Generative Artificial Intelligence: On the Path to Changing the World Parameters | Models | Artificial Intelligence

Few technologies have received such differentiated evaluations as generative artificial intelligence. The so-called generative artificial intelligence refers to an artificial intelligence system that can generate text, images, or other information based on prompts. ChatGPT, which was born in November last year, is the most well-known representative among them.

Adopters believe that it is as important as the invention of the Internet, type printing and electricity. Gates, the co-founder of Microsoft, said: "Artificial intelligence has the ability to change the world." The scaremongers believe that it is "one of the greatest risks of human civilization in the future." Truss CEO Mask speculates that it may control or even destroy human civilization. Disdainful people believe that "its intelligence is not even as good as a dog," Turing Award winner Yang Likun asserts, "Within 5 years, no one will use ChatGPT."

During the writing process of this article, ChatGPT was asked to write an article introducing generative artificial intelligence. The article title it generates reflects the intertwined emotions of human love and fear - "Unleashing the creativity of generative artificial intelligence: changing technology and triggering threats to humanity.".

Excitement and confusion

The pursuit of generative artificial intelligence comes from its powerful capabilities that exceed human expectations. The performance of the large model GPT-4 based on ChatGPT in various tests is surprising: successfully passing the unified lawyer exam, with scores higher than 90% for human candidates; Participated in the Biology Olympiad and defeated 99% of human candidates; Still ranked at the top of college exams in courses such as art history and biology.

The fear of generative artificial intelligence comes from people's confusion about it - where its capabilities come from, even developers cannot fully understand. This is totally different from the past programs. Traditionally, programmers have designed the program in their minds before they began to write code, which can explain what the machine is going to do. However, the machine learning models based on artificial intelligence are essentially "black boxes", and researchers do not fully understand how it generates content. Professor Zhao Xing from the Big Data Research Institute of Fudan University said, "People throw data into models without clear expectations of what comes out." It is evident that at the technical level, generative artificial intelligence has endogenous uncertainty.

Regardless of people's attitudes towards generative artificial intelligence, its fiery is an indisputable fact. Just two months after ChatGPT launch, it reached 0.1 billion monthly active users, making it the fastest growing app ever. Following ChatGPT, major Chinese and American technology companies have launched their own generative artificial intelligence products. PitchBook data show that in 2022 alone, 78 investments were made in the generative AI circuit, with a total financing amount of more than $1.37 billion, almost equal to the total financing in the past five years. According to ResearchandMarkets forecasts, the global digital content market will grow by $519.83 billion from 2020 to 2024, with a compound annual growth rate of 15%.

Since the emergence of the iPhone, people's belief in a new technology that can change the industry has never been so profound. Cloud computing companies are following suit, developing and selling products based on big models; E-commerce companies hope to change their business models through generative artificial intelligence; Social media platforms are filled with posts written by robots

In the face of the risks that generative artificial intelligence may bring, the call for regulation and standardization is gradually increasing. The Italian Data Protection Agency announced on March 31 that it has temporarily suspended the use of ChatGPT and has launched an investigation into ChatGPT's alleged violation of data collection rules. The European Data Protection Commission announced on April 13th that a special working group is being established to assist EU countries in responding to ChatGPT and promoting cooperation among EU countries. On April 11, the National Internet Information Office of China issued the Measures for the Management of Generative AI Services to standardize the development and application of generative AI in China.

The changed production logic

People's understanding of artificial intelligence has been reshaped. A practitioner in the domestic consulting industry told reporters that the artificial intelligence discussed in the industry in 2023 and 2022 are no longer the same. Previously, the artificial intelligence industry was more based on algorithms to create knowledge graphs, while after 2023, artificial intelligence means big models, generative artificial intelligence, and natural language interaction.

In fact, this change in definition implies a change in the logic of human future production patterns.

On the one hand, generative artificial intelligence can enhance the integration and penetration of artificial intelligence technology into traditional industries, as well as the quality and efficiency of transformation. Liu Chen, a researcher at Bank of China Research Institute, told reporters that just like ChatGPT can provide basic work such as intelligent consulting, document writing, code writing, and translation, generative artificial intelligence can also cover basic content modalities such as text, audio, images, and videos, and can also perform cross modal generation, thereby empowering various segmented industries to promote productivity improvement. In the future, from perception to cognition, the creative ability of artificial intelligence will continue to enhance, entering into human daily life and endowing all things with intelligence.

On the other hand, compared to professionally generated and user generated content, generative artificial intelligence has lower costs and richer content. Artificial intelligence has achieved a transformation from screening and recommendation to the role of content creators, and will become a new generation of industry wide productivity tools.

The new technology route represented by OpenAI adopts large model training with billions of parameters, which will also be closer to human thinking when understanding requirements. For example, in the past, when we had conversations with smart speakers, we often encountered situations where the voice assistant said they didn't know, because the knowledge graph behind it was relatively static. In the future, with the support of generative artificial intelligence technology, intelligent speakers can make comprehensive judgments based on external information and provide answers in a probabilistic manner, which is more in line with the characteristics of human interaction with the world.

In addition, generative artificial intelligence is also expected to accelerate the development of emerging fields such as digital content creation, metaverse, and virtual humans.

Transfer of domestic artificial intelligence hotspots

In the past six months, domestic large models have attracted widespread attention. A large model refers to a machine learning model with a large number of parameters that can handle large-scale datasets during the training process. Large models are used to construct generative artificial intelligence models such as GPT.

On March 16, Baidu launched ERNIE Bot, which carries Wenxin's big model, to benchmark ChatGPT. Baidu calls ERNIE Bot "a new generation of knowledge enhancement big language model", which has the ability of human feedback to enhance learning, knowledge enhancement, retrieval enhancement and dialogue enhancement.

After Baidu, a large number of giant companies followed up and released big models. On April 8th, Huawei updated the Pangu model. On April 10th, Shangtang Technology launched SenseChat. On April 11th, Alibaba launched Tongyi Qianwen. After that, 360, ByteDance, iFLYTEK, JD, Tencent, Meitu... almost all Chinese Internet giants have gone down one after another. For Internet giants, this "big model war" has to be fought, because it has a direct bearing on the market's judgment of enterprise development prospects.

Looking at the world, China is by no means lagging behind in the number of large models. According to the "Research Report on China's Artificial Intelligence Large Model Map" released by the Artificial Intelligence Development Research Center of the Ministry of Science and Technology at the end of May, 79 large models with a parameter scale of over 1 billion have been released in China, indicating a great potential for a "hundred model war".

In contrast to the enthusiastic market demand, the user experience of domestic large models did not impress the four seats. For Baidu, which currently has the fastest development of artificial intelligence technology in China, there are still many immature aspects. The market is more reflected as being average.

In terms of training data, there is also a certain gap between domestic large models and ChatGPT. Geek circles speculate that the data size of ERNIE Bot is about 4TB, while ChatGPT is estimated to exceed 45TB. Industry insiders believe that although many companies have announced the integration of large models, there is still some distance between the practical application of universal large models.

Domestic Generative Artificial Intelligence Still on the Way

The switch in the field of artificial intelligence is sudden, and many domestic large models have not been fully trained yet. The ChatGPT that people see now is actually the result of years of accumulation. In 2019, when OpenAI founder Sam Altman took over the company completely, OpenAI's funding chain was on the brink of collapse, and the industry was almost not optimistic about the General Motors big model race. Because the application scenarios of such large models were not yet clear at that time, supporting such a large-scale model product was very costly, which was a huge challenge for any company.

The later story is widely known, as Microsoft successfully "fed" ChatGPT with billions of dollars, reigniting people's interest and expectations for big models.

Currently, China is developing rapidly in the field of artificial intelligence application technology, such as computer vision, intelligent speech, natural language processing, etc. However, the pace of development in basic technology fields such as underlying algorithm theory and development platforms is slow. Liu Chen believes that this is largely due to China's insufficient scientific research capabilities in the field of basic technology and insufficient reserves of low-level technical talents.

In the middle of June, Wu Hequan, an academician of the CAE Member and an authoritative expert in the field of communication in China, said publicly that there were not few technicians engaged in AI research in China, but there was still a shortage of talents with architecture design ability and generative AI data training facilitator level. In the 2022 list of the most influential scholars in artificial intelligence released by AIMiner, the United States has an absolute advantage in the number of top researchers in the field of machine learning, ranking first in the world, while China only ranks fourth.

This poses a series of difficulties in the localization process of generative artificial intelligence, from research and development to actual application of basic technologies. For example, Liu Chen said that China's deep learning framework represented by PaddlePaddle is developing into an open source platform that is more suitable for industrial needs and more popular with Chinese developers. However, due to the lack of basic capabilities, and a technology needs at least three to six months from the laboratory to the industrial landing, the technology and functional experience of domestic frameworks cannot meet the needs of developers, which makes it difficult to foster a prosperous AI development and application ecosystem.

Perhaps the popularity of generative artificial intelligence is not so important. Only by starting from the source can we truly seize the initiative of the wave of artificial intelligence technology.

At this World Artificial Intelligence Conference, more than 30 domestic and foreign models were unveiled at the exhibition, including Baidu ERNIE Bot, Shangtang Ririxin, iFLYTEK Spark, Huawei Pangu, Tsinghua University ChatGLM, Fudan University MOSS, etc., as well as the main and sub forums focusing on the fundamental issues such as the technology evolution and evaluation standards of large models. Some companies, such as Huawei, have announced the release of an integrated solution for large model training and promotion, which will help solve industry pain points such as high development and application costs, long cycles, and high technical difficulties for large models. These efforts will promote the better consolidation of the foundation and coordinated development of domestic large-scale models, and jointly build a sound development ecosystem.