Year-End Review: The Rapid Development of AI Large Models in 2023
-
In January this year, Baomai Fitness at Beijing Changying Hualian Shopping Center updated its lightbox advertisements, with one notable change: the models in the visuals were replaced by AI-generated digital avatars.
Rewind to a year ago, when ChatGPT ignited the AI large model frenzy. People were filled with imagination about the development prospects and application scenarios of general large models, while also harboring concerns about the associated technological risks and ethical issues.
Yet, in just one year, even a small gym has started using AI in its promotional lightboxes. As we marvel at the rapid development of AI large models and their swift integration into daily life, it's worth taking a moment to reflect on what exactly has transpired over this past year. Just two months after its launch, OpenAI's chatbot ChatGPT has become the fastest-growing consumer application in history.
A research report released by UBS shows that ChatGPT had an estimated 100 million monthly active users in January 2023. The report cited data from analytics company Similar Web, stating that approximately 13 million unique visitors used ChatGPT daily worldwide in January 2023. UBS analysts wrote in the report: "In the 20 years of internet development, we can't think of any consumer internet application that has grown faster than this."
According to data compiled by World of Engineering, it took iTunes 6.5 years to reach 100 million users, Twitter 5 years, Meta (Facebook) 4.5 years, and WhatsApp 3.5 years. The explosive popularity of ChatGPT has made the market aware of the opportunities presented by AI large models, and this trend has also swept across China. Various internet and technology giants are actively promoting the construction and release of their own AI large models.
After the wave of large model enthusiasm triggered by ChatGPT, universities in China were the first to release products.
In February 2023, Fudan University released MOSS, a ChatGPT-like conversational large model, and officially open-sourced it two months later, making it the first plugin-enhanced open-source conversational language model in China. Domestic expectations for MOSS were high, and within less than 24 hours of its opening, the MOSS server was overwhelmed by the instantaneous access pressure.
MOSS is mainly used in scientific research. According to Aotou Finance, researchers plan to combine Fudan University's achievements in artificial intelligence and related interdisciplinary fields to equip MOSS with multimodal capabilities such as drawing, voice, and composition, as well as enhance its ability to assist scientists in efficient research.
Despite this, MOSS has set a positive precedent for the development of large-scale dialogue models in China, leading to a wave of large models from major tech companies. On March 16, 2023, among the major internet companies, Baidu took the lead by releasing its large-scale AI model—ERNIE Bot (Wenxin Yiyan).
From an external perspective, Baidu's early release was seen as a natural move. "Baidu Search has been around for many years, and information retrieval is closely tied to big data processing. Baidu has inherent advantages in the investment and application development of large AI models," analyzed an industry observer. "On the other hand, Baidu was one of the first internet giants to establish a clear direction in AI development and has conducted extensive research in this field, giving it a more solid technical foundation."
The debut of ERNIE Bot appeared somewhat rushed. The launch event did not feature the live demo that the public had anticipated, instead showcasing a pre-recorded demonstration video. Despite this, Baidu still managed to take the lead in the fierce competition among large AI models. Since 2019, Alibaba has been conducting research in related fields. In April 2023, Alibaba introduced the Tongyi Qianwen large model, representing a milestone in its large model research.
Tongyi Qianwen is primarily used to empower various internal products of Alibaba, such as DingTalk, Taobao, and Tmall Genie. According to Aotou Finance, after integrating Tongyi Qianwen for testing, DingTalk can automatically generate work plans, summarize meeting minutes and create to-do lists, and even generate mini-programs from functional sketches.
The greater significance of Tongyi Qianwen lies in providing a direction for the commercial application of large AI models. In the AI era, if every enterprise possesses a dedicated large model with industry-specific capabilities, it will bring substantial market growth to large AI models. If Alibaba's Tongyi Qianwen is considered a "weak application" providing technical empowerment for enterprise digital transformation, then Iflytek's release of the Spark Cognitive Model signifies the advent of the era of strong AI model applications.
On May 6, 2023, Iflytek launched the "Spark Cognitive Model" and demonstrated its applications in industries such as education, office work, automotive, and digital employees. The products unveiled include the Iflytek AI Learning Machine T20 series, Iflytek Hearing and Writing, digital employees, smart cockpits, and other AI application products.
According to Aotou Finance, based on the "Spark Model + Education" scenario, Iflytek has introduced products such as Teacher Assistant, Education Digital Base, and Spark Language Companion APP, covering both "teaching" and "learning" aspects. Among them, the Spark Teacher Assistant can intelligently generate scientific and systematic teaching designs, flexible and practical teaching activity designs, and courseware tailored to teaching needs through conversational and generative interactions, focusing on the lesson preparation scenario. This indicates that AI large models have been deployed in consumer-facing applications that are closer to people's daily lives, further broadening their application scenarios.
By the end of May 2023, the "China AI Large Model Map Research Report" was released, showing that China had unveiled 79 large models with over 1 billion parameters by the time of the report's publication.
As the competition among hundreds of models intensifies, Light Year Beyond encountered unexpected changes with its leadership team member Wang Huiwen taking medical leave, while Meituan provided a safety net for the team. On June 29, 2023, Meituan announced the completion of its equity acquisition of Light Year Beyond for approximately 2.065 billion RMB. The announcement revealed that the total consideration included approximately $233 million in cash, debt assumption of about 367 million RMB, and a symbolic cash payment of 1 RMB.
Meituan stated that after the acquisition, it would continue to support the Light Year team in their exploration and research in the field of large models. However, as of now, neither Meituan nor Light Year Beyond has produced significant research outcomes.
The developments at Light Year Beyond serve as a reminder to practitioners that the AI large model field is a 'long and arduous track,' and entering it with mere enthusiasm might lead to 'total loss.' In July 2023, an AIGC product went viral - Miao Ya Camera, an AI face-swapping software launched by Alibaba's digital entertainment division. Users can upload photos to generate AI images with different styles and backgrounds.
WeChat data shows that public interest in Miao Ya Camera peaked on July 25 and 27. According to Qimai Data, after the app's launch on July 28, its download rankings surged rapidly, reaching the top spot on China's charts by August 12.
While Miao Ya Camera gained explosive popularity, it also raised concerns about data security and personal privacy. Many users questioned whether the digital avatars created through the app pose information leakage risks. The potential exposure of high-precision, high-quality photo data could bring immeasurable harm to users. The emergence of new technologies and applications inevitably brings controversy. However, we should not restrict technological development due to disputes or risks. Instead, we should use institutional measures to mitigate risks and provide broad space for technological advancement.
On August 15, 2023, the "Interim Measures for the Management of Generative Artificial Intelligence Services" (referred to as the "Interim Measures") officially came into effect. These measures aim to promote the healthy development and standardized application of generative artificial intelligence, safeguard national security and public interests, and protect the legitimate rights and interests of citizens, legal persons, and other organizations.
At the same time, 11 companies and institutions had their large model products pass the first batch of filings under the "Interim Measures." These include Baidu's ERNIE Bot, ByteDance's Skylark model, Baichuan AI's Baichuan model, Zhipu AI's ChatGLM, the Purple Mountain Taichu model developed by the Institute of Automation of the Chinese Academy of Sciences, SenseTime's SenseChat, MiniMax's ABAB model, and Shanghai AI Lab's InternLM, as well as large model products from Huawei, Tencent, and iFlytek. After obtaining the first batch of filings under the Interim Measures, Tencent released its Hunyuan large model on September 7, 2023. It is reported that the Hunyuan large model has a parameter scale exceeding 100 billion, with pre-training corpus surpassing 2 trillion Tokens. It has already been integrated into over 50 Tencent services, including Tencent Cloud, Tencent Advertising, Tencent Games, Tencent Financial Technology, Tencent Meeting, and Tencent Docs.
Alongside the release of the large model, Tencent also announced that the Hunyuan large model is officially open to the public through Tencent Cloud. Various industries can access Hunyuan via API calls or use it as a base model to build large model applications for different industrial scenarios.
Among domestic internet giants, Tencent's release of its own large model was relatively late. However, with the launch of the Hunyuan large model, major internet companies have taken the lead in this field, gradually 'closing the door' on the domestic general-purpose large language model race. On October 31, 2023, the 2023 Apsara Conference kicked off in Hangzhou. During the event, Alibaba Cloud released its 100-billion-parameter large model Tongyi Qianwen 2.0. In 10 authoritative evaluations, Tongyi Qianwen 2.0's comprehensive performance surpassed GPT-3.5 and is rapidly closing the gap with GPT-4.
According to Aotou Finance, Tongyi Qianwen 2.0 has shown significant improvements in complex instruction understanding, literary creation, general mathematics, knowledge retention, and hallucination resistance.
Since the release of the "Interim Measures," the "Hundred Model War" has entered its second phase. Leading tech giants continue to enhance their large model versions and capabilities, while later entrants focus more on industry-specific and vertical large models. In the competition of the second half, some will inevitably rise to prominence while others will falter, though no one wishes to be the latter.
In November 2023, shortly after the release of ChatGPT-4 turbo, OpenAI staged a dramatic power struggle.
OpenAI CEO Sam Altman received an email from the board notifying him of his dismissal, followed by an official statement from OpenAI announcing, 'Altman will step down as CEO and leave the board.' Altman negotiated with the board, hoping to return as CEO and restructure the board, but the two sides failed to reach an agreement.
Then the situation took a turn. Microsoft subsequently announced that it would accept Sam Altman and have him lead a new advanced AI research team. Meanwhile, over 700 OpenAI employees signed a joint letter demanding the board's resignation and Altman's reinstatement. With the backing of investors and employees, Altman regained control of the company.
From an external perspective, there appear to be two main reasons for the internal changes at OpenAI: first, disagreements among executives over technical direction, and second, declining product performance data and ongoing financial losses. Data shows that from January to May 2023, ChatGPT's global traffic growth rates were 131.6%, 62.5%, 55.8%, 12.6%, and 2.8% month-on-month, respectively, showing a declining trend each month. From June to August 2023, the situation further deteriorated, with growth rates dropping to -9.7%, -11.2%, and -3.2%, respectively, marking three consecutive months of decline.
In fact, commercialization issues continue to trouble pioneers in the AI large model industry, indicating that the commercial prospects of large models still require ongoing exploration.
AI large model learning requires vast amounts of data, which is highly likely to constitute infringement. On December 27, 2023, The New York Times filed a lawsuit against Microsoft and OpenAI, marking the first instance of a major U.S. media company taking legal action against AI technology firms for copyright infringement.
The New York Times argues that their works involve substantial intellectual and financial investments, and the unauthorized use of these works by AI technology companies has caused significant harm to others.
The development of new technologies is never without challenges, and AI large language models are no exception. Over the past year, AI models have advanced rapidly, but these achievements may also include an overextension of future potential. In the new year, the emerging industry should be met with less criticism and more trust.