Summary of Key Points
This conversation is an open exchange between Tang Daosheng, Senior Executive Vice President of Tencent Group, and Yao Shunyu, Chief AI Scientist. The main topics include Yao Shunyu's role at Tencent after joining the company, the essence of the "second half" of AI development, Tencent's new AI strategies (the integration of models and products, and the upgrade of Hunyuan 3), as well as responses to claims that Tencent's AI progress is slow. Yao Shunyu has established cross-team trust through concrete actions (such as helping Yuanbao adapt to DeepSeek) and has gained the authority to coordinate the integration of Tencent's AI models and products. He believes that the focus of the "second half" of AI development lies in identifying the right problems, rather than just focusing on technical methods. The strength of Tencent's product scenarios is what attracted him to join the company. Tencent's approach to AI has shifted from a competitive landscape (where different business units developed similar products) to one of collaboration, where product capabilities can be integrated with each other. The upgrade of Hunyuan 3 focuses on infrastructure, data, and decision-making based on experience. Finally, he addresses concerns about slow progress by stating that AI is a long-term, multi-faceted marathon that has just begun.
1. Yao Shunyu at Tencent: Building Trust to Gain Authority
After joining Tencent, Yao Shunyu's responsibilities expanded from managing model development to overseeing the entire AI infrastructure, making him the key figure in coordinating the integration of AI models and products (except for WeChat). His ability to gain this authority stems from his efforts to build trust across teams:
- When Hunyuan's pre-trained models were not yet ready, he assigned the team with the strongest post-training capabilities to help Yuanbao (Tencent's chatbot) adapt to DeepSeek. Many in the model development team were skeptical: “Why help others when our own models aren’t fully developed?”
- However, Yao Shunyu recognized that maintaining Yuanbao’s user base (DAU) was crucial for future model iterations and long-term collaboration. This action demonstrated that the model team considered issues from a product perspective, laying the foundation for trust in the launch of Hunyuan 3.
- This conversation also sends a signal to the outside world that Yao Shunyu has gained the trust of senior management and is a core decision-maker in Tencent’s AI strategy.
2. The "Second Half" of AI: Lack of "Tools," Not "Problems"; Tencent’s Strength Lies Here
Yao Shunyu argues that the "second half" of AI development is characterized by the overuse of tools, but what really matters are well-defined problems. He explains:
- For decades, AI focused on developing specialized tools for specific tasks (e.g., AlphaGo for Go, dedicated models for translation).
- Now that pre-trained and post-training techniques have matured, large models have become versatile tools that can be applied to various tasks, but what’s scarce are meaningful problems. Companies need to clearly define the specific issues they want to solve with AI.
- One of the main reasons he joined Tencent was the abundance of valuable problems: Tencent has a wide range of product scenarios, such as WeChat, Yuanbao, and WorkBuddy, each presenting real AI challenges (e.g., how to make chatbots more user-friendly or office tools more intelligent). These are precisely what the "second half" of AI development requires.
3. Tencent’s Transformation: From Competitive Isolation to AI-Driven Collaboration
Tencent’s previous approach involved different business units developing similar products in competition, but this is changing with the advent of AI:
- Yao Shunyu emphasizes that LLMs (large language models) require versatility, such as being able to handle coding tasks, which requires a combination of coding data, communication skills, and reasoning abilities.
- Tencent’s product scenarios are now beginning to interact with each other. For example, features from Yuanbao (chatting and search capabilities) can be integrated into other products like ima or WorkBuddy, and the data generated by these products can be used to improve the models, creating a networked system of collaboration.
- This collaborative approach enables Tencent’s AI capabilities to be utilized more effectively as a cohesive whole.
4. The Upgrade of Hunyuan 3: No Secrets—Infrastructure, Data, and Experience-Based Decision-Making
Yao Shunyu reveals that there are no secrets behind the upgrade of Hunyuan 3. The key improvements include:
- Rebuilding the infrastructure: Overhauling the underlying systems for pre-training and reinforcement learning to make model training more efficient.
- Optimizing data: Using more realistic and diverse data, improving data quality (e.g., focusing on users’ natural, vague queries rather than precise benchmarks).
- Experience-driven decision-making: Making critical choices (e.g., hiring, model iteration pace, resource allocation) based on experience and judgment. For instance, decisions about when to release a preview version or allocate more resources to certain data types are made based on team intuition and feedback.
5. Is Tencent’s AI Really Slow? It’s a Long-Term Marathon
In response to claims that Tencent’s AI progress is slow, Yao Shunyu points out:
- AI is a long-term endeavor, not a short-term opportunity for quick profits. It’s similar to the PC era of the 1970s; ChatGPT won’t be the only dominant technology, and there will be many new opportunities in the future (e.g., coding assistants, multimodal interactions, embodied intelligence).
- AI development is diverse, with multiple potential directions. It’s normal for Tencent to have taken some detours, but the key is to be honest about feedback, continuously adapt, and maintain patience.
- Tang Daosheng adds that Tencent’s rich product scenarios (e.g., WeChat’s context, Yuanbao’s user data) provide the necessary resources for AI development. The current focus is on adjusting the pace of progress, not on being slow overall.
This conversation signals to the outside world that Tencent has identified a core leader for its AI strategy and has established a collaborative approach. For users, future Tencent products (such as WeChat and office tools) will become increasingly intelligent due to a unified AI framework and collaborative development.