Huaxia Times reporter Lu Xiao reports from Beijing
AIGC has been burning this fire for over two years. Compared to early roll parameters, roll applications have more players in the second half and competition is more intense.
On November 12, Robin Lee, the founder of Baidu, believed at the 2024 Baidu World Conference that agents are the most mainstream form of AI applications and will soon usher in an explosion point. In fact, the productivity revolution triggered by AIGC has attracted attention from the outside world for its research and development progress and application implementation. It not only affects the performance of large model companies, but also their valuation.
At present, big model manufacturers are increasing the popularity of AIGC applications, and making big models accessible to ordinary intelligent agents has become their competitive focus.
Intensify the popularization of applications
Enhancing trust is a necessary path for the popularization of large-scale model applications.
If you were to ask me, what has been the biggest change in the AI industry over the past 24 months? My answer must be that big models have basically eliminated illusions, and their accuracy in answering questions has greatly improved, making AI usable and trustworthy from 'serious nonsense' Robin Lee said so on November 12.
Industry insiders believe that hallucinations have always been one of the limitations that hinder the widespread application of AI big models, and the key to solving this problem lies in RAG technology, namely retrieval enhancement, when communicating with reporters from Huaxia Times. However, the development of multimodal technology has increased the difficulty of solving the problem of model illusion. Robin Lee also believes that the RAG at the text level has been done well, basically eliminating the illusion of the big model; But in terms of multimodal aspects such as images, the combination with RAG is not enough, which is also the direction that Baidu hopes to focus on breaking through.
According to him, the iRAG (image based RAG) technology developed by Baidu for enhanced retrieval can combine Baidu's billion level image resources with basic model capabilities, generating images with an overall effect far exceeding the native system of iRAG, removing the machine flavor.
Reducing the entry threshold while making big models more trustworthy is also one of the prerequisites for accelerating the popularization of big model applications. Baidu's codeless tool "Miaoda" has become the latest development in the industry. Robin Lee said that Seconds enables everyone to have the ability of a programmer, and can make applications when they can speak, which greatly improves the working efficiency of human beings. "We will usher in an unprecedented era in which people can make money only by thinking."
Baidu took the lead in launching ERNIE Bot in the domestic market last March. Wang Haifeng, chief technology officer of Baidu, disclosed in his speech on November 12 that the number of users of ERNIE Bot has reached 430 million. Robin Lee also disclosed on the same day that as of the beginning of November, the daily average adjustment volume of Baidu Wenxin's big model had exceeded 1.5 billion, an increase of 7.5 times compared with May, and an increase of about 30 times compared with the 50 million times first disclosed a year ago. He believed that this steep growth curve represented the outbreak of big model applications in China in the past two years.
From Baidu's own perspective, AI is also playing a greater role in its various businesses. According to Baidu's Q2 earnings call this year, the proportion of revenue contributed by AI to intelligent cloud has increased to 9%, an increase of two percentage points compared to the previous quarter. In addition, 18% of Baidu's search results are generated by artificial intelligence, an increase of seven percentage points compared to May.
Collective effort intelligent agent
The breakthroughs in a series of key technologies have prepared for the application of large-scale model racing, among which the intelligent agents that are more easily perceived by C-end users have become the collective focus of players in the application field.
Entering Q4, mobile phone manufacturers such as Vivo and Honor have successively sparked a trend of using intelligent agents to order coffee. Microsoft has also launched 10 business intelligence agent plans to "take over" a series of tasks such as human sales, service, finance, and supply chain teams. More big model manufacturers are looking forward to becoming an agent development platform: Tencent announced the launch of the open agent platform Tencent prototype in May this year, and Ant launched the Alipay agent ecological co construction plan in September this year, and launched the agent development platform "treasure box".
Baidu has invested quite early in the field of intelligent agents. According to the reporter's understanding, Baidu released the intelligent agent mechanism in October last year. During this year's Q2 earnings conference, it also revealed that the average daily distribution of intelligent agents in the Baidu ecosystem exceeded 8 million times in July, which is twice as many as in May. At present, the Wenxin intelligent agent platform has attracted 150000 enterprises and 800000 developers.
Robin Lee said in his speech on November 12 that in different historical periods of human information technology change, applications appear differently: in the PC era, they are software and websites; In the mobile era, it was a series of apps and accounts that could be followed; In the era of AI, the main form of application is intelligent agents. "The difference is that intelligent agents are more like people, more intelligent, more like your sales, customer service, and assistants. Intelligent agents may become a new carrier of content, information, and services in the AI native era." He also said that Baidu is not trying to launch a "super application", but to help more people and enterprises create millions of "super useful" applications.
It should be mentioned that while big model manufacturers are gearing up for the other side of intelligent agents, AI big models are still in their early stages, and it will take time to wait for a large number of mature and valuable intelligent agents to emerge.
Gao Wen, an academician of the CAE Member, said in his speech at the 2024 World Conference on Artificial Intelligence in July this year that today's artificial intelligence is mainly low level intelligence, which is based on statistical models. "There is a bit of illusion of medium level, which is equivalent to reading 300 Tang poems and chanting without writing poems. Practice makes perfect."
Zhang Xiaorong, the director of the Institute of Deep Science and Technology, also believes that currently, although everyone can create and use intelligent agents, the fun and value of intelligent agents are not fully demonstrated, and the overall user experience needs to be improved.
He told a reporter from Huaxia Times that in practical applications, how to make users feel valuable and generate stickiness is a key challenge for the development of intelligent agents. In addition, it also includes security and ethical issues, computing resource consumption, complex tool usage, and multi-agent interaction mechanisms. He also believes that intelligent agents face cost challenges in the commercialization process, especially issues such as error loops and high token consumption that occur during agent interactions.