Today, some media reported that 11 large models in China will successively pass the "Interim Measures for the Management of Generative Artificial Intelligence Services", and the first batch will be opened to the public from August 31. Among them, 5 in Beijing and 3 in Shanghai will be the first to go online, 2 in Guangdong Province and 1 in other provinces and cities will also be opened one after another. It is reported that the approved companies in Guangdong are Huawei and Tencent, and iFLYTEK is a product approved in other regions.
Among them, there are 5 enterprises and institutions in Beijing, including:
- Baidu (Wen Xin Yi Yan)
- TikTok (lark big model)
- Smart Spectrum AI (GLM Big Model)
- Chinese Academy of Sciences (Zidong Taichu Grand Model)
- Baichuan Intelligent (Baichuan Big Model)
3 corporate organizations in Shanghai, including:
- Nikchi New Model
- MiniMax(ABAB model)
- Shanghai Artificial Intelligence Laboratory (Scholar General Large Model)
Once these products are approved, the large-model products of the above-mentioned organizations can be downloaded from the website/app store and used directly.
The approved large-model products of the enterprises were launched today
As early as 00:02 today, Wen Xin's words took the lead in opening up to the whole society. Users can download the "Wenxin Yiyan APP" in the app store or log on to the "Wenxin Yiyan official website" to experience. At the same time, enterprise users can directly log in to the official website of Baidu Intelligent Cloud Qianfan Model Platform and call Wenxin's one-word ability.
It is reported that Baidu will also hold a press conference in October to open a batch of newly reconstructed AI native applications, allowing users to fully experience the four core capabilities of generative AI's understanding, generation, logic and memory.
On March 16 this year, Wen Xin opened the invitation to test, but the "recording" behavior at the press conference once disappointed users. However, after several months of iterative upgrading, the recently upgraded Wenxin Model 3.5 continues to rank first in more than ten authoritative evaluations at home and abroad.
Li Yanhong said that when Wen Xin Yiyan opens up its services to hundreds of millions of Internet users on a large scale, it can get a lot of human feedback in the real world, which will further improve the basic model and iterate Wen Xin Yiyan at a faster speed to create a better user experience.
At 1:44 a.m., Beijing Zhipu Huazhang Technology Co., Ltd. ("Zhipu AI") also announced that Zhipu AI's intelligent assistant "Zhipu Qingyan" was fully opened. At present, Zhipu Qingyan already has the capabilities of general question and answer, multi-round dialogue, creative writing, code generation and virtual dialogue, and will open up multimodal generation capabilities in the future. Users can download it through the Apple Store App Store, Android mainstream stores, or search for "Zhipu Qingyan" in the WeChat mini program to experience its functions.
Similarly, SenseTime officially released news to the outside world today that the company's large-model application "Discussion SenseChat" will be fully opened to the majority of users from now on. Its basic model is the "Scholar · InternLM-123B", with 123 billion parameters, can be registered through the official website of SenseChat.
Among the approved lists, Baichuan Intelligent was established the shortest, founded by Wang Xiaochuan in April this year. Just four months later, Baichuan Intelligent has successively released three general large language models, and its Baichuan-53B parameter magnitude released on August 8 reached 53 billion. Today, Baichuan Intelligent also said that it took the lead in filing through the "Interim Measures for the Management of Generative Artificial Intelligence Services" and opened services to the whole society. Users can feel the experience brought by the "Baichuan Big Model" in the fields of knowledge Q&A, text creation and so on.
In addition, iFLYTEK Spark tentatively plans to open its services to the public on September 1, and 360 Wisdom Brain is also preparing for its official launch. The list of other enterprises that have passed the filing in the first batch is also expected to be approved by the relevant local management departments within one week.
Humans will babble, and so will artificial intelligence
In the past six months, the success of AIGC and large models has brought unprecedented "intelligence emerged", and people are full of expectations for the coming era.
However, as tech giants flock to the AI track, people are excited to experiment and discuss the power of AI, and thus lament whether it may replace human labor, the problem of AI illusion cannot be ignored, becoming an obstacle to the further development of AI. Yann LeCun, one of the world's three giants of deep learning, even asserted in a previous speech that "GPT models will not live for 5 years."
With the controversy surrounding the AI illusion, how much role the big model can play in the industry and whether it will produce side effects has also become a focus issue.
In a word, the nonsense of artificial intelligence is the so-called "machine illusion".
Specifically, AI illusion is content generated by large models that appears reasonable and logical on the surface, and may even be intertwined with real information, but in fact there is wrong content, citation sources, or statements. These erroneous contents are presented in a persuasive and credible way, making it difficult to discern the false information in them without careful verification and fact-checking.
Currently, AI hallucinations can be divided into two categories: Intrinsic Hallucination and Extrinsic Hallucination.
The so-called internal hallucination refers to the contradiction between the content generated by the large model and its input content, that is, the generated answer is inconsistent with the information provided. Such errors can often be found and corrected relatively easily by checking the inputs and generating the content.
External illusions, on the other hand, are more complex, referring to errors in generated content that cannot be directly verified from the input. This error typically involves the model calling data, text, or information outside of the input, causing the generated content to be misstated. External hallucinations are difficult to identify easily, because while the generated content may be false, the model can be presented in a logical, coherent and methodical way, making it difficult to doubt its authenticity. In layman's terms, AI is "making up information".
At present, major big language model technology companies, including OpenAI, have unanimously said that they are working to improve the "illusion" problem and make large models more accurate.
The "illusion" has not yet been resolved, but the policy comes first
With the continuous development of artificial intelligence technology, generative artificial intelligence services have gradually become the focus of attention. This technology can have a huge impact in many areas, but it also brings some management challenges. Especially since the explosion of large models this year, there have been relevant policies at the regulatory level. Generative AI products involve user data security, and in addition to algorithm filing, security assessment needs to be reported to relevant departments before they are officially launched.
As early as April 11, the Cyberspace Administration of China issued the Measures for the Administration of Generative Artificial Intelligence Services (Draft for Comments), which stated that before using generative AI products to provide services to the public, they should report the security assessment to the state internet information department in accordance with the "Provisions on Security Assessment of Internet Information Services with Public Opinion Attributes or Social Mobilization Capabilities", and perform the procedures for algorithm filing, modification, and cancellation in accordance with the "Provisions on the Administration of Internet Information Service Algorithm Recommendations". At the same time, content generated by generative artificial intelligence should be truthful and accurate, and measures should be taken to prevent the generation of false information.
On June 20, the Cyberspace Administration of China issued an announcement on the filing information of deep synthesis service algorithms, publicly releasing domestic deep synthesis service algorithm filing information. According to the list attached to the announcement, a total of 41 deep synthesis service algorithms were included in the list, covering many popular applications such as Meituan, Kuaishou, Taobao, Damai, Kugou Music, and Baidu Wenxin Yige.
On July 13 this year, the "Interim Measures for the Management of Generative Artificial Intelligence Services" was released, clarifying the management principles and specifications of generative artificial intelligence services and promoting the healthy development of the technology. The Measures define generative AI services as the process of providing users with intelligent and personalized services based on deep learning, natural language processing, computer vision and other technologies through model generation and data training, which is applicable to all institutions and individuals providing generative AI services.
Now, from filing to approval, domestic large-model enterprises finally have clear policy access. The first approved enterprises can prioritize the launch of AI products for toC users, which means that they can dominate user accumulation, data acquisition and product polishing.