Are Tencent, Alibaba, and ByteDance All Building Worlds for AI? Behind This Lies the Life-and-Death Battle of the Second Half of the AI Era
Hello everyone, I'm your financial journalist. Today, we're going to discuss a topic that might sound a bit like science fiction, but it's actually profoundly changing the landscape of the tech industry: Why are Tencent, Alibaba, ByteDance, as well as overseas giants like OpenAI and Anthropic, so eagerly building “worlds” for AI agents (intelligent systems)?
If you follow AI news, you've probably noticed that recent advancements have been about which models are smarter or respond faster. However, this time the competition is about something more fundamental and crucial—the environment in which these AI agents operate.
To make it easier to understand, I'll first summarize the key points from this in-depth report from “Ji Zi Guang Nian” and then break it down into five aspects to explain the logic and opportunities behind this trend in plain language.
📝 Summary of Key Points
In simple terms, the development of AI has entered its “second half.” In the first half, the focus was on problem-solving abilities (such as coding or writing articles). Now, the focus has shifted to the ability to define problems and to function effectively in complex real-world scenarios.
To enable AI to perform complex tasks like fixing bugs, booking flights, or managing databases, it’s not enough to just feed it a bunch of static data. What’s needed is an interactive environment where it can learn from mistakes and receive real feedback.
Currently, companies like Tencent’s Hunyuan, Alibaba’s Tongyi, ByteDance’s Seed, and overseas giants are investing heavily in creating these virtual worlds. This is not just a technical race; it’s also giving rise to a new infrastructure market. The company that can provide high-quality, large-scale, and evolving AI training environments may become the “new Scale AI” (a giant in data annotation).
---
🔍 In-Depth Analysis: Five Dimensions to Understand the “Building of Worlds”
1. **Why the Sudden Focus on Building Worlds? Because AI Needs to Start “Working”**
Previously, when we interacted with AI, it was like consulting an encyclopedia—you asked a question, and it answered without any action. But now, we want AI to perform tasks on its own.
Imagine asking an intern to fix a software issue. You can’t just give them a manual with instructions; you need a computer, a code repository, and a working environment where they can test, make changes, and see the results. This is exactly what AI agents need.
Yao Shunyu, Tencent’s chief AI scientist, put it plainly: “Without a good environment, agents can’t perform various tasks.”
Over the past six months, Tencent’s Hunyuan team has published six papers on this topic. They realized that traditional training data is static documents; what’s needed now is dynamic, interactive data. For example, AI should not only know how to code but also understand how the code affects the system and what error messages look like.
In simpler terms: Instead of teaching AI to memorize recipes, we’re teaching it to cook in a real kitchen.
2. **How Does Tencent’s Hunyuan Build These Worlds?** Multiple Teams Working Together
Tencent’s Hunyuan is making significant progress with multiple teams working on different aspects:
- Long-Horizon Tasks: They’ve extended test tasks from a few minutes to several hours, creating 37,000 such tasks at an average cost of just 5 cents each. This shows that building these worlds can be scaled and done efficiently.
- Mobile Environments: The Hunyuan Vision team has developed “PhoneWorld,” which simulates apps like Tencent Health and Maps, allowing AI to perform real-world tasks like booking vaccines or switching between apps.
- Evolving Environments: They’re creating environments that automatically upgrade as AI’s capabilities improve, ensuring there’s always something new for it to learn.
- Integration into the Training Process: They’ve automated tasks, building sandboxes, and validating results directly into the AI training pipeline. This means AI can practice in these virtual worlds in real-time during training.
Note: There’ve been organizational changes within Tencent, with Yao Shunyu taking charge of both large language models and multimodal models, emphasizing the importance of building these environments as a company strategy.
3. **Alibaba and ByteDance Are Also in the Race, and Even More Intensively**
Tencent’s efforts are not isolated. Alibaba and ByteDance are also investing heavily, with unique approaches:
- Alibaba’s Tongyi: Their research paper, co-authored with Zhejiang University and Peking University, focuses on using environments to enhance general intelligence. They’ve created over 800,000 verifiable software engineering environments.
- ByteDance’s Seed: Their “Agent-World” includes 2,000 environments and 19,000 tools, with dynamically generated tasks that help AI learn from its mistakes.
Common Goal: All these efforts are aimed at moving from static data to dynamic, interactive environments.
4. **The Biggest Challenge: 99.7% of Environments Are Ineffective**
A striking statistic from the report: Out of 47,678 publicly available environments collected from GitHub and Hugging Face, only 127 passed the rigorous screening process. The reason for such a high rejection rate is:
- Poor Quality: Many environments contain bugs or have inconsistent testing criteria.
- Too Simple or Fake: Small models may not be affected, but large models can exploit flaws in the environments to pass tests, without truly learning.
- Instability: Different hardware configurations can significantly impact AI performance. Stable environments are essential for effective training.
5. **The Next Big Opportunity: Building High-Quality Environments**
The data annotation market, once dominated by large models, is now seeing new growth in “environment construction.” Startups like UniPat AI (led by Alibaba, valued at $2.5 billion) are focusing on providing integrated solutions. Companies like Neptune (acquired by Mercor) are replicating real software environments for AI training. Large companies like Tencent and Alibaba are also building their own environments to maintain control and efficiency.
The Future Opportunity: The key is to understand both AI needs and how to create realistic, complex business environments. Companies that combine industry knowledge with AI environment expertise could become the next giants, providing simulated environments for AI training.
💡 Conclusion: From Problem Solving to Problem Definition
Yao Shunyu said last year that the focus of the AI second half is to shift from solving problems to defining them. This is now becoming a reality. By building complex environments, we’re helping AI learn to identify, analyze, and solve problems in uncertain situations.
For consumers, this means AI assistants will become more than just chat companions; they’ll handle complex tasks, code, and manage schedules and finances. For investors and entrepreneurs, “AI training environments” represent a burgeoning infrastructure sector. While data annotation was once a costly endeavor, environment construction is more practical and closely tied to industry needs.
In summary: Instead of feeding AI with data, we’re building entire worlds for it. The more realistic, complex, and stable these environments are, the better AI will perform in the real world.