The era of large models: quantitative and qualitative changes in the advertising system

The Violent Aesthetics of Data and Computing Power

In late '22 and early '23, the world was shocked by the emergence of generative AI. I, too, spent a lot of time researching and thinking about this matter: on the one hand, it exceeded my expectations in terms of effect; On the other hand, removing the bells and whistles on the surface, OpenAI is actually extremely focused on one thing: Scaling Law. It means that OpenAI uses a large amount of data and computing power, uses a general architecture (Transformer), and then focuses on All In, and uses "brute force solutions" to achieve the emergence of intelligence. In the vernacular, Sacling Law is "quantitative change leads to qualitative change", or "the violent aesthetics of data and computing power".

Scaling Law is an important technical belief in the field of deep learning, and in 2019, someone (Rich Sutton, "The Bitter Lesson") made an incisive summary, which basically means: algorithm carving is not very useful, more data - more computing power is king.

Scaling Law has also always been my technical intuition. The "Next Generation Advertising System" project (Advertising System 2.0), which began in 2021, is based on this idea: using larger models, more data, and stronger computing power to estimate advertising effects (CTR, CVR, etc.). And, after generative AI came out, the first thing I did was to pool Tencent's computing power; Build the infrastructure (machine learning platform) so that teams can use this computing power centrally and efficiently. So, on top of this, there is the "Tencent Mixed Yuan" model, and now I am also in charge.

Generative AI has taught me a lot. On the one hand, it strengthened my belief in Scaling Law, and on the other hand, it also made me think about how to further use the technology and thinking of large language models into the advertising system. The recently released "Advertising System 3.0" is an initial attempt.

Make the advertising system "really understand"

The most surprising thing about generative AI is that it understands. Generative AI can understand what the user is saying, and instead of "parroting", it gives people the feeling that "it really understands" and "it has intelligence". Let's not argue about whether generative AI is really intelligent; But what we know is that we didn't understand the previous advertising system at all. The core of advertising system 3.0 is to find a way to make the advertising system "understand more". Only when the advertising system really "knows" and "understands" the products, advertising creatives and users can it improve the certainty of delivery and reduce the "metaphysics" of delivery.

The first thing to do for the ad system to "know what to do" is the "new ad ID". The advertising ID is the lowest logic of the ad system and the starting point for the ad system to understand each creative. The multi-layered and complex structure of advertising plans, ads, and creatives left over from the past has resulted in extremely complex advertising IDs, and the data associated with each ID is sparse or even combative, making it difficult for the system to understand. And, due to the uncertainty of the system, various optimizers are building a large number of new creatives, which is called "heaping infrastructure" in the industry: the basic operation is to make a little fine-tuning of the material and then "bet once". This is actually exploiting the loophole that "the big model doesn't understand advertising and doesn't understand the product".

In the form of a complex ID system + a large number of pile of material infrastructure, the audience's behavior is greatly diluted. So there's very little data on each ad, and it's not achievable to expect to make good predictions with overstretched data. So what we have to do is to "stack" ads with the same product - similar materials through the new ID system, so that there will be more data when the model predicts, and the advertising will have higher stability and better results.

We have seen that after the launch of the advertising system 3.0, the number of advertisements of Tencent Advertising has dropped from 7.7 million to about 700,000, and the stability and certainty of delivery have been significantly improved.

Second, the premise of "aggregation" is the "understanding of advertising content". That is, the advertising system can analyze the creative to understand what exactly is going to sell in each ad and what types of people are more suitable for, so that similar ads can be classified. Here, the system understands not only the title of the ad, the copy, but also the images and videos. Behind this is the "multimodal understanding ability" of the "Tencent Hybrid Model".

Looking ahead, we still have a lot to do to improve system understanding. The takeaway from Transformer is that it is the simplest and most versatile architecture to accommodate a variety of different forms of data. Therefore, we should also use the data of views, clicks, and add-ons in the advertising domain with other types of basic training data to build a model with a more general architecture, which may allow the advertising system to achieve "intelligent emergence" in the future and further open the ceiling of performance.

From managing the process to managing the endgame

Return to the essential needs of consumers

As the advertising system is upgraded, the capabilities of the model will become stronger and stronger, but the model will never be able to solve all problems end-to-end. Optimizers, designers, and delivery agencies are not going away in the process, but the nature of their work is beginning to change.

Think back to the previous work of an optimizer: many times it was necessary to complete the delivery operation at high speed and produce mass production of materials in batches...... In this process, the work of the optimizer itself is disconnected from the demands of the end consumer: they do not have the time and are not asked to understand the product and the consumer's demands. With the enhancement of AI capabilities, optimizers will be freed from these simple and repetitive tasks and think and make decisions from the perspectives of products, business models, and consumers.

At the end of the day, advertising is not an end in itself, the purpose of advertising is the final sale.

As a brand and agency, it is necessary to liberate from the process of paying attention to advertising, and think more about how to meet the essential needs of consumers: product, brand, and business model. That is, "from the management process to the management end".

First, there's the material itself. At present, AIGC materials in some industries have accounted for 20%; However, this does not mean that designers lose their jobs, but requires designers to use their ability to understand products and consumers to select and co-create a large number of AI materials. Get the content that best suits your brand and impresses consumers the most. In addition, brands and agencies should feed back and guide the future creative production of AI/large models based on the conversion data of materials.

Second, business models and links. "Generalized goods" - "selling things" have very different business models and links in different industries: for example, automobiles and real estate sales must be retained first, and the education industry must first convert low-cost courses and then full-price courses...... The essence of these business models is to constantly explore the needs of consumers. But the model itself does not directly understand these business models; Therefore, if we want the model to be accurate, we need to clearly define the link and optimization goals, so that the model can "understand" the business model and final requirements from beginning to end.

Third, comprehensive data. Whether it is a material or a link, the core is to let the model understand the product, understand the material, and understand the consumer demand. At the heart of it all is data. If our sales process is purely offline and unrecorded, then it is difficult for a smart woman to cook without rice: the model cannot understand the business link with no or lack of data, let alone the role of advertising in the overall sales process. Therefore, for the model, it is very important to obtain the full link data for the optimization of the effect. Taking live e-commerce as an example, only with a complete data of the whole process of appointment, viewing, liking, add-on, payment, logistics, and return can the model truly understand the demands of consumers. Therefore, for brands and agencies, the first thing to do is to digitize the sales process, and then the second step is to effectively cooperate with the rest of the platform to fully realize the business value.

At the same time, as a platform, its responsibility is to better cooperate with advertisers and agencies: to provide a stable advertising experience, continuously improve the advertising effect, create a more effective data cooperation model, and provide better and efficient creative tools...... And most importantly: to do the underlying technological innovation. I think the main line of this road is to continue to move forward on scaling law, and give more diverse, larger, more accurate and complete data to the advertising model; At the same time, it provides resources such as computing power and latency to the key model estimation links more intensively. I believe that if we continue along this road, the advertising model will also achieve a qualitative leap through the accumulation of quantity.

Redefining the value of people in the era of AI

Taking a step back, we can already clearly feel that the work in the AI era must be completed by people and AI. The natural question is: what should humans and AI do? How should humans and AI work together?

In a word, AI is good at a lot of parallel repetitive work; People are better at work with high uncertainty, strong innovation, and insight into human nature. AI can do a lot of information gathering, but decisions still need to be made by humans.

So back to the advertising industry, with the improvement of model capabilities, I hope that a large number of optimizers and operators can be liberated from inefficient and repetitive work, and really think about how to meet the demands of consumers and advertisers: a clearer brand image, more attractive material content, a smoother conversion link and business model, better products, and better services...... AI can play a role in scenarios that would otherwise require a lot of human labor: such as the creation and modification of materials, the modification of bids, data mining and analysis......

Finally, AI will become a management issue. Managers in the advertising industry are about to think about the career development of optimizers, designers, and traditional media in the new situation; and model-centric organizational architecture issues in the marketing domain. If we look a little further: most of the work in the future will have to be done by teams that mix AI and people. So how should we, as executives, manage such a hybrid team? Which jobs can be done better by AI, and which jobs are suitable for humans? Should AI be left to manage AI? Should AI be left to manage people? What decisions can be issued to AI? What must be done manually? …… These questions are worthy of deep consideration by every manager, because these things will become a reality in the next 3-5 years. In addition, we also need to "empathize" with AI, and find the most suitable work for AI from the perspective of AI. In this way, we can enable the company to evolve its business to "AI native" and give full play to the maximum and appropriate value of AI.

Jiang Jie | wen

Jiang Jie, Ph.D., joined Tencent in 2012 and is currently the Vice President of Tencent. As Vice President of Tencent's Enterprise Development Business Group, Jiang Jie is responsible for the overall technical management of Tencent's advertising platform products. He is also the Vice President of the Technology Engineering Business Group, managing Tencent's AI Lab, data platform, database platform, machine learning platform and billing platform.

Dr. Jiang has more than 10 years of experience in massive computing, distributed architecture, data mining, machine learning, etc., and has been invited to give keynote speeches at the China System Architects Conference and the China Cloud Computing Conference. Jiang Jie, as the person in charge of Tencent's general model, officially released the "Hybrid Yuan" model in September 2023.