active intelligence”, such as –
1. The 10 friends I have chatted with most frequently and have positive feelings recently, and the circle of friends they posted was automatically liked within 5 hours;
2. When my friend posted two consecutive circles of friends saying “overtime is very tired”, I reminded me to send him greetings;
3. The boss in the work group asked about the work progress, reminded me to reply in time, and gave three different reply options;
…….
This plausible imagination first requires extremely high levels of intelligence and context length at the technical level, and more importantly, this social action leads to “intelligent inflation”.
What do you mean?
If WeChat really achieves the above-mentioned “intelligence”, then if you like me, how do I know if you took the initiative to like it or if AI helped you like it?
Once this AI function related to core social networking is launched, the trust crack will begin to spread, and WeChat may be backlashed by excessive “intelligence”, and Zhang Xiaolong, who has a high social awareness, will never let this happen.
05
But this does not mean that WeChat cannot make more intelligent attempts at the existing level, prompt expert Li Jigang recently posted an immediate –
Indeed, you must know that now WeChat already has a “reminder” function, so you can completely withdraw from a “smart version of reminder”.
The above is actually a minimalist example, in fact, we can also open up some possible brain holes –
Now after long pressing, there will be some common functions for writing death, and theoretically, these common functions can also be intelligent.
For example, long press and call out “WeChat AI” on any interface, automatically identify the context locally based on the content of the page, guess the user’s possible intentions, and intelligently recommend different actions.
In addition, almost all generative AI today is triggered passively, that is, users need to take the initiative to respond.
WeChat, an app with a high level of user interaction, can actually do some “active intelligence” attempts, although it has extremely high requirements for intelligence ceilings, privacy schemes and interaction design.
In this sense, WeChat’s AI has actually just started, and the more arbitrary statement is –
In the matter of WeChat AI, the product god Zhang Xiaolong has not yet made a move!
The current WeChat’s scattered AI and paste AI are just Brother Long practicing his hands, and this mysterious man is still observing, contemplating, and holding back his big moves. (For more information about WeChat’s ability to post late, please refer to my other analysis article “WeChat has a rare ability to flow for a long time”)
Now is the addition stage, and the next step is the multiplication stage.
06
What is Multiplication?
We can get a glimpse of Tencent’s Q1 earnings call –
In last night’s Tencent conference call, Tencent’s management almost announced the WeChat version of Agent when answering analysts’ questions, and the original words were expressed like this –
“In the WeChat ecosystem, we see an opportunity to create a unique Agent AI that deeply integrates various unique components of the WeChat ecosystem.
These components include social graphs, communication and community capabilities, content ecosystems, such as WeChat official accounts, video accounts, etc., as well as millions of mini programs in WeChat, which cover various information flows, transaction operations, and applications across multiple verticals.
This Agent AI can use WeChat’s unique social, content, and service system to integrate information and execute operations in different scenarios, forming a more differentiated product than the general-purpose Agent AI.
This Agent AI, exclusive to the WeChat ecosystem, has a unique competitive advantage compared with other more general-purpose intelligent agents on the market. ”
Although no timetable is given, this blueprint is pragmatic and grandiose, and this is the comfort zone of Zhang Xiaolong, the product god.
Intuitively judging, the WeChat version of Agent is the AI product with the real temperament of Zhang Xiaolong——
That’s right, from “generation” to “action”, WeChat is in a sense an “operating system” due to its social, content, and mini program ecology.
Therefore, ordering takeout, taking a taxi, booking a hotel, and placing an e-commerce order can be completed entirely within WeChat……. which provides a great imagination space for WeChat to be an “all-round version of life Agent”.
In my opinion, the mainstream scenario of the current so-called general AI agent is PC, and the mainstream population is productivity-centered knowledge workers, and the proportion of these people in the pan-AI population is actually not large.
At this stage, if there is only one product that can enable domestic agents to move from PC scenarios to mobile scenarios, from professionals to the masses, then this product is likely to be WeChat.
More importantly, if WeChat is willing to increase the input bandwidth of its large model, the upper limit of intelligence is likely to go to a big level. (For a specific analysis, see Wei Xi’s other long article “Short-term Overvalued, Long-term Undervalued – WeChat Access to DeepSeek’s Cold Thinking”)
07
Things to point out:Tencent’s conference call was “we think there is an opportunity to create …..” and did not say “we are coming soon”.
This shows that the current progress of the WeChat version of the Agent is still in a very early stage, and there is no way to give a clear timetable, at least for now.
I expect that even if various agents are in full swing and competing to release them, the WeChat version of Agent will most likely not be launched within the next two Q.
Not to mention, on March 7 this year, Apple acknowledged through a statement to Daring Fireball that the Apple Intelligence version of Siri enhancements, such as cross-app manipulation capabilities, has been postponed and is scheduled to be launched within the “next year.”
Basic Apple Guy’s analysis pointed out that these features will be delayed until iOS 19 and beyond, hinting at an extension to 2026.
One fact can be seen – Apple and WeChat are not currently fast in embracing AI on the main thread.
In addition to the above-mentioned privacy and data security constraints, another important reason is –
At the level of technical routes, they all chose “application first”.
In other words, Apple and WeChat themselves do not chase the “upper limit of intelligence” at the model training level.
Apple itself has been reported by the media many times to train its own models, but judging from the limited information, their goal is not to hit the peak of intelligence, but more to match their own functions through small-sized models.
Although WeChat has a self-developed large model that belongs to the same Tencent system, the two are not the same business group.
Although from the recruitment information, WeChat is also training its own model, we can intuitively judge from the relevant recruitment job descriptions of the WeChat business group that the center of this training is likely to work at the optimization level such as “post-training” and “fine-tuning”.
WeChat recruitment job description updated on May 14
It is difficult for us to judge whether this strategy is wise in the long run from God’s perspective today, after all, “insight into one’s own strengths and be one’s own longboard” is a logical choice.
However, it must be pointed out that both have to bear the cost of “take-it-it-the-get-ism” – the pace of running-in is long and the product iteration is slow.
OK, to sum up, Apple and WeChat are both national technology consumer-grade products, and in the process of embracing AI, the pace has been flattened due to two core reasons – the first is the constraints of privacy and data security, and the second is the pragmatism strategy of “application first”.
epilogue
OpenAI’s Sam Altman said in a recent interview with Sequoia Capital –
“The platonic ideal state is a very tiny reasoning model with a trillion tokens of context that you put your whole life into：Every conversation you’ve ever had in your life, every book you’ve ever read, every email you’ve ever read, everything you’ve ever looked at is in there.”
“The platonic ideal is to have a very small reasoning model with a trillion tokens of context that puts your whole life into it, every conversation in your life, every book you read, every email you read, everything you read is there.”
If this is really a picture of the future, then in the process of getting closer to this picture, at least for the time to be seen, Apple and WeChat are two mountains that cannot be bypassed.
