Whoever can seize the opportunity of the AI operating system model will become the "leader" of the technology industry.
Author丨Dong Zibo
Editor丨Lin Juemin
AI has been popular for a year, but the biggest market opportunity is not in the big model itself.
In 2023, there will be many entrants who will devote themselves to AI, and under the "group model flurry", players can probably be divided into several categories.
The next strategy in the era of large models is to make a single model, which is intended to squeeze open the market door with technical force, one-sidedly focus on model capabilities and parameters, and insist on benchmarking the frontier of technology to make the most "bull X" large model.
This kind of phenomenon was the hottest in the first half of last year, and the craze set off by ChatGPT also heated up many people's brains, but in fact, they always found that in order to win the technology, they could not escape the scaling law of large models, and the investment was bottomless, and the technical gap could not be made up for a while.
As the capital market gradually cooled, last year, many of this group of people fell in front of the door of idealism, defeated by the waste of productivity caused by "repeatedly making wheels".
The middle strategy is to grasp both models and products, and use large model capabilities to drive the development of AI products, so as to find a place for artificial intelligence to actually land and generate value.
Someone once said that the current market idea of the large model is to put a good truffle in an oil-paper bag and wrap it in front of customers - no matter how good the quality is, blowing it to the sky, diners don't know how to eat it, and they don't know how to eat it;
Experienced chefs will supplement truffles with steak, cod, venison and other already delicious ingredients to maximize the advantages of truffles and "remake" those "traditional" dishes.
It is also true to create a killer application of AI: let AI play to its strengths and re-empower the original scenario with intelligence. If you can find such a right combination point, it is easy to win out the next byte and Pinduoduo company.
In the final analysis, the above two points are not the best policy. The larger pattern is to use a larger system to incorporate countless "killer applications" into its own downstream, promote the transformation of the industry and the world with technology, and mine a rare opportunity in the torque of the gears of the times for decades -
An opportunity to change the world with an AI operating system.
1
Every time human-computer interaction changes,
There is an opportunity on the scale of a trillion dollars
In the past year, in the process of communicating with many elites in the AI industry, AI technology reviews have always heard one voice:
Human-computer interaction is the "golden key" that makes a super technology company born. On April 16, at the Baidu 2024 Create AI Conference held in Shenzhen, Shen Dou, executive vice president of Baidu Group and president of Baidu Intelligent Cloud Business Group, told such a story:
(Shen Dou, Executive Vice President of Baidu Group and President of Baidu Intelligent Cloud Business Group)
From the earliest days, the interaction between humans and computers still relied on wiring boards, and workers manually plugged and unplugged cables, turned knobs, and human-computer interaction was even a "manual work";
With the advent of assembly language, assembler, high-level programming language, and compiler, humans can communicate with computers in a "language" in a "similar" way;
The timeline is advancing again, and soon, it is difficult for a single machine computing power to face the rapidly growing market demand, cloud computing came into being, and human beings are no longer satisfied with the "dialogue" with one machine at the same time, but to cooperate with the "cluster" formed by multiple machines on demand, which pushes human-computer interaction to a higher level;
Today, the emergence of large AI models has once again brought new opportunities for human-computer interaction - taking programming as an example, programmers no longer have to start with linguistics, but can directly use natural language, and with the help of AI, they can achieve the functions they want;
In the same way, programming is no longer a "battle of wits and courage" between humans and computers, but a "wish" for AI.
With the generous help of artificial intelligence, the vision of "everyone is a developer" is not far off.
Looking back at this process, just like a cultivator turning over the soil in the spring, countless vitality has emerged, and every new revolution in human-computer interaction has brought an opportunity for the trillion-dollar scale of the technology industry.
Seizing this opportunity, Apple made the first personal computer, and then led the trend in the mobile Internet era; Microsoft made the first visual operating system, and used office software to lay the foundation for today's office flow; and with the blessing of cloud business, Amazon has soared nearly 300 times in the secondary market in the past 20 years, starting from an e-commerce platform and becoming one of today's top technology companies......
If you look closely, you will find that the birth of these trillion-dollar technology companies is inseparable from their efforts to seize an opportunity to change the way humans interact with each other in the wave of technological innovation.
As mentioned at the beginning, the current momentum of large models is booming, and if AI is simply regarded as a technology or a product, then the true potential of AI has not yet been fully explored.
On the contrary, it is only possible to leverage this trillion-dollar opportunity by integrating AI capabilities and matrices into an operating system, managing the runtime pool of hardware resources and software services, providing external service interfaces, and bringing new changes to human-computer interaction.
And how easy is it to make such an operating system and win the championship in the battle of human-computer interaction?
2
Do AI operating system,
High-performance heterogeneous computing is a "hard threshold"
"Big models are not a free lunch. At the conference, Shen Dou preached. If you want to make the AI operating system run, computing power is a topic that can never be crossed.
At the beginning of last year, in the assertion of a group of analysts, computing power is not a key factor: after all, computing power is a "money game", as long as there is enough money, you can smash enough cards, this problem will be solved.
However, a year has passed, and more and more people have discovered that as the magnitude of model parameters becomes larger and larger, the magnitude of money required for the "money game" is also beyond the scope that most companies can afford.
In this way, it is difficult to come up with a convincing answer to the problem of computing power by simply "throwing money".
Even Sam Altman, CEO of OpenAI, the "leader" of large AI models, has more than once publicly complained about the lack of computing power in his hands. At least for now, the issue of computing power is strongly restricting the speed of AI development.
The calculation of large models also needs to "reduce costs and increase efficiency", and reducing the waste of computing power is an important way.
How to make the Wanka cluster "work together" like a card, so that the performance can be scaled linearly without interruption between tasks, which poses a huge challenge to the design, scheduling, and fault tolerance of computing power clusters of AI operating systems.
Baidu, which has recently launched the AI operating system "Wanyuan", is passing the effective training time on the Wanka cluster to account for more than 98.8%, and the linear acceleration ratio and bandwidth effectiveness have reached 95%, which has given the industry new hope.
(Baidu Intelligent Cloud released a new generation of intelligent computing operating system - Wanyuan)
However, only solving the problem of wasting computing power does not seem to be enough to complete the "cost reduction and efficiency increase" of large models once and for all - in China, the impact of the "embargo" is still continuing, and it has almost become a must to use a variety of heterogeneous computing chips with different performance to make up the 10,000 calorie computing power required to carry large models.
Heterogeneous computing is a huge problem to interconnect chips from different manufacturers, generations, and performances.
In the past, it was common knowledge that heterogeneous computing could only solve different tasks with different chips, and it was almost impossible to put a variety of chips into the same task, especially training tasks.
In order to rise to the challenge and solve the efficiency problem of heterogeneous computing, we have to need a little "black technology".
To make an AI operating system, Baidu's confidence in the computing power layer is that they have solved this "impossible" problem, shielded the differences between different chips, and gave users the freedom to choose chip combinations.
On Baidu's heterogeneous computing platform "Baige", under a single training task, tens of thousands of chips from multiple manufacturers cooperate with each other, and the performance loss of 100 calories is 3%, and the loss of 1000 calories is no more than 5%, and the efficiency of heterogeneous computing power collaboration is improved by hard technology.
It is precisely in this way that the "bottleneck" problem of chips has been solved by Baidu: it can not only get rid of the dependence on a single chip, create a more flexible supply chain, and minimize the risk, but also make the cost more controllable.
Baidu's solution, one is to separate the optimization and implementation of the underlying communication and the upper-level operator, let professional people do professional things, and be responsible for the development of the underlying acceleration library AIAK, so as to build a "big stage" for heterogeneous computing, so that all chips can run through Baidu's integrated communication library;
The second is to do a good job in the implementation of the parallel framework in the acceleration library, optimize the parallel strategy, and automatically realize the parameter setting of different parallel strategies of TP (tensor model parallelism), MP (model parallelism) and PP (pipeline parallelism) through adaptive algorithms, quickly grasp the parallelization strategy, and deal with the underlying settings of the training task.
The third is in the inter-card communication, the integration of network protocols, the GPU level is mainly through NVLink, and when the computing power out of the machine, Baidu mainly uses RDMA, for some special chips, Baidu also has some quite specific implementation strategies.
Since the coexistence of multiple chips is a reality in the short term, it is better to take the initiative to embrace it than to complain. Baidu makes "Baige" compete for the stream, and the main focus is to actively embrace the diversified chip ecology and achieve the ultimate in heterogeneous computing.
Without the hardware of the "cloud", the implementation of the AI operating system has become a "castle in the sky" - and the integration of cloud intelligence and collaborative software and hardware optimization are gradually becoming the consensus of many powerful players.
And to run to 10,000 cards, to let 10,000 cards run to the same mission, or a "multinational force" composed of multiple chips, is an insurmountable gap for any company.
Today's Baidu has run a road - the heterogeneous road of one cloud and multiple cores, single task and multiple chips;
(Industry-leading single-task, one-cloud, multi-core large model training solution)
If other players want to run through this system, and then build a new AI operating system on this system, years of accumulation and implementation in the field of AI and cloud computing are essential.
3
Tool chain, "low cost" has "high difficulty"
If computing power is the "core" of an AI operating system, then the toolchain is the "middle layer" that links developers and AI.
In the final analysis, no matter how strong the computing power and large model are, users will not be able to use the large model into the scene, and the AI operating system will only be empty talk.
At this level, the AI operating system is no longer just a technical challenge. How to understand the vast market and complex scenarios has also become a question that must be answered.
On the one hand, for users with different budgets and needs, AI operating systems must be flexible enough to provide a variety of solutions that can solve problems most efficiently.
Although the market demand for large-scale native applications has been increasing since last year, the "impossible triangle" of "performance, speed, and price" in use has also discouraged many people.
If only one model is used, the problem cannot be solved if the parameter scale is small, and it is difficult to level the cost and meet the requirements of high concurrency if the parameter scale is large.
This also requires the provider of the AI operating system not to blindly hold a consistent line of thought, and to provide a choice of pedestal models with different specifications in different scenarios, so as not to let users spend money in the operating system.
Taking Baidu as an example, on the basis of Wenxin 4.0, which has already gained a lot of recognition, the Qianfan team has trained three lightweight models of different magnitudes: ERNIE Speed, Lite, and Tiny, respectively, to meet the different needs of different users for large models.
Under the model routing service provided by ModelBuilder, for tasks of different difficulty, the AI operating system Wanyuan can also independently select the most suitable model to complete the call, achieve the optimal valence ratio, and reduce the inference cost by 30% when the effect is basically the same.
In addition to its own models, Qianfan also supports third-party model products such as Llama3 and Baichuan, focusing on a "what customers want, we provide." ”
On the one hand, large models make human-computer interaction move towards natural language, and they also need the support of a complete set of AI capabilities and tools.
A better AI operating system makes it easier for users to develop their applications, even by entering a short description of the app's features, or even writing a single line of code, to get an application that actually solves their needs.
Under Baidu's Wanyuan system, the built-in AppBuilder and AgentBuilder two application development platforms have not only made the above imagination fall into reality, but also the SDKs of the two platforms also support secondary development, so that the meticulous personalized needs of developers can be met.
Not only that, apps developed with Baidu AppBuilder can be published to Baidu Search, WeChat official account and other platforms with one click, making it no longer difficult to distribute apps, and can also be integrated into your own system through APIs or SDKs.
On the other hand, the complex scenarios make it necessary to fine-tune the model and improve the tool component system.
In Baidu Wanyuan, the number of officially selected components supported by development has increased from 11 last time to 54. Among them, there are various types of components of large models, components of AI capabilities, plug-in tools, and components of digital humans, and they also support all access at one time, eliminating a lot of cumbersome procedures.
Just like building blocks, users can combine different components, and after assembling them, they can become a workflow to complete the customization of large models that meet their own use needs.
If a person has been engaged in the To B business for many years, most of them will find that the underlying logic of To B is always "unpretentious": how to spend small money and do big things.
"Simplicity is the ultimate sophistication. Shen Dou started his speech at the Create AI Developer Conference with this sentence, which is intended to reduce the burden on developers and provide users with a minimalist development experience.
At present, the number of customers served by Baidu's entire large model platform has increased by 10,000 in more than a month, and has exceeded 85,000. The number of models fine-tuned by Qianfan has risen to 14,000, and the number of applications developed has now exceeded 190,000.
The cheap and easy-to-use toolchain has allowed Baidu's AI operating system to have a shallow pool of users. The accumulation of users, how to do the ecology, has become Baidu Wanyuan's "One more thing".
4
To do ecology, we must "be like Beichen"
The success of an operating system is not just a technical success.
More than ten years ago, smartphones have just come out, and Android and IOS are not the only ones that dominate - a large number of operating systems such as Symbian, Blackberry, Windows, etc., are in full bloom and competing fiercely. And at the end of the fight, only Android won.
In essence, developers still have to win the world.
And how to keep developers in the operating system, how to let developers find users, and make money? This also tests the ability and resources of AI operating system developers in the market field.
In the era of AI, the concepts of "customer" and "partner" from the perspective of the market are blurred by the further lowering of the development threshold. For the developer ecosystem of AI operating systems, one point is to bring them in, and the other point is to keep them.
The introduction relies on direct incentives and the convenience of joining the platform. This requires that the AI operating system be open enough and invest enough money to incentivize and hold developer competitions to allow more people to join the ranks of AI-native application development.
In the battle for developers, the first-mover advantage of AI operating systems is obviously more obvious - on the one hand, it can first obtain a larger group of developers and let them complete products on the platform earlier and obtain benefits, and on the other hand, it can reach B- and C-end users of AI-native applications earlier, which can also make developers stick and stay on the AI operating system earlier.
When it comes to the first-mover advantage, we have to mention Baidu. Whether it is in large model technology, large model ecology, or AI operating system, Baidu is the leader of a number of followers in China.
On Qianfan's app store, 300 AI-native apps have been put on the shelves, and the first batch of online apps have already begun to get a share of profits.
In order for people to stay, developers must be able to continuously obtain commercial income and positive feedback on the platform, and finally realize the "fortune" of using the AI operating system, which is the core value that the AI operating system can provide to developers.
For example, the best one sold on the Qianfan platform is a presentation assistant called ChatPPT, which has sold thousands of orders so far at a price of about 100 yuan.
And one company, in just one month, has released more than 20 applications in Qianfan, covering memorizing ancient poems, writing essays, marketing, and drawing, and the net profit in a month is millions.
If simple work can bring in good income, why wouldn't developers want to stay on the platform?
Although there have been good results, it is not easy to reverse the user's habits.
On the one hand, it should let developers see the value of AI, and on the other hand, it should actually bring productivity improvement and economic benefits to developers.
And Wanyuan's ecosystem is not only that:
On the other hand, Wanyuan Link developers have made AI native applications prosperous through constantly updated capabilities and interfaces;
Next, Wanyuan links chip manufacturers, expands the team of chip adaptation, and provides developers with more simple and easy-to-use heterogeneous computing power;
To the left, Wanyuan can link enterprise users, so that they can build their own AI operating system based on Wanyuan's personality;
To the right, Wanyuan links to the intelligent computing center to promote more efficient computing solutions to more users.
This road, as it is said in the Analects, "For example, the North Star dwells in its place, and the stars share it." "With a better ecosystem, every link in the industrial chain will be surrounded by AI operating systems, and then create a broader ecosystem with AI as the axis.
5
Conclusion: 1+1>2
Technology, products, and market matrix should be made into an operating system in the AI era, and none of the three should be missing.
As the first major manufacturer in China to propose the concept of AI operating system, Baidu's Wanyuan does not seem to be a new technology or new product, but more like another integration of Baidu's intelligent cloud technology product system.
Some people may ask: this is just a new concept, what kind of operating system is cobbled together? If you want to talk about AI OS, Windows with New Bing and Copilot integration is a real AI operating system.
In fact, Wanyuan's proposal is likely to be Baidu's most exciting new move in 2024: the integration of many technical products in the operating system is a good demonstration of Baidu's ambition in the field of artificial intelligence as an AI giant.
And the reason why it did not choose the form of Windows, but chose To B, Baidu Wanyuan benchmarked against Microsoft's Azure, doing what they are better at, and finding a field closer to money.
Last year, many people talked about AGI, dreaming that in the future, artificial intelligence will change the world just like what is written in science fiction.
And there are still many people who "boringly" polish technologies and products, go deep into industries and scenarios, and only do the most practical things - they also want to change the world, and they want to change the world little by little today.
The operating system is another explosive opportunity for human-computer interaction in the technology world. To seize this opportunity, it is necessary to use a complete set of systems to form a scale effect, so that 1+1>2.
In the future, using AI operating systems to provide full-stack To B services and solve users' problems in one stop will inevitably be the general trend of excellent AI manufacturers, and it is also the inevitable direction of the transformation of cloud manufacturers.
In order to join this competition, other cloud manufacturers need not only to have enough high-quality large models as support, but also to have deep cloud technology and scenario accumulation, and to be able to do full-link support of "model-development-market". Like a bucket, there can't be a shortcoming.
Regard Wanyuan AI operating system as a fertile soil, and the applications that grow on it—whether developed by Baidu or by other developers—can enter many scenarios on the B-end and C-end in a variety of forms, thereby driving the company's revenue in the cloud field.
In this way, Wanyuan can be called the most solid cornerstone of Baidu Intelligent Cloud.
"Our ultimate success," Hou Zhenyu, vice president of Baidu, said in an interview, "is that we hope that our ModelBulider can produce more models, and our AppBulider can produce more applications, which is our biggest ideal." ”
(In the face of this trillion-dollar AI opportunity, AI Technology Review will continue to follow up and observe, and interested readers are welcome to add the author's WeChat: william_dong to discuss cognition and share gossip.) ）
Without the authorization of "AI Technology Review", it is strictly forbidden to reprint it in any way on the webpage, forum, and community!
Please leave a message in the background of "AI Technology Review" to obtain authorization for reprinting on the official account, and you need to indicate the source and insert the business card of this official account when reprinting.