Summary of Key Points
This news article highlights the divergent approaches taken by two prominent players in China's AI industry: ByteDance and DeepSeek. ByteDance (led by Zhang Yiming) is pursuing a closed-source, multi-modal strategy with a focus on developing comprehensive super-apps, while DeepSeek (led by Liang Wenfeng) adopts an open-source approach, focusing on single-modal capabilities and a streamlined ecosystem. The public statements made by these two low-profile leaders—especially regarding the priority of AI programming and their attitudes towards large-scale model development—have amplified this contrast. However, both companies are ultimately aiming for AGI (Artificial General Intelligence), albeit through different pathways. This divergence is similar to the competition between OpenAI and Anthropic, and it is expected to drive progress throughout the industry.
I. Fundamental Differences: Closed-Source vs. Open-Source
The choice between closed-source and open-source represents a fundamental difference in the approach each company adopts, akin to choosing different foundations for building a house:
- ByteDance’s Closed-Source Approach: The model code is not made public, allowing ByteDance to control the iteration pace independently. This strategy has advantages such as maintaining full control over user data and content ecosystems (e.g., using user conversation data from DouBao to train models) and enabling seamless integration of models into applications like JianYing and FeiShu. However, it limits the ecosystem's scalability, as ByteDance must rely on its own efforts to promote products.
- DeepSeek’s Open-Source Approach: Even their most advanced models have open-source code, allowing global developers to contribute to improvements and adapt the models for various hardware platforms. This strategy results in a larger ecosystem, with many companies using DeepSeek’s models for private deployments, but it comes at the cost of limited access to user data, which is a disadvantage when pursuing multi-modal capabilities (e.g., video and image processing).
Liang Wenfeng has directly criticized ByteDance’s closed-source approach, questioning its benefits. While ByteDance has not explicitly stated its reasons, its decision to remain closed-source is likely driven by its substantial user base and content assets, which open-source models may not fully leverage.
II. Technical Directions: Multi-Modal Versus Single-Modal
The technical focus of the two companies is also vastly different:
- ByteDance: Emphasizing multi-modal capabilities, ByteDance has praised its Seedance video generation model as distinctive and leading. Their goal is to create AI systems that can handle various modalities (text, images, videos, audio), with plans to develop models with over 5 trillion parameters (setting a new domestic record). In short, they aim for AI that can perform a wide range of tasks.
- DeepSeek: Focusing on single-modal capabilities and coding agents, DeepSeek believes that the current priority is on developing AI programming assistants. They specialize in visual understanding (e.g., image analysis) and have achieved significant success with their DeepSeek V4 Flash model, which was once the most frequently used model globally. The lack of access to large amounts of data makes it challenging for them to develop multi-modal capabilities.
III. Product Strategies: Super-App Matrix Versus Streamlined Ecosystem
The product strategies reflect these technical differences:
- ByteDance: Developing a comprehensive “DouBao” super-app that integrates features from FeiShu and other services, aiming to create a one-stop AI platform. They envision DouBao as the AI equivalent of WeChat, offering a wide range of functions.
- DeepSeek: Avoiding the super-app approach, DeepSeek focuses on building a lightweight ecosystem where their models can be used by developers and companies for private deployments. Their strategy is to provide a framework that others can build upon, allowing for a diverse range of applications.
IV. Commercialization: Diverse Revenue Models
The approaches to generating revenue also differ significantly:
- ByteDance: Generates higher annual revenues (estimated at $4 billion) from various sources, including MaaS (Model as a Service), GPU cloud leasing, Seedance video generation services, and premium DouBao subscriptions. However, the high costs associated with research and development across all business segments pose challenges to profitability.
- DeepSeek: Generates lower annual revenues ($400–500 million, one-tenth of ByteDance’s amount), primarily from API fees. They charge only for enterprise private deployments and offer free apps for individual users. DeepSeek claims that their API pricing model can recoup costs in about 10 months, with shorter payback periods due to recent price increases.
In summary, ByteDance is large and comprehensive but more resource-intensive, while DeepSeek is smaller and more streamlined.
V. Industry Implications: Divergence as a Driver of Progress
This competition is reminiscent of the rivalry between OpenAI (closed-source, super-app approach) and Anthropic (open-source, streamlined approach), as well as that between Intel and AMD or Boeing and Airbus. Different approaches can spur industry development:
- ByteDance’s closed-source strategy is suitable for large companies that can integrate existing technologies, while DeepSeek’s open-source model suits smaller teams that can thrive in a competitive market.
- The two companies are likely to learn from each other; for example, ByteDance may eventually make its models open-source, and DeepSeek could improve its visual understanding capabilities.
Ultimately, the AI industry benefits from both types of companies—those with broad reach (like ByteDance) and those with deep expertise (like DeepSeek). Together, they will contribute to the advancement of AGI in China.
VI. The Common Goal: AGI
Despite their differences, Zhang Yiming and Liang Wenfeng share a common vision: achieving AGI. Zhang Yiming aims to build foundational barriers for AI, while DeepSeek’s six-stage roadmap focuses on enhancing model capabilities. Both companies believe in long-term commitment and the potential of AI to benefit humanity. This competition is not about elimination but about coexistence and mutual growth, as the future of AI requires diverse perspectives and innovations.
Conclusion
While their paths differ, both ByteDance and DeepSeek are committed to realizing the dream of AGI. Their approaches complement each other, much like how iOS and Android have jointly propelled the development of mobile technology. Together, they will drive China’s AI industry towards achieving AGI.