Author | Liu Baodan

Editor | Huang Yu

The rapid iteration of AI large model technology has raised higher requirements for AI governance, and Alibaba has begun to think more about how to improve AI governance.

Recently, at the seminar on the development and governance innovation of large model technology, Qian Lei, Vice President of Alibaba Group, stated that in order to effectively govern large models, improve their controllability, and reduce application risks, Alibaba has comprehensively deployed security measures around the entire lifecycle of large model research and development and application.

First, enhance the endogenous security capabilities of the model. Emphasize source control during data collection and processing, conduct strict data screening and safety labeling; during the model training and design phase, add secure corpora; conduct comprehensive security assessments before the model goes live to ensure the qualification rate of various risk items in generated content.

Second, strengthen external fence capabilities. After deploying the model for application, further enhance the security of the large model in application scenarios through external fences and verify through daily inspections whether the results can meet expectations and whether reliability, trustworthiness, and usability can meet Alibaba's requirements.

Alibaba has also jointly released the "Report on the Development and Governance Practices of Large Model Technology" with the China Electronics Standardization Institute, marking the third consecutive year that Alibaba has published research results related to AI governance, further deepening Alibaba's experience summary in technology research and governance.

Yuan Yuan, the director of the Alibaba Research Institute, believes that the governance approach of AI large models involves analyzing the three core development engines—computing power, algorithms, and data—and providing a development path suitable for China. For instance, with data, synthetic data can help models break through current capacity limits while also reducing security concerns, provided that it aligns with the characteristics of the real world.

Since the end of 2022, when ChatGPT gained global popularity, AI large models have become a field where tech giants are eager to invest, and Alibaba is a core participant in this arena.

In terms of AI large models, Alibaba insists on the collaborative development of cloud and AI, evolving cloud from IT and DT infrastructure towards AI infrastructure. According to Qian Lei, Alibaba will continue to increase investment: first, to enhance the capabilities of self-developed foundational models of the Tongyi series, and second, to reduce the cost of model application development by improving AI infrastructure and related supporting systems.

The API call price for "Tongyi Qianwen" has decreased by 97% over the past year. At the same time, Alibaba adheres to full-scale and full-modality open-sourcing, driving the transformation of AI capability into productivity across various industries. Currently, more than 0.3 million enterprises have accessed the Alibaba Tongyi Qianwen large model.

Data from the Hugging Face Community shows that derivative models developed based on the "Tongyi" model have surpassed 0.08 million, attracting over 6 million AI developers, with downloads on global open-source platforms exceeding 0.1 billion times.

In Alibaba's view, while advancing cutting-edge technologies for large models to promote AI applications, it is also crucial to pay high attention to AI security and to view the security issues brought by large models responsibly.

However, AI large models are still in a rapid iteration stage, and the key to solving safety and governance issues lies in balancing development and risk, much like changing parts on a highway, which requires extreme wisdom and professionalism.

Greater challenges lie ahead. As AI large models truly enter people's lives, issues concerning AI security and governance will surface more problems, which all AI-related companies must face.

Never has a technology been able to match human intelligence like AI, but this also means that once AI is misused, its destructive power is equally worth being cautious about. The path of AI governance has just begun, and Alibaba has already attempted to provide its own answers.