Article | Liu Yiqin Zheng Keshu Editor | Xie Lirong

Source: Caijing Eleven

On August 31, including Baidu's Wenxin Yiyuan, Zhipu Qingyan, Sensetime SenseChat, and other large-scale model applications finally obtained the first batch of entry tickets issued by the regulatory authorities, and can provide services to ordinary individual users. This is a key milestone. Large model companies that started at the beginning of the year can finally open to individual users in China.

The regulatory authorities have issued licenses and filings for large model service providers based on geographical divisions.

In Beijing, the first batch of 5 large models have been filed under the "Interim Measures for the Management of Generative Artificial Intelligence Services," including Baidu's Wenxin Yiyuan, Douyin's Yunque, Baichuan Intelligence's Baichuan Large Model, Zhipu Qingyan under AI company Zhipu, and Zidong Taichu under the Chinese Academy of Sciences.

Shanghai, Guangdong, and other places also have large model products approved for launch. Sensetime, Shanghai Artificial Intelligence Laboratory's Shusheng Large Model, and the previously opened Minimax belong to the Shanghai filing. Tencent and Huawei belong to the Guangdong filing. iFlytek's Xinghuo Large Model and Alibaba Cloud's Tongyi Qianwen Large Model have also claimed to have completed the filing. A total of 12 approved large model enterprises have been mentioned. Most companies have announced this news, but the relevant persons in charge of Douyin and Tencent have told "Caijing Eleven" that there is no response regarding the filing and open application of large models.

As of now, the mobile app stores can download Wenxin Yiyuan and Zhipu Qingyan, while Sensetime SenseChat and Baichuan Intelligence are only available for web use. "Caijing Eleven" compared the three applications (Wenxin Yiyuan, Qingyan, and SenseChat) and found that Wenxin Yiyuan's database is relatively up-to-date, with data up to 2022, and can also answer some information from 2023. Zhipu Qingyan's database is updated to 2021.

If asked "What major events have happened in the past week," Wenxin Yiyuan answered with several major events from a month ago (July 2023); Zhipu Qingyan said it could not answer because there is no relevant information in the training data; and Sensetime SenseChat answered with three events from different years (2013, 2016, 2023).

The "Interim Measures for the Management of Generative Artificial Intelligence Services" was implemented on August 15, 2023, requiring that for generative artificial intelligence services with public opinion attributes or social mobilization capabilities, safety assessments should be conducted in accordance with relevant national regulations, and algorithm filings, changes, and cancellations should be carried out in accordance with the "Regulations on the Management of Algorithm Recommendations for Internet Information Services."

"Caijing Eleven" learned that the relevant companies have filed with the relevant departments in their respective regions, and each local department chooses the release time on its own. The threshold for filing is not high, and more large model companies will complete the filing in the future.

The shift from the initial licensing system to the current filing system means that the threshold for large models to provide services to individual users has been lowered. However, the entry threshold for this industry has actually increased—each interaction by users is feeding new data to the large models, and the competition threshold for later entrants will be significantly raised.

Is it too late to open to individuals now?

Currently, several large models that have been launched are free trial clients for ordinary individual users, including Wenxin Yiyuan, Zhipu, Sensetime, and others. It is currently uncertain whether other companies that have completed the filing will also release free applications for the consumer market.

Most of China's large model manufacturers have previously focused on the B-end market for enterprises, providing basic capabilities of large models to enterprise users, providing computing resources, or developing vertical large models. The logic is simple: first, without regulatory filing permits, they cannot follow ChatGPT in opening to individual users in advance; second, the business model for the B-end market is clearer.

OpenAI, now valued at $29 billion, first opened to individual users when releasing ChatGPT. A person close to OpenAI told "Caijing Eleven" that OpenAI's strategy is to cover enough users in the short term through C-end applications to achieve market share in the education market. In addition, the more users there are, the more feedback and new data the large models will receive, which can help test their effectiveness and capabilities, and facilitate further optimization to achieve commercialization in the B-end.

However, it is a well-known fact that the cost of large model language applications is high, and the influx of a large number of users will directly lead to a sharp increase in costs, mainly in computing power consumption. At the beginning of this year, the daily operating cost of ChatGPT was about $100,000. By April of this year, Dylan Patel, Chief Analyst of semiconductor research company SemiAnalysis, said that the daily operating cost of OpenAI's ChatGPT could be as high as $700,000.

The high cost also brings high returns. OpenAI's C-end layout has been effective. In April of this year, OpenAI completed a $10.3 billion financing. Recently, OpenAI has launched an enterprise version of ChatGPT and will also launch customized database analysis tools in the future.

The high cost of computing power consumption is probably an important factor that all entrants have to consider. As of the time of writing, many companies including Baidu, Alibaba, iFlytek, Sensetime, Zhipu, and Baichuan have not responded to "Caijing Eleven" on whether they will continue the long-term strategy of free applications for ordinary individual users.

A person in the large model industry told "Caijing Eleven" that the relevant companies are constantly adjusting and optimizing their large models, and they have not yet reached a "mature" stage. Opening C-end applications is also an opportunity for testing. Especially for start-up companies, attracting a large number of users can also help with subsequent financing.

Wenxin Yiyuan was released on March 16, 2023, and then entered a small-scale internal testing and application phase for enterprise users. This time, in addition to fully opening Wenxin Yiyuan to individuals, a person in charge of Baidu told "Caijing Eleven" that Baidu's strategy is to take multiple approaches. Various user-side products under Baidu are being reconstructed with large models, including Baidu Search, Baijiahao, Baidu Wenku, and input methods, and will gradually launch related AI functions. In addition, the B-end strategy is also important. Enterprise users can call the capabilities of Wenxin Yiyuan through the Qianfan Large Model Platform website.

Baidu Wenku has already launched the "AI Makes PPT" creation function based on large models. Baidu stated that after being open for 12 hours, it has been experienced by over 1 million people. In the B-end market, Baidu has completed the first phase of application scenario co-creation with some enterprises.

Zhipu AI, Baichuan Intelligence, and Minimax are representative start-up companies in the new wave of AI large models. Zhipu AI was established in 2019, with a core team from Tsinghua University, and has completed four rounds of financing. The latest round was an investment from Meituan in July 2023. Baichuan Intelligence was established only four months ago by former Sogou CEO Wang Xiaochuan.

Minimax was established in 2021 and has disclosed three rounds of financing, with the latest round being an investment from Tencent in June 2023. Last October, Minimax launched a chat application called Glow, when there were no related filing requirements. Another AI company, Lingxin Intelligence, launched the chat application "AI Utopia" in December 2022, which has now been upgraded to "AiU" and is a product developed in deep cooperation with Zhipu AI and is currently in normal use.

Competition Enters the Second Half

At the beginning of this year, OpenAI and a series of American technology companies made intensive moves in large models, giving a great stimulus to the Chinese technology and venture capital circles.

Several industry insiders told "Caijing Eleven" a common view: in the field of general large models, the competition among Chinese companies is relatively homogeneous, with algorithms based on Transformer and data coming from open-source datasets and web crawling data. There may be some differences in computing power, as large companies have accumulated computing power and have the strength to purchase or lease more computing power. However, leading start-up companies have also received good financing this year, and there won't be much difference in computing power.

An enterprise user who has extensively tested GPT-4, Wenxin Yiyuan, Minimax, and Zhipu told "Caijing Eleven" that he found that the overall performance of GPT-4 is better than the other three. If GPT-4 is rated 100, the others are around 70-80. However, GPT-4 needs to call overseas servers, which is relatively more troublesome.

In addition, in terms of growth rate, the evolution speed of domestic large model applications has been very fast, and the gap with GPT-4 is rapidly narrowing.

The continuous open sourcing and open use of general large models means that the industrial competition of large models has entered the second half, and the gap between early entrants and later entrants has further widened—general large models have a certain first-mover advantage, as each user interaction is feeding new data to the large models, and the competition threshold for later entrants will be significantly raised.

An investor who has long been focused on AI commented that the recent attitude of capital towards large models has clearly cooled, and this is due to several reasons: First, large models are very expensive, and there are not many institutions that can invest, and those that can have already entered the market; second, everyone has found that the commercialization path of OpenAI is not clear enough, and the commercialization of domestic large models is still in the PowerPoint stage.

Currently, it can be relatively certain that the new direction is vertical large models and multimodal large models.

Compared to general large models, the parameter size and model volume of vertical large models will be smaller, and enterprise users can choose as needed, with lower costs. The key to competition lies in having sufficient vertical domain data, which is the biggest threshold. In addition, it is also necessary to have a sufficient understanding of the vertical industry, to discover the pain points of industry users or have the ability to create new demands.

Multimodal is another direction. The reproduction of single-modal large models (only generating text or images) has already been proven, and multimodal is a new path to prove the technological barriers of tech companies. On August 17, 2023, OpenAI acquired the start-up company Global Illumination to prepare for multimodal by supplementing video technology capabilities.

Several investors and practitioners who focus on AI mentioned that the biggest problem in the current large model industry is that the commercialization path is not clear enough. Whether facing investors or enterprise users, they are usually asked the same question, "What can you do that other large models cannot?" This question is difficult for many people to answer.

The commercialization of large models is based on solving real problems. Currently, the industry is in a "vaguely correct" state regarding large models. "There are too many question marks, and it can only be confirmed that large models are an important direction, but it is still uncertain how to do it and who can do it." A investor mentioned, so the strategy of large companies is not only to do it themselves, but also to invest externally, such as Meituan, ByteDance, Tencent, etc., "Since the direction is correct, try multiple paths."

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。