2024-08-06
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
Produced by NetEase Technology's "Attitude AGI" dialogue column
Author: Zong Shuxian
Editor: Ding Guangsheng
Twelve years have passed since the first Internet Security Conference. This year's Internet Security Conference has upgraded its conference name from ISC to ISC.AI, and Internet security has officially entered the AI era. "In the AI era, cyber attacks have a greater impact on personal safety and society. The Microsoft blue screen incident reminds the public that digitalization and intelligence have made the Internet world particularly vulnerable."360Group Founder and CEOZhou Hongyiexpress.
He pointed out that magic must be fought with magic,Large ModelTo solve the security problem of big models, this is the starting point of "AI + security". At present, 360 is actively laying out the two main business lines of security big models and "search + big models", providing free services to small and medium-sized enterprises and individual users, and building free security big models. Zhou Hongyi said in an interview that 360 does not intend to directly get involved in the fierce competition of super big models, but chooses to focus on the research and development and application of security big models. 360 will cooperate with 15 top big model companies in China, open up its own scenarios, and promote the popularization and application of big models.
In other words, 360 is positioned to do three things well: First, as a security vendor, it must build a big security model. On the one hand, it must use the big security model to solve security problems, and on the other hand, it must solve various security challenges encountered in the development of the big model. Second, 360 itself has 2C user scenarios, such as 360 browser, 360 search, 360 children's watches and cameras, and it must use big models to empower its products. Third, it must promote the popularization of big models, so that enterprises and governments can train and use their own professional big models very cheaply.
Zhou Hongyi admitted that the views in the industry are converging. It is no longer about models, but about applications. When the big model becomes "invisible", the technology will be mature and it will have to work "behind the scenes".
Regarding the hotly discussed AI search opportunities, 360, as an important player in the search market, Zhou Hongyi also gave his own judgment. He pointed out that at present, whether Search or Chat is better, it is possible that they are evenly matched, or one may be replaced by the other in the future. AI search is an important scene entrance that no one can miss. "OpenAI's SearchGPT proves that OpenAI has also been shaken."
"But it still has many problems to solve, such as the easy generation of 'hallucinations', 'data poisoning', and challenges in commercialization. Advertisements cannot be placed in it, and it is unrealistic to ask users to pay. These are common problems faced by everyone."
The following is a selection of transcripts from NetEase Technology's conversation with Zhou Hongyi (with adjustments that do not change the original meaning):
Q: Some people believe that the market will be dominated by a few super large models in the future, while others believe that it will be dominated by many small vertical models. What do you think about the direction of this industry landscape and how will 360 position itself?
Zhou Hongyi:I think both paths will exist.Large companies with financial resources and capabilities like OpenAI certainly want to develop super general artificial intelligence. However, with the emergence of open source big models represented by Lama3 and Qianwen 2, big models will not be monopolized by only 30 companies. Big models will become important functional nodes and functional components in the digitalization of all enterprises and governments. Therefore, in the future, there will be many vertical industry big models or vertical big models within enterprises and governments. At the same time, there is another trend of big models. The so-called thousands of households, plus everyone's mobile phone, everyone's computer, everyone's car, and every family'srobotThere will always be big models, so big models will be everywhere and work collaboratively in the future.
360 positions itself to do three things well: First, as a security vendor, it needs to build a big security model. On the one hand, it uses the big security model to solve security problems, and on the other hand, it needs to solve various security problems that will be encountered in the future during the development of the big model. Second, 360 itself has many 2C user scenarios, such as 360 browser, 360 search, 360 children's watch and camera. We will use professional big models to empower products. Third, promote the popularization of big models, so that enterprises and governments can train and use their own professional big models at a very low cost.
In a word, large models cannot be made by a few people, but must be introduced into thousands of households and industries. Models must be made small before they can be used in personal products.
Q: From your perspective, which existing industries will be disrupted or transformed by the development of the large model industry?
Zhou Hongyi:I think the big model will bring about huge changes to all industries, but it is not necessarily called subversion. I think it would be better to use the concept of "reshaping". More importantly, it can directly improve efficiency and reduce costs.For example, we divide the capabilities of big models into several levels. The basic capabilities are more reflected in personal use, such as office, education, learning, language processing, graphics generation, and short video generation. Second, there are many business capabilities within the enterprise, such as helping enterprises achieve knowledge management, intelligence analysis, and business decision-making. Third, big models will also be combined with autonomous driving, robots, traditional basic scientific research, biomedical pharmaceuticals, and agricultural industries, which will produce new explosive effects.
Q: How can we become an all-around champion like ChatGPT, and how can we face up to our own shortcomings?
Zhou Hongyi: Frankly speaking, if we follow the path of OpenAI and develop super artificial intelligence, I don’t think every company needs to do it. China can’t produce 100 companies because we don’t have enough computing power. We think that China, led by Alibaba, can follow up on the basis of Qianwen and Baichuan through the open source community, and the follow-up will be very fast.
But this is not the same as the professional big model I mentioned. The professional big model I mentioned is to promote big models into enterprises and governments, and promote industrial revolution, not for individuals. This only requires my professional ability to be improved and combined with the scene. Therefore, the professional big model will develop very richly in the future. In the future, there will definitely be more than 10,000 professional big models playing a role in the Ten Thousand Model War.
Recently, the industry has reached a consensus that we should stop focusing on models and focus on applications. In fact, in the future, the big model technology will be "invisible", and the technology will be mature. The big model is more like an electric motor, which determines the development direction according to different fields and works behind the scenes.
Question: "Building a security + AI, an innovative ecosystem of security + AI can accelerate upgrades." Can you tell me more about this?
Zhou Hongyi: AI+security is mainly about building a big security model to solve two problems: one is to use big model capabilities and AI capabilities to improve the pain points and rigid needs of today's traditional network security. To put it bluntly, the biggest pain point of traditional network security today is the lack of professional network security talents. They cannot keep up with the requirements of the times in terms of attack detection, attack resistance, situation handling, and automated operations. Now we can fill the talent gap by training big security models and using AI to automatically solve the problems of attack detection, attack resolution, and operational handling. Second, many countries are now aware of the problem of AI security governance. With the development of AI, it will bring many new security threats and challenges beyond network security and data security. It may still be content security now, but the human security issues that will arise in security in the future are now unknown. From now on, someone has to solve the problem of AI security.Therefore, we need to use magic to deal with magical thinking, and use big models to solve the security problems of big models. This is the starting point of our AI+security.
Q: Some people say that AI search will replace traditional search earlier. How can 360's AI search be real, reliable and easy to use?
Zhou Hongyi: AI search may represent a real breakthrough point. Even OpenAI has started to work on AI search. Using chat to solve problems has also caused doubts today. For example, when two of us chat, sometimes it is very efficient, but sometimes it is difficult to understand, especially when we have to write complex prompt words, which is difficult for me. Therefore, search must be combined with AI.Everyone can use search. It does not require long input, just a few keywords will do. AI search uses a big model to guess the search intent, decomposes the information into multiple steps of search, retrieves the results from multiple searches, and then aggregates the results. This is the role of the big model.
Big models + search can easily lead to hallucinations, but now we have a search + big model, and the data is mainly searched, which is equivalent to doing search enhancement and knowledge alignment. Of course, there will be problems here, such as someone poisoning the data, and then we need to add content recognition to identify some absurd content. There are also big models here, and the big models are combined to do these things, and the problems you mentioned can be solved.
In addition, many AI searches are relatively simple, just rewriting the search results. What we are doing now is that each search intent may be completed by an intelligent agent for you, calling multiple large models to complete complex tasks.
AI search is becoming more and more complex. The biggest challenge is to solve the commercialization problem. It is impossible to get users to pay for the search without advertising, because the search is free. This is why Baidu and Google are hesitant, and why GPT is competing with Google.
So the 360AI search is not based on a big model, because a big model simply cannot do the job.To make an inappropriate analogy, this is more and more like what we call an intelligent agent framework orWorkflowSoftware. Therefore, the industry's recognition of big models is a process, and the security industry is even slower to realize it. Now, the leading big model giants in China are no longer just talking about big models. Knowledge determines the capabilities of big models, and whether the capabilities of big models can be well utilized does not depend on the intelligent agent framework. After the intelligent agent framework generates the big model capabilities, whether it can bring direct help to the business depends on integrating it into the workflow.
Q: What do you think is the development trend of AI search? There is a voice saying that AI search+Agent+Work Flow is a trend. What do you think?
Zhou Hongyi: To make good use of AI now, it is definitely not enough to rely on a big model alone, because no matter how powerful the big model is, it cannot complete the task without an agent, an intelligent body, and a WorkFlow. The big model can enhance tools, knowledge, reasoning, and memory. The most important thing is to use the agent framework to solve the problem that the big model can only think quickly. People have two systems, one for fast thinking and the other for slow thinking.It is actually unrealistic to push the big model of fast thinking to the forefront and try to solve all problems. The most important thing about the Agent framework is to solve the problem of slow thinking. No matter what AI is doing, whether it is AI search or search + AI, these functions are indispensable.
The reason why I think that search + AI is better than big model + search enhancement in terms of user experience is thatFirst, users have already developed the habit of searching and do not need to be educated, but it may be difficult to communicate using the prompt words of the big model. Second, the expectation of search is different from that of the big model chatbot. The expectation of search is to find answers. Third, the difficulty is different. If you do not define the scenario and use part of the big model's capabilities to try it out, you can improve the search experience very well.
Our AI search traffic is growing very fast. Now 360AI search has over 10 million monthly active users. I think the big model is very powerful, but it must be combined with various scenarios to empower each scenario. This is the best user experience. Recently, I noticed a phenomenon that the US startup Perplexity is growing very fast. OpenAI's SearchGPT proves that OpenAI is also shaken. Which is better, Search or Chat? It is possible that they are evenly matched, and it is possible that one will be replaced by the other in the future, butAt least it shows that AI search is an important scene entrance that no one can miss.
Q: What do you think about large models being free?
Zhou Hongyi: The general big model is not free. The cost of the general big model is too high, and it is provided in the form of public cloud. With such a high cost, Token is already very cheap, but it certainly cannot be free. The office suite we launched is supported by the reasoning ability of the general big model, which also has costs. There is a free use period, but it is not free.
There are two conditions for the free security big models I mentioned: First, what is provided is not general big data, but six professional big models. Users who have purchased my 360 enterprise security products, such as my security brain, my terminal security, and my security device services, can upgrade for free. Second, these special big models are not publicly deployed on the cloud, but privately deployed within the user's enterprise, so there is no computing cost, so of course they can be free. The two products are different, one is a toB product, and the other is a toC product.
Q: I would like to ask about the input-output ratio. I wonder if the specific mode of cooperation with these big models is resource exchange or specific price. With the increase in user growth or corresponding terminal download volume, will the ROI of this matter show a significant growth in the future? What are the considerations for choosing this cooperation model?
Zhou Hongyi: First of all, each company has a quotation for tokens, and the fee for each thousand words on the token is already very low. We pay the amount of tokens we use. They support me with their capabilities, which is equivalent to the fact that my base model platform is very powerful. I will try my best to bring them more traffic and more users, and find ways to create more products so that users can use them more times every day.
The current situation of the domestic big model is that each company has both advantages and disadvantages, and there is no overwhelming lead. If you use Baidu, Alibaba, and Doubao, you may form the strongest team by combining their capabilities. Does China need dozens of manufacturers to develop a general AGI? I think it is impossible, so I don't support it.
So my current thinking is that everyone is training models, but without expanding applications, the model cannot be used no matter how powerful it is. So don't always say that you should encourage application developers to use APIs to do shells. Shells can be deep or shallow. The office membership we do, including browsers, is deep shells. We can't call API calls shells, as it seems very low-level. They make APIs just to hope that someone will use them.
About Attitude AGI
NetEase Technology has launched a series of dialogue columns "Attitude AGI". In the past three years, AI has been changing rapidly, the global technological order is being restructured, and the road to AGI may be approaching quietly. This column is themed on AGI and will talk to 100 AI experts, entrepreneurs, and investors, trying to clear the fog and witness the arrival of the AGI era with everyone. The tenth episode is a dialogue with Zhou Hongyi, the founder of 360.
Previous reviews:
The first issue of "Attitude AGI":Dialogue with Kai-Fu Lee: Big Model Entrepreneurship Has Been Running Wild for One Year, and the Gap Between China and the United States Has Been Narrowed. I Have Not Cashed Out for Ten Years
The second issue of "Attitude AGI":Dialogue with Wang Xiaochuan: We will not follow the price war of large models
The third issue of "Attitude AGI":Dialogue with Dai Wenyuan: The price war of large models does not solve the core problem
"Attitude AGI" Issue 4:Wang Zhongyuan, President of Zhiyuan Research Institute: GPT4 is not the end of domestic large models
Attitude AGI Issue 5:Dialogue with Zhu Xiaohu: In 5 years, there will be no independent large-scale model companies because there is no business model
Attitude AGI Issue 6:Dialogue with Wu Shichun of Meihua Venture Capital: It is very difficult for the "Five Tigers of China's Large Models" to succeed
Attitude AGI Issue 7:Dialogue with Yueran Innovation CEO Li Yong: How to compete with giants in large-scale model applications
Attitude AGI Issue 8:Dialogue with Zhang Peng, CEO of Zhipu AI: Video generation is the only way for AGI, and super apps will come "unknowingly"
Collect and organize the latest news information and make it available for free for everyone to read
