Summary of Key Points
Tencent has recently undergone several significant changes in the field of AI:
1. Under the leadership of Chief AI Scientist Yao Shunyu, the focus of Tencent's large-scale AI models has shifted from competing for rankings to enhancing user experience. The HY3 model has been developed by optimizing data quality.
2. AI products, such as intelligent agents and coding tools, have seen rapid growth, leading to a surge in token consumption. However, due to limited computing power, Tencent has chosen to collaborate with chip manufacturers rather than developing its own chips.
3. AI has transformed the company's organizational structure, making research and development teams more agile and compact. The approach to development has shifted from pre-defined functional designs to result-driven approaches.
4. While AI coding has greatly improved efficiency, it has also introduced new challenges, such as the potential for AI-generated code to contain errors (known as "AI hallucinations") and the need for extensive code review.
5. The commercialization of C-side AI solutions still faces high costs, so Tencent is prioritizing improving the product experience at this stage.
Detailed Explanation
1. Yao Shunyu's Influence: Making Tencent's Large-Scale Models More Practical
Previously, Tencent's large-scale AI models may have focused more on external rankings. With Yao Shunyu in charge, the emphasis has shifted to user satisfaction. For example, he has collaborated with C-side applications like "Yuanbao" to ensure that the models better understand user needs. They have also eliminated unnecessary training data, retaining only high-quality information, which has resulted in the HY3 model being used by 80% of Yuanbao users. Tang Daosheng praised him for simplifying complex models to focus on a more user-centric experience.
2. Insufficient Computing Power: Collaborating with Chip Manufacturers
AI requires substantial computing power, especially as Tencent's AI products (such as intelligent agents) are being used increasingly frequently, with token consumption doubling monthly and exceeding 5 trillion per day. Although Tencent's available GPU resources are not sufficient to meet demand, the company does not plan to develop its own chips. The lack of global manufacturing capacity makes it more practical to collaborate with multiple chip manufacturers.
3. AI-Driven Organizational Change
In the past, internet companies developed products using a traditional "pre-made" approach: designing functions, then researching and developing them, and finally testing them, with a hierarchical structure. In the AI era:
- Smaller Teams: Teams of three to five people can handle development tasks efficiently. For instance, the WorkBuddy team operates without a traditional management hierarchy; there are only groups within departments.
- Transparent Processes: Business modules, code, and user feedback are shared openly, ensuring smooth information flow despite hierarchical barriers.
- Result-Oriented Approaches: The focus is on desired outcomes rather than predefined functions. In the future, even a single-person company could create complete products with AI assistance.
This flat structure eliminates inefficiencies associated with lengthy reporting processes. For example, if an engineer leaves, the team can easily continue working because all information is accessible.
4. AI Coding: Programmers as Commanders
When Tencent started with AI coding in 2022, only 30% of the code generated by AI was acceptable. Now, over 90% of the code in many teams is created by AI. Using tools like CodeBuddy, programmers can direct multiple AI assistants to develop different components simultaneously. The "demand throughput rate" (the ability to handle user requests) has become a key indicator of innovation speed. Although the cost-effectiveness of using tokens is still being evaluated, the business growth resulting from these investments far exceeds the initial investment.
However, AI also brings challenges, such as the potential for errors in the generated code and the need for extensive human review. The industry is beginning to implement constraints on AI to ensure the quality of the code.
5. Challenging Commercialization of C-Side AI
Some C-side AI applications have started charging users, but it is not easy to make this a profitable model. The operational and computational costs of AI services are high, and advertising models are not viable due to the variability in user usage. Therefore, Tencent is not competing on the speed of commercialization but is focusing on improving the product experience first.
Furthermore, with AI becoming a form of digital outsourcing, the role of employees has evolved. While AI can handle routine tasks, humans remain responsible for final quality control and accountability, such as reviewing and correcting AI-generated code.
Conclusion
Tencent is being reshaped by AI, adapting its models, organizational structures, and commercial strategies to meet the new demands of this technology. Despite challenges related to computing power, costs, and potential errors, the overall direction is to use AI to improve efficiency and enhance user experience, rather than merely pursuing technological showcase or quick profit generation.