Tencent Chief AI Scientist Yao Shunyu made his first public appearance on June 5 at the Tencent Cloud AI Industry Application Conference, where he discussed the company's Hunyuan 3 large language model and responded to criticism that Tencent has fallen behind in AI development. Yao, who joined Tencent from OpenAI and leads the Hunyuan model team, participated in a dialogue with Tencent Senior Executive Vice President Tang Daosheng covering AI model strategy, product development, and the future of intelligent agents. The 28-year-old scientist addressed external skepticism by characterizing AI competition as a marathon rather than a sprint, stating that the world cannot rely solely on ChatGPT as a single dominant application.
Yao Shunyu Details Hunyuan 3 Development Approach
Yao described three core improvements in Hunyuan 3 during the conference dialogue. "There's actually no secret. Building large models today is to some extent a rather tedious task. We should focus on getting the infrastructure right, getting the data right — the algorithm part is relatively simple," Yao stated. The improvements included rebuilding the entire infrastructure for pre-training and reinforcement learning, comprehensively upgrading data and evaluation systems with greater emphasis on defining real problems and improving data quality, and making taste-driven decisions across hiring, model development cadence, and tradeoffs.
Yao, who proposed the ReAct architecture during his doctoral research, emphasized that many decisions in model development are "taste-driven" rather than formula-based. "A lot of decisions are actually very taste-driven," he repeated multiple times during the dialogue. His doctoral thesis from 2019, titled "From Next Token Prediction to Digital Automation," explored language agents during the GPT-2 era. "At that time it was still the GPT-2 era, and it couldn't even generate continuous paragraphs — there were many rough edges," Yao recalled. "But I felt at the time that GPT was a very elegant thing, and outputting the next token is an extremely simple yet very general task. I believed it had the potential one day not just to output the next token, but to automate everything in this world."
Yao identified coding agents as the most essential form of intelligent agents because they are Turing-complete. He outlined Tencent's three-part strategy for agent development: emphasizing comprehensive system design, fully utilizing data flowing back from product lines, and maintaining sufficient imagination for exploration. On model development strategy, Yao shared three directions: maintaining comprehensive data systems despite coding becoming the most important track, leveraging product-line feedback data through co-design experience, and exploring next paradigms in technology and product evolution.
Regarding cost-performance tradeoffs, Yao stated that performance is the prerequisite for cost-effectiveness. "Many people find that using a strong model is cheaper than using a poor model, because it gets things done correctly faster," he explained. He suggested that achieving strong performance with relatively smaller models and maintaining robust performance across most tasks "may be more valuable in today's China."
Tencent Defines AGI Organizational Requirements
Yao introduced the concept of "AI's second half" in a blog post last year, a term now widely used in the industry. He explained that the core of this concept lies in a fundamental shift: while the past decades of AI development focused on "finding methods," now that methodologies have matured, "finding good problems" has become more difficult. "In the past we invented methods like AlphaGo to play Go, but it could only play chess. We made a special model for translation, but it could only do translation. But with pre-training and post-training, we now have a universal hammer that can hit any nail," Yao explained. "The more difficult thing is actually finding good problems to solve."
Yao stated that Tencent's vast product portfolio and scenarios provide authentic problem sources for AI technology, which was his second reason for joining the company. He identified culture as his primary motivation. "When I first chatted with President Tang and other executive leaders, my first impression was that everyone was very honest about what we're doing well and what we're not doing well — very straightforward without covering things up," Yao said. "Tencent overall operates based on trust rather than metrics. This candid, low-ego, pragmatic culture, along with commitment to long-termism, is crucial for building a long-term AI organization."
According to Yao, the most important task in AI's second half is establishing a long-term AGI-based organization in China. This organization requires constructing a "balanced triangle": solid foundational technology, products that create value, and a spirit of frontier exploration.
Yao and Tang Address Pace Criticism at Conference
Yao stated during the dialogue that AI is a long-term game and the second half has just begun. "I don't think ChatGPT and Claude Code will be the only super applications — that would be a very bleak world. Today is like the 1970s when PCs first emerged — there are still many, many things that need to be done," Yao said. He predicted the future will become more diverse rather than more singular, noting that "coding agents are just getting started, and multimodal, embodied intelligence — many, many new things are just beginning to happen."
"In the past, models and products underwent much exploration and took many detours. I think that's normal," Yao stated. "The more important thing is whether we can honestly face ourselves, whether we can be real, whether we can see feedback and then change, and maintain patience. That's the most important thing in the second half."
Tang Daosheng stated that Tencent welcomes external criticism and suggestions. "We are a company with very diverse business formats. Sometimes we may be fast, sometimes slow, and we'll fail in some areas. But this is a marathon, and we believe models will continuously iterate, user needs will keep changing, and new product formats will emerge," Tang said.
FAQ
What did Yao Shunyu announce at the June 5 Tencent conference?
Yao Shunyu, Tencent's Chief AI Scientist, discussed the Hunyuan 3 large language model development approach at the Tencent Cloud AI Industry Application Conference on June 5. He detailed three core improvements: rebuilding infrastructure for pre-training and reinforcement learning, upgrading data and evaluation systems, and implementing taste-driven decision-making across hiring and model development. Yao also responded to criticism about Tencent's AI development pace by characterizing the competition as a marathon rather than a sprint.
Why did Yao Shunyu join Tencent from OpenAI?
Yao stated during the conference dialogue that culture was his primary reason for joining Tencent. He described his first impression of Tencent leadership as "very honest" and "straightforward without covering things up." Yao explained that Tencent operates based on trust rather than metrics, with a "candid, low-ego, pragmatic culture" and commitment to long-termism that he considers crucial for building a long-term AI organization. His secondary reason was Tencent's vast product portfolio providing authentic problem sources for AI technology development.
