news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

On August 2, the 12th Internet Security Conference (ISC.AI 2024) opened, the conference was themed “Building a safeLarge Model, leading a new revolution in the security industry" with a deep focus on the two major areas of security and AI. At the scene, we interviewed360President of ZhinaoZhang Xiangzheng, and had an in-depth exchange on topics such as 360 Smart Brain’s understanding of security issues and product planning.

Zhang Xiangzheng introduced that the Zhinao team is responsible for general model capabilities and special capabilities for core business scenarios, such as intelligent reading, browser, text summary, video/subtitle summary, mind map generation, search word guessing, intent recognition and summary, etc. At the same time, the Zhinao team is also responsible for the work of combining API and AI security with large model capabilities.

The theme of this conference is "Building a Safe Big Model". When talking about the topic of security, Zhang Xiangzheng said that security issues are divided into several parts. The first part is the system deployment level, which is related to traditional network security, including the Agent framework, vector database, and PyTorch framework, which itself has potential vulnerability risks; the second part is whether the generated content meets regulatory requirements, including the alignment of values; the third part is the output result error. Suppose 10% occurs, I don’t know when this 10% occurs, which often leads to whether the generated results should be trusted. The fourth part is related to the Agent framework. The big model is only the scheduling hub, which can access many third parties or many databases within the company or enterprise. If automated operations are performed, it may affect the security of other systems. For example, when used for embodied intelligence, some dangerous actions may be performed.

In terms of security, the difference between ToB and ToG is also quite obvious. Zhang Xiangzheng introduced that the tolerance for security issues is different. For example, in government departments, training data needs to provide a private deployment solution, and all fine-tuning and incremental training must be done in the internal network environment of government agencies. In the ToB scenario, a typical example is education. There is a poem "The Moon Shines on My Bedside", and there are two versions online. The product manager was very panicked, wondering why the introduction of such a famous poem is different from the textbook. Children do not have good discrimination ability, which leads to very high requirements for the credibility of the results.

Talking about the topic of small models, Zhang Xiangzheng pointed out that we have explored the application of small models on computers and tested related solutions internally, but there are too few laptops that meet the requirements and the quantity is not that large. In addition, we have also considered the acceptance, and we believe that there is still a long way to go for the commercialization of small models.

Regarding the hotly debated topic of AI search, Zhang Xiangzheng responded that compared with traditional search, AI search is more distinctive in that it can perform multi-step reasoning or multiple keyword searches, breaking down complex problems, and automatically performing multi-step searches in steps before fusion, or performing another search on the results after the first step of reasoning, and then handing them over to the big model for fusion. In the future, if AI search capabilities become stronger and stronger, user habits will migrate, but it cannot be said to be a completely new product form, which is related to user usage habits. (Dingxi)

This article is from NetEase Technology Report. For more information and in-depth content, follow us.