Has Ali Tongyi become the "Whampoa Military Academy" of AI?

Since last year, major technology giants have adjusted their strategies and stepped up their efforts to recruit talent: Tencent's "Qingyun Plan" attracts top talent with "industry top 0 salary" and "no upper limit on job level", and plans to add 28,000 internship positions to enrich reserve forces; ByteDance has launched the "Top Seed Plan", which specifically opens positions for top PhDs in cutting-edge fields such as large models and multimodality. Compared with new forces, experienced AI core backbones have become the top priority for all parties in this competition.

The flow of talent at Alibaba's Tongyi Lab is particularly eye-catching during this wave.

Recently, the resignation of Bo Liefeng, the head of the applied vision team of Alibaba Tongyi Lab, has caused a round of heated discussion. This is the third time that Alibaba's AI core department has lost a key figure, following Yan Zhijie, the head of the voice team in February and Zhou Chang, the head of large model technology in August 2024.

On the one hand, the Qianwen series of models dominate the list, but on the other hand, the leading figures of the research team have left several times. These core talents with Alibaba's technical genes have either joined other large companies or set up their own companies, becoming new variables in the industry's secret war. Tongyi has a bit of the flavor of the new generation of "AI Huangpu Military Academy".

Before Tongyi, the industry-recognized AI "Huangpu Military Academy" was Microsoft Research Asia (MSRA). The two do have many similarities in the path of cultivating top talents. Since its establishment in 1998, MSRA has provided many key figures in the field of AI in China and even the world, including Kai-Fu Lee, Harry Shum, Zhang Hongjiang, Lu Qi, and Wang Jian. MSRA is the first generation of organizations supported by powerful technology companies and devoted to long-term technology research.

The reason why Tongyi can become a new "gold mine" for talent export lies in its strong parent support and long-term strategic layout.

Back in October 2017, Jack Ma announced the establishment of the DAMO Academy at the Yunqi Conference, with an initial investment of 100 billion yuan, determined to build it into a "research institute dedicated to exploring the unknowns of science and technology and driven by human vision". Under the leadership of then-dean Zhang Jianfeng (nickname: Xingdian), five major laboratories such as machine intelligence and data computing were quickly rolled out. Behind this move is the era ambition of Internet giants to try to replicate the new generation of "Microsoft Research Asia". Jin Rong, Si Luo, Hua Xiansheng, Wang Gang and other scientists who are famous in their respective fields, nicknamed "Thirteen Sweeping Monks" internally, laid the initial talent foundation for the DAMO Academy.

Compared with the AI ​​teams of other large companies, Tongyi Lab, which was incubated by DAMO Academy, has established advantages from the beginning in terms of top talent density, cutting-edge technology exploration and systematic research system. Compared with ByteDance today, which has set up an independent Seed large model team by drawing from internal resources and supplemented by external recruitment, Tencent has made improvements on the path of gain based on the existing AI Lab. Objectively speaking, the layout of Tongyi Lab is indeed earlier, which has brought more abundant accumulation.

DAMO Academy was one of the first teams in China to sense and fully invest in the big model trend. Before OpenAI's GPT series exploded around the world, DAMO Academy had already started big model research and launched early exploratory models such as M6. This strategic path of "early follow-up and continuous investment" enabled Tongyi to come up with competitive products more quickly in subsequent big model competitions.

The subsequent evolution of Qwen further verified the importance of Tongyi's earlier accumulation. Today, it has not only achieved results comparable to or even surpassed the world's top models (such as GPT-4 and LLaMA series) on multiple authoritative benchmarks, but more importantly, Alibaba had already begun to use open source as an important technical competition strategy during the period of DAMO Academy, and even made many commercial attempts based on open source, which made Alibaba's decision to open source multiple versions of the Qwen series (from 1.8B to 72B, and even closed-source versions with larger parameters) smoother.

Many developers have reported that Qwen's performance in code generation, long text understanding, multi-round dialogue, etc. is impressive. This "vote with your feet" from front-line developers indirectly confirms the accumulation of the Tongyi team in model architecture design, training data construction, and optimization and tuning.

In addition to model performance, a major feature of the Tongyi team is that its research direction covers a wide range, covering almost all key dimensions of AI technology - from natural language processing to computer vision, from speech recognition to multimodal fusion, and has accumulated technical talents in each segment. These earlier technical research and development accumulations that are more closely integrated with other real business scenarios of Alibaba have made Tongyi a talent pool that has been verified in actual combat. When ChatGPT suddenly broke out, these truly experienced talents became popular.

Unlike when MSRA was born, when China's AI industry was still in its infancy and focused more on basic research and talent enlightenment, Tongyi grew up in a stage where the competition for AI industrialization has become increasingly fierce. This determined that Tongyi's model research and development was more closely centered around market demand and commercialization from the very beginning. Therefore, the technical talents that flow out of Tongyi often have "plug-and-play" practical capabilities and rich engineering experience. Once they join the core competitor camp or set up their own business, they will undoubtedly become potential variables that leverage the existing industry structure.

For example, Zhou Chang, as one of the technical leaders of Tongyi Qianwen big model, was deeply involved in the core links of Qwen series models from architecture design to training optimization. The discussion and subsequent disputes over competition caused by his departure at that time came from the prediction that this scarce experience might bring changes to the competitive landscape. The understanding of the big model technology stack, the know-how of training, and the judgment of the direction of industry technology development are valuable assets for any company committed to the research and development of big models. As the former head of the Intelligent Voice Laboratory of Alibaba Damo Academy, Yan Zhijie has more than 20 years of R&D experience and deep technical accumulation in the fields of speech recognition, speech synthesis and natural language processing; Bo Liefeng, as the head of the applied vision team of the former Tongyi Laboratory, has rich practical experience in the field of computer vision, especially in industrial vision, 3D vision, and multimodal content understanding. According to public information, the technical achievements of his team have been applied in projects such as Alibaba Cloud Industrial Brain and City Brain. These are very different from the Whampoa Military Academy during the MSRA period. Today, Tongyi has left key figures who can directly affect the industrial landing pattern.

For Alibaba, the outside world is worried about the chain reaction. Will the loss of these "general-level" talents further affect the subsequent flow of some members of their original team, or affect the research and development direction and progress of key projects? After Bo Liefeng left, the leadership handover of the Tongyi Applied Vision Team has become a new focus of attention, and the official has not yet announced a clear successor.

But from a broader industry perspective, the "Whampoa Military Academy" will appear in every period of technological innovation, which has become a necessity. Behind this is the key role of experience spillover in the overall maturity of the industry. The reasonable flow of talents itself is a catalyst for technological innovation and ecological prosperity. The evaluation of an algorithm engineer is quite representative: "People who come out of Tongyi are very valuable when looking for jobs in Silicon Valley." This indirectly confirms the market recognition of Tongyi talents.

For large companies or startups that try to quickly make up for their shortcomings by "poaching" core talents or even copy a "small Tongyi", poaching is not an easy road. Tongyi's unique AI research atmosphere based on the soil of "research institute", as well as the massive real business scenarios, strong computing support and long-term strategic investment provided by Alibaba Group, are the key factors for its success.

This makes the competition more complicated. Zhou Jingren, CTO of Alibaba Cloud and head of Alibaba Tongyi Laboratory, once emphasized: "Now model research and development is not only an innovation at the model and algorithm level, but also a long-term system project. It requires patience and firm investment. Doing it today and stopping tomorrow will have a great impact on the team. Alibaba's investment in Tongyi Qianwen is very firm." This basically shows that it is difficult to build real core competitiveness by poaching alone. What needs to be done behind it is a complex and huge system project.

The outflow of talent from rivals may bring tactical advantages to the recipient in the short term, but in the long run, changes in the industry landscape depend more on the clarity of each player's own strategy, the thoroughness of execution, and the ability to build a sustainable innovation system. For each company, the real test is not how to simply "retain" or "grab talent", but how to consolidate and enhance their unique culture and competitiveness in continuous technological innovation and an open ecosystem. This game is much more complicated than a "war for talent".