In Depth: China Debates Pros and Cons of Open-Source AI Models
Listen to the full version
When discussing the development of large language models (LLMs) Zhou Hongyi, CEO of Chinese cybersecurity firm 360 Security Technology Inc., dismissed claims that open-source LLMs are inferior to closed-source ones, calling such assertions “nonsense.”
Zhou’s comments, which were delivered during the 27th Harvard College China Forum in April, were widely seen as a response to Baidu Inc.’s CEO Robin Li, who had previously said that it did not make sense for LLMs to be open-source and that closed-source LLMs would gain an expanding edge in capabilities as technology advances.
Download our app to receive breaking news alerts and read the news on the go.
Get our weekly free Must-Read newsletter.
- RELATED
- DIGEST HUB
Digest Hub
Back
- Zhou Hongyi criticized claims that open-source large language models (LLMs) are inferior to closed-source ones, countering opinions from Baidu's CEO Robin Li.
- Examples include Alibaba and Beijing Academy of Artificial Intelligence developing open-source LLMs, promoting community and cost-effective benefits for developers.
- Baidu maintains a closed-source stance, offering toolkits like AppBuilder, ModelBuilder, and AgentBuilder, while concerns remain about dependency and data sharing with the tech giant.
The development of large language models (LLMs) remains a contentious topic among major Chinese tech leaders. Zhou Hongyi, CEO of 360 Security Technology Inc., criticized claims that open-source LLMs are inferior to closed-source ones as “nonsense” during the 27th Harvard College China Forum in April, indirectly responding to Baidu Inc.'s CEO, Robin Li. Li believes that closed-source LLMs will prevail as technology advances, while Zhou advocates for open-source LLMs [para. 1][para. 2][para. 4].
Baidu is one of the early Chinese companies to offer a local alternative to OpenAI’s ChatGPT [para. 3]. The debate extends to the advantages and disadvantages of making LLM source codes available to external developers for creating specialized applications [para. 4]. Contrary to Baidu's stance, companies like Alibaba and Beijing Academy of Artificial Intelligence (AI) along with startups such as Beijing ModelBest and 01.AI are supporting open-source LLMs [para. 5][para. 6][para. 7].
A real-world application of open-source LLMs was demonstrated by Liu Ming, a designer in Zhejiang province, who found it cost-prohibitive and risky to use closed-source models. He adopted an open-source LLM integrated with Nvidia GPUs and open-source software from GitHub for his design tasks. Eventually, Liu deployed his system on Alibaba's open-source Qwen-72B model, which he found to match the accuracy of closed-source alternatives [para. 8][para. 9][para. 10]. This success highlighted the importance of free resources provided by open-source models [para. 10].
Similarly, Associate Professor Xue Dong from East China University of Science and Technology used open-source LLMs to develop MindChat, a chatbot offering psychological services. After assessing the costs and data privacy issues with closed-source models, Xue's team utilized LLMs developed by Alibaba, Beijing Baichuan Intelligent Technology, and a team including SenseTime and the Shanghai AI Laboratory [para. 12].
The ecosystem around open-source LLMs has been bolstered by platforms like Alibaba’s ModelScope, which offers over 4,000 deployable AI models and has engaged 4.4 million developers worldwide. Zhou Jingren, CTO at Alibaba Cloud, emphasized that their strategy is to build an open ecosystem rather than keeping LLMs proprietary [para. 14][para. 16]. Despite open-sourcing their models, Alibaba also aims to promote their cloud services [para. 18].
In contrast, Baidu’s CEO Li argues that transforming open-source LLMs into profitable ventures is challenging due to the significant investment required. Baidu has introduced three AI development toolkits—AppBuilder, ModelBuilder, and AgentBuilder—to facilitate creating AI applications using their closed-source LLM ERNIE [para. 19][para. 20][para. 21]. However, reliance on Baidu's platforms raises concerns about data ownership and accessibility to Baidu's APIs [para. 23][para. 24].
Acknowledging these nuances, market observers believe that effective AI applications necessitate a blend of computing resources and comprehensive development toolkits [para. 27]. Chen Ran, formerly of GitLab China, identified a business opportunity in this arena. He now leads OpenCSG, a platform similar to Hugging Face, providing tools for LLM development and deployment. OpenCSG also includes a computing resources trading platform and has attracted over 500,000 registered users, expected to generate significant revenue [para. 28][para. 29][para. 30].
Ultimately, the discussion surrounding open-source versus closed-source models continues to evolve, guided by the availability of resources and development tools that cater to the needs of AI developers [para. 27].
- 360 Security Technology Inc.
- 360 Security Technology Inc. is a Chinese cybersecurity firm led by CEO Zhou Hongyi. The company is involved in the development of large language models (LLMs) and is an advocate for open-source LLMs, opposing the view that closed-source models are superior.
- Baidu Inc.
- Baidu Inc. is a Chinese tech giant that insists on keeping its large language models (LLMs) closed-source. CEO Robin Li believes open-source LLMs are less practical for developers due to high costs and data sensitivity. To aid developers, Baidu offers toolkits like AppBuilder, ModelBuilder, and AgentBuilder, which rely on its proprietary ERNIE LLM. Baidu also provides an all-in-one LLM development platform integrating both closed-source and open-source models.
- Alibaba Group Holding Ltd.
- Alibaba Group Holding Ltd. has made many of its large language models (LLMs) open-source. It launched ModelScope in November 2022, an open-source Model-as-a-Service platform offering over 4,000 AI models to 4.4 million global developers. Alibaba Cloud supports this ecosystem with extensive free computing power, promoting its cloud services. The company's CTO, Zhou Jingren, emphasizes building an open ecosystem for all developers.
- Beijing ModelBest Intelligent Technology Co. Ltd.
- Beijing ModelBest Intelligent Technology Co. Ltd. is a startup focused on building ecosystems for open-source large language models (LLMs). The company is part of a growing movement in China supporting open-source LLMs, which contrasts with the closed-source approach adopted by tech giant Baidu.
- Beijing Baichuan Intelligent Technology Co. Ltd.
- Beijing Baichuan Intelligent Technology Co. Ltd. is mentioned as a developer of open-source large language models (LLMs). In July 2023, a team led by associate professor Xue Dong at East China University of Science and Technology used its open-source LLM to develop a data-driven chatbot called MindChat, indicating its active role in the open-source AI community.
- SenseTime Group Inc.
- SenseTime Group Inc. is involved in developing open-source large language models (LLMs). In July 2023, a team including SenseTime and the Shanghai Artificial Intelligence Laboratory contributed to the launch of MindChat, a data-driven chatbot providing psychological assessment, diagnosis, and treatment services.
- Shanghai Artificial Intelligence Laboratory
- The Shanghai Artificial Intelligence Laboratory collaborated with SenseTime Group Inc. to develop an open-source LLM, which was utilized by a team led by associate professor Xue Dong at East China University of Science and Technology for creating the MindChat chatbot.
- November 2022：
- Alibaba launched ModelScope, an open-source Model-as-a-Service platform.
- June 2023：
- Liu Ming chose to use a Chinese-developed open-source LLM with Nvidia Corp. graphics processing units.
- middle of 2023：
- Xue Dong’s team picked some open-source LLMs to develop a data-driven chatbot, MindChat.
- July 2023：
- Xue Dong’s team launched three versions of MindChat developed based on open-source LLMs.
- end of 2023：
- Liu Ming formally deployed his inference engine on Alibaba’s open-source Qwen-72B model.
- As of now：
- OpenCSG has been operating for around four months, amassing over 500,000 registered users.
- April 2024：
- Zhou Hongyi delivered his comments dismissing claims that open-source LLMs are inferior to closed-source ones at the 27th Harvard College China Forum.
- GALLERY
- PODCAST
- MOST POPULAR
Sign in with Google
Sign in with Facebook
Sign in with 财新