List of high-quality authors
Reporter丨Zhang Feitao Trainee丨Han Qing
Produced by丨Aotou Finance (thesankei).
In January this year, Bolmiles Fitness, located in Beijing's Changying Hualian Shopping Center, updated its lightbox advertisement, which is different from the past, and the model on the screen has been replaced by an AI-generated digital human.
Back to a year ago, ChatGPT set off a boom in AI large models, and people are full of imagination about the development prospects and application scenarios of general large models, but also full of worries about the technical risks and ethical issues they face.
However, in just one year, a small gym has used AI on the promotional light box, and when we are amazed at the rapid development of AI large models and the rapid penetration of life, we should look back and see what happened this year.
January: ChatGPT exceeded 100 million monthly active users, becoming the fastest-growing consumer app in history
Just two months after its launch, ChatGPT, OpenAI's chatbot, has become the fastest-growing consumer app in history.
According to a research report released by UBS, ChatGPT has an estimated 100 million monthly active users in January 2023. In January 2023, about 13 million unique visitors worldwide used ChatGPT every day, the report said, citing data from analytics firm Similar Web. "In the 20 years of development in the Internet space, we can't think of a consumer Internet application that has been able to grow faster than it," UBS analysts wrote in the report. ”
According to data compiled by World of Engineering, it took six and a half years for iTunes to reach 100 million users, five years for Twitter, four and a half years for Meta (Facebook), and three and a half years for WhatsApp.
The explosion of ChatGPT has also made the market see the opportunity of AI large models, and this wind has also blown to the Chinese market, and various Internet and technology companies have promoted the construction and release of their own AI large models.
February: Fudan University released MOSS as a prelude to the 100-model war
After the craze for large models caused by ChatGPT swept in, universities were the first to release products in China.
In February 2023, Fudan University released MOSS, a chatgpt-like conversational large model, which was officially open-sourced two months later, becoming the first plug-in-enhanced open-source conversational language model in China.
China is full of expectations for MOSS, and it has not been more than 24 hours since it opened, and due to the excessive pressure of instantaneous access, the MOSS server has been crowded.
MOSS mainly plays a role in the scope of scientific research, and Aotou Finance understands that researchers plan to combine Fudan's research results in artificial intelligence and related interdisciplinary disciplines to endow MOSS with multimodal capabilities such as drawing, speech, and music composition, and strengthen its ability to assist scientists in efficient scientific research.
Despite this, MOSS also made a good start for the development of the domestic dialogue model, and then the model between the big manufacturers "you sing and I will appear".
March: The release of Wenxin's words was a little hasty
On March 16, 2023, among many major Internet companies, it took the lead in releasing a large model - Wenxin Yiyan.
In the eyes of the outside world, it is reasonable to be the first to release. "After so many years of searching, information retrieval and big data processing are closely related, and there are inherent advantages in the investment and application research and development of AI large models. Industry observers said, "On the other hand, it is one of the Internet companies that established the development direction of AI earlier, and has also done a lot of research in this area, and it is more solid in terms of technical reserves." ”
Wenxin Yiyan's first launch was a little hasty, and there was no actual demonstration of the actual machine expected by the public at the press conference, but only a pre-recorded demonstration. Despite this, it still came out on top in the 100-model war.
April: Alibaba released Tongyi Qianwen to take the lead in transforming internal products
Since 2019, Alibaba has started research in related fields. In April 2023, Alibaba launched the Tongtongyi Qianwen model, which is the phased achievement of Alibaba's large-scale model research.
Tongyi Qianwen is mainly used to empower Alibaba's internal products, such as DingTalk, Tmall Genie, etc. Aotou Finance learned that after DingTalk is connected to the Tongyi Qianwen test, it can automatically generate a work plan, a summary and to-do list after the meeting minutes, and a functional sketch and automatically generate a small program.
The greater significance of Tongyi Qianwen is that it provides a direction for the commercial application of AI large models, and in the AI era, if each enterprise has an exclusive large model with industry capabilities, it will bring huge market increments to the AI large model.
May: The era of strong application of iFLYTEK Xinghuo large model is coming
If Alibaba's Tongyi Qianwen is just a "weak application" of Alibaba Cloud to provide more technology empowerment in the digital transformation of enterprises, then the Xinghuo cognitive model released by iFLYTEK marks the advent of the era of strong application of AI large models.
On May 6, 2023, iFLYTEK released the "Spark Cognitive Model" and released the application results of the large model in education, office, automobile, digital employee and other industries, including iFLYTEK AI learning machine T20 series, iFLYTEK hearing and writing, digital employee, intelligent cockpit and other artificial intelligence application products.
Aotou Finance has learned that based on the "Spark Model + Education" scenario, iFLYTEK has launched products including teacher assistants, educational digital bases, and Spark language companion apps, covering both ends of "teaching" and "learning". Among them, Xinghuo Teacher Assistant can intelligently generate scientific system teaching design, flexible and practical teaching activity design, and courseware that meets teaching needs for teachers through conversational and generative interaction, focusing on lesson preparation scenarios.
This means that the AI model has been applied to the C-end that is closer to people's daily life, and the application scenarios have been further broadened.
June: The battle of 100 models is white-hot, and Meituan is personally down
At the end of May 2023, the "Chinese Intelligent Large Model Map Research Report" was released, and the report showed that as of the release of the report, 79 large models with parameters above the scale of 1 billion parameters in China have been released.
When the war of 100 models entered the white heat, there was a sudden change light years away, Wang Huiwen, who led the team, was sick and retired, and Meituan took care of the team.
On June 29, 2023, Meituan issued an announcement to about 206.5 billion yuan **, completed the acquisition of equity beyond Light Year. According to the announcement, the total consideration of the acquisition includes about 2$3.3 billion, debt assumption of about 3RMB6.7 billion, and RMB1 in cash.
Meituan said that after the completion of the acquisition, it will continue to support the Lightyear team's exploration and research in the field of large models. However, so far, there are few research results from Meituan or light years away.
What happened light years away also reminded practitioners that the AI model is a "track with long slopes and thick snow", and it may be "lost in blood" if it only enters with enthusiasm.
July: The Wonderful Duck Camera exploded and AIGC caused data security problems
In July 2023, an AICG product will become popular, this is the wonderful duck camera launched by Ali Da Entertainment, which is an AI face swap software, and users can get AI-generated pictures of different styles and backgrounds after uploading**.
WeChat data shows that the attention of Miao Ya Camera peaked on July 25 and July 27; Qimai data shows that after the Miao Ya camera app was launched on July 28, the ranking climbed rapidly in a short period of time, and won the first place on the list on the 12th.
At the same time as the explosion of the Miao Ya camera, data security, personal privacy and other issues have attracted attention, many users question, through the Miao Ya camera generated by the digital clone, whether there is a risk of information leakage, high-precision, high-quality information once leaked, may bring immeasurable harm to users.
The emergence of new technologies and new applications will inevitably be accompanied by controversy, but technological development should not be restricted by controversy or risk, but should be avoided by the system and given a broad space for technological development.
August: The Interim Measures come into force to bring AI into regulation
On August 15, 2023, the Interim Measures for the Administration of Generative AI Services (the "Interim Measures") came into effect. The measures aim to promote the healthy development and standardized application of generative AI, safeguard the public interest, and protect the legitimate rights and interests of citizens, legal persons and other organizations.
At the same time, there are 11 enterprises and institutions of large model products for the first batch of the "Interim Measures" for the record, namely Wenxin Yiyan, ByteDance research and development of the lark model, Beijing Baichuan Intelligent Technology launched the Baichuan model, Beijing Zhipu Huazhang Technology launched Zhipu Qingyan, the Institute of Automation of the Chinese Academy of Sciences research and development of Zidong Taichu, Beijing SenseTime Technology research and development of SenseChat, The ABAB large model launched by MiniMAX and the general large model of scholars developed by the Shanghai Artificial Intelligence Laboratory. and large-scale model products launched by Huawei, Tencent and iFLYTEK.
September: Tencent's mixed yuan model was completed
After obtaining the first batch of filings for the Interim Measures, on September 7, 2023, Tencent released the hybrid model. It is understood that the hybrid model has a scale of more than 100 billion parameters, a pre-trained corpus of more than 2 trillion tokens, and has been connected to more than 50 Tencent businesses such as Tencent Cloud, Tencent Advertising, Tencent Games, Tencent Financial Technology, Tencent Meeting, and Tencent Documents.
At the same time as the release of the large model, Tencent also announced that the hybrid model is officially open to the public through Tencent Cloud, and thousands of industries can call the hybrid element through APIs, or use the hybrid element as the base model to build large model applications for different industry scenarios.
Among the domestic Internet giants, Tencent's own large model release is relatively backward, but after the release of the mixed yuan model, various Internet manufacturers took the lead in this field, and the domestic general large language model track slowly "closed".
October: Tongyi Qianwen performance exceeds GPT-35 The battle of 100 models enters the second half
On October 31, 2023, the 2023 Apsara Conference opened in Hangzhou. During the conference, Alibaba Cloud released the 100-billion-level parameter model Tongyi Qianwen 20, among the 10 authoritative evaluations, Tongyi Qianwen 20. The comprehensive performance exceeds GPT-35. Catching up with GPT-4 is accelerating.
Aotou Finance learned that Tongyi Qianwen 2The 0 parameter has been significantly improved in the ability of complex instruction comprehension, literary creation, general mathematics, knowledge memory, and hallucination resistance.
Since the release of the "Interim Measures", the 100-model war has entered the second half, and the major manufacturers that have been in the first position have continuously improved their versions and capabilities of large models, and the follow-up filings have focused more on industry large models and vertical large models.
In the second half of the competition, there will definitely be people who stand out, and there will eventually be people who will fall into the sand, but no one wants to be the latter.
November: OpenAI's sudden "palace fight" The prospect of commercialization of large models needs to be explored urgently
In November 2023, Open AI, which had just released ChatGPT-4 Turbo, staged a "palace fight scene".
Open AI CEO Sam Altman received an email from the board of directors to expel him, and then Openal issued a statement saying that "Altman will resign as CEO and leave the board."
Altman and the board of directors negotiated a return to the position of CEO and a reorganization of the board of directors, but the two sides did not reach an agreement.
Then things took a turn for the worse. Microsoft has since announced that it has accepted Sam Altman and put him at the head of a new team of advanced AI researchers, and more than 700 Open AI employees have signed a joint letter asking the board to resign and bring Altman back. With the support of the owner and employees, Ultraman regains control of the company.
In the eyes of the outside world, there are two reasons for OpenAI's internal changes, one is the technical route dispute between executives, and the other is the deteriorating data and continuous losses of products.
According to the data, from January to May 2023, the month-on-month growth rate of ChatGPT's global visits is as follows: 8%, showing a downward trend month by month; From June to August 2023, the situation deteriorated further, with the growth rate falling to -9., respectively7%、-11.2%、-3.2%, down for three consecutive months.
In fact, the issue of commercialization is still plaguing the pioneers of the AI large model industry, which indicates that the commercialization prospect of large models still needs to be continuously explored.
December: The New York Times sues Microsoft for OpenAI's long road to large-scale development
AI large model learning requires a large amount of data, which is likely to constitute infringement.
On December 27, 2023, the New York Times filed a lawsuit against Microsoft and Open AI, becoming the first major U.S. company to sue an AI technology company for infringement.
The New York Times argues that the work contains a large investment of intellectual and financial resources, and the unauthorized use of the work by AI technology companies has caused huge losses to others.
The development of new technologies has never been smooth sailing, and the same is true for AI models, which have developed rapidly in the past year, and these achievements may also include overdrafts for the future.