The AI of the big factory is stepping into another river
Today's craze for large models is full of misconceptions.
One of the most popular is that "AGI" (Artificial General Intelligence) is about to be implemented, and the way to do it is to solve an extremely important task that you don't even know you need to solve now. The vehicle for implementation is a dialog box that is smarter than a person.
Therefore, everything needs to be turned upside down and started all over again.
This can be a great future, but it's not something that is going to happen right away. A series of recent events have brought people back to their senses. For example, GPT-5, which is a benchmark, has been repeatedly delayed, such as Zuckerberg's rejection of a single all-round AI product like ChatGPT for the first time; For example, a list known as the "AI Graveyard" went viral, including 738 AI projects that have died or ceased to function.
At the same time, several important AI products are reflecting the direction of another wave: instead of tearing down and rebuilding, they are incorporating large model capabilities into systems that already have a large number of users, so as to greatly improve the user experience; It is not to create something out of nothing, but to use a large model to really use the accumulated resources to better serve the existing needs of users.
Last month's WWDC announcement was immediately considered disappointing, largely because expectations for Apple to release an omniscient model were too high, but the subsequent market feedback such as the stock price surge reversed people's judgment and reflected a rethinking of the AI route that Apple represents.
Instead of having its own all-in-one model, Apple has built a three-tier model system: a local model for simple tasks, a private cloud model to ensure encryption and security, and a third-party model to provide more capabilities. This system is built for its complex ecosystem, to augment its own existing capabilities, and to improve the user experience on top of existing needs.
This is also why everyone later understood that OpenAI is not a role in eating Apple, even if ChatGPT is currently the strongest large-scale model product, it cannot "catch" Apple's user needs, and only Apple can serve them.
Google's thinking at the annual conference Google I/O is the same, leaving aside the "futures" product Astra that is coping with OpenAI's pressure, and its release is more about integrating Gemini into its existing product line of hundreds of millions of users, rather than updating a separate new Gemini app itself.
There is also a similar trend in the recent update of some domestic national products. At the just-concluded World Artificial Intelligence Conference (WAIC), Alipay focused on the latest AI application - the intelligent assistant integrated in the Alipay App, you can find it in the Alipay homepage drop-down, and the services that used to be completed with multiple clicks, such as ticket booking, food ordering, consultation and registration, can be done more easily by speaking.
Unlike those sci-fi scenes that are too late to experience, Apple gave the example of "Let's say one of my meetings is rescheduled to late afternoon, and I wonder if that will affect my ability to attend my daughter's show on time", which will be available in a subsequent Siri update.
Similarly, Alipay is also concerned about how AI can help people solve life problems, and among the realized functions displayed by the intelligent assistant, they include "help me order a large cup of Starbucks Ice American", "help me pay 200 yuan phone bill", "how much money I spent last month", "help me find out the flight from Shanghai to Beijing after 7 p.m."
Since April this year, Alipay has been testing this new intelligent assistant on the homepage, it is not an "AI native application" that focuses on chat communication, but more like an AI life steward integrated into the ecology of the Alipay platform, not only "has a brain and a mouth to talk", but also "has hands and feet to do things".
In addition to "life partners", there are also "work partners" - in January this year, the "AI super assistant" launched by DingTalk has become the entrance to almost all functions of DingTalk. Earlier, Microsoft also embedded Copilot (AI assistant) into all its office applications such as Word and Excel.
These are typical self-reinventions of platforms or apps. They have not left behind the daily needs of hundreds of millions of users, they will not change due to the emergence of large AI models, but new technologies will make a big difference in meeting the needs.
This kind of self-reinvention of the product, starting from the needs of users, does not seem to be so "show-off", and even a "stupid" job, which requires both ecological protection and even systematic secondary development.
For example, Apple Smart needs to process users' personal information in a secure environment based on a large model, so as to further "judge whether the user's schedule conflicts". In order to complete the task of booking tickets, Alipay's intelligent assistant is backed by systematic ecological and technical support to form a closed loop, and it is necessary to personalize the processing of users' personal information on the basis of privacy protection.
These companies are also often the ones that care most about privacy and data protection, so you'll see features released with a "boring" introduction to the technical guarantees of data privacy, which of course doesn't have a Scarlett-Johansson-accented AI avatar that catches the eye, but it's crucial for people who rely on these services in their daily lives.
AI should be integrated into users' life scenarios and provide more humane services. As for the large model itself, there is no need to make a noise, it has to come out and lift all the tables, it can hide in the back.
"Make AI as easy as scanning QR code payment" is a very vivid statement. A QR code simplifies many tedious processes, but also hides many complex technological innovations. This sentence proposed by Alipay clarifies the new direction of many national-level applications using large models, and it is also another way to "AGI" - not only All in AI, but also AI in All.
Three waves of large-scale model landing
Behind the "coincidence" of many national-level products, it is the product of putting the large model in a longer technological development perspective.
From the perspective of technological evolution, the breakthrough of large models can be regarded as a new stage in the long development process of machine intelligence, rather than a self-contained "genesis" moment, abandoning the past and subverting everything.
To a certain extent, we can regard the maturity of the Internet infrastructure, the resulting large amount of data and data processing technology, the algorithm model that has been advanced due to the abundance of data, the further birth of recommendation algorithms, etc., and the eventual prosperity of the mobile Internet as part of a whole uninterrupted process, and the explosion of today's large model provides the ability to completely release the accumulated technology and data assets in the past.
This also means that the large model is a very important but still difficult "brain" to work with, and it needs to be supported by a whole system to help the system complete the upgrade.
It's already very different from what people try to do at the beginning. In terms of how to implement large models through applications, there have been three waves in less than two years.
The first wave is AI chat apps. But people tend to forget that ChatGPT is an "accidental" product, originally just a demo to show what the model is capable of, and OpenAI itself was not prepared, and no one expected the series of changes it would trigger.
So in the first stage, the shocked people pinned all their fantasies on a magical dialog box - since it could show intelligence, then my conversation with him should solve everything. As a result, various large models have been turned into products in the form of a dialog box, and they have been launched one after another.
The second wave is to use simple methods such as Prompt to turn this dialog into an expert in some vertical scenarios, and then rebuild the corresponding vertical application. GPT store and the like are products of this stage.
Now the big manufacturers are rationally entering the third stage: based on the first two stages of trying, they find that they want to reinvent themselves simply by relying on a dialog box, and the implementation of large models needs to be integrated with existing systems, using existing technologies and resources to serve users, rather than completely starting over.
A look at Microsoft's series of attempts can better understand such trends. As the biggest promoter of OpenAI, Microsoft also hoped to save Bing through a universal dialog box for the first time after the emergence of ChatGPT, becoming the entrance to the future AI era.
However, Bing, which is blessed by ChatGPT, has a mediocre effect in competing for market share. After that, Microsoft quickly embraced OpenAI's GPTs concept, and integrated the GPT store function in the Copilot Pro service it provided to B-end users for the first time, but just 3 months later, it decided to go offline again.
In the end, Microsoft's AI strategy is to integrate large models into existing products and ecosystems - from Office Copilot to Recall, the most famous product in the latest AI PC, all of which focus on allowing large models to tap their potential in existing complex resources and then truly land.
This technical route is now defined by many as "AI Agents", and there is a clear consensus that AI Agents increasingly emphasize the ability of large models to intelligently invoke existing tools, services, and computing resources.
The big model isn't everything, but it can call everything really intelligently. Therefore, it also needs everything. And where is this "everything"?
In those national-level products that have been integrated into life and the thousands of complex scenes it has served.
The services that Apple showed "large model enhanced Siri" can bring are based on its system-level invocation capabilities that combine various software and hardware. What Alipay intelligent assistant is doing is also a system-level project.
Take the "smart ordering" function being tested by Alipay's intelligent assistant as an example: when the user says "please order me a large Starbucks iced latte", the large model first allows the AI to have the ability to perceive the screen and "see" the Mini Program page, and then quickly complete all the steps that used to be clicked by the user through the simulation execution ability, and after the user confirms and pays, he can go to the nearby offline store to pick up the coffee.
Alipay did not choose to simply call the data interface, but chose this intelligent service technology called ACT (Transformer for Actions), hoping to connect millions of small programs behind it and even thousands of digital life services on the platform through AI in the future.
We can imagine that the intelligent assistant of the future can help us book tickets, register, and plan our travel itinerary with a single sentence...... AI can also change from a "general" model to a "useful" gadget to serve more ordinary people.
It can be seen that whether it is Apple or Microsoft, or Alipay or DingTalk, platform products at home and abroad are no longer obsessed with creating new needs from a God's perspective.
They decided to let the big model out of the fantasy all-round dialog box, let the AI into the ecosystem, and let the AI fall into life - this may not be forward-looking, but the path of the big model landing that everyone can see can promote the true potential of this technology to begin to realize gradually.