With 100 million monthly active users within 2 months of launch, ChatGPT became the fastest growing consumer app in history.

How to understand the "explosion" of this phenomenon-level application?

What challenges are you facing in the development of new technologies?

How will artificial intelligence lead to new industrial changes around the world?

While the fiery new technology brings great changes, it also requires a calm and cold thinking.

There is still a long way to go before maturity

　　Recently, ChatGPT has become popular all over the world almost overnight.

Because of its logical and smooth dialogue and interaction capabilities, it has attracted much attention and discussions.

Before it turned out, the development of artificial intelligence was slow, and it was even interpreted as "getting cold", but now the new technology has once again attracted widespread attention from all walks of life, various investment institutions have heard the news, and related concepts have been hyped in the capital market.

　　"The outside world always regards ChatGPT as the point of this wave of innovation. In fact, this breakthrough point is not ChatGPT, but a large language model." Former executive vice president of Microsoft (Asia) Internet Engineering Institute, Xiaoice, who has long been engaged in the research and development of AI intelligent dialogue robots Li Di, CEO of the company, told reporters that the so-called big language model is an artificial intelligence model that uses large-scale text data for training and can generate natural language.

ChatGPT is not the technology itself, it is just the product and attempt of the "big model".

　　The success of the "Large Language Model" means that the technical bottleneck that was previously stuck in the development of artificial intelligence natural language processing has been broken, and it also indicates that this technical route will produce many innovations in the next five years.

　　Ma Lei, a researcher at Beijing Zhiyuan Artificial Intelligence Research Institute, said: "The investment community and the Internet industry are urgently looking for new growth points and innovation points. widespread attention."

　　Not only that, but the technical maturity of ChatGPT still has a lot of room for improvement.

Many artificial intelligence technology experts admitted that the debugging of large language models represented by ChatGPT is still in the stage of "manual workshop", and its basic technical standards and models have not yet formed an industry consensus.

　　Li Di said that "debugging" is the most important thing in this technology.

After training the model, to really make a model more usable and at the same time conform to business rules, at least 9 times the cost of capital and time will be spent on debugging.

　　"Why is the logic and accuracy of ChatGPT's English output results much greater than Chinese results? It is because OpenAI spends a lot of debugging energy on English data rather than Chinese content." He said.

　　The world's major business giants have accelerated their deployment.

Internationally, Microsoft has released the latest versions of the Bing search engine and Edge browser supported by ChatGPT.

Microsoft CEO Nadella said: "The search engine has ushered in a new era."

The US search giant Google immediately followed up and demonstrated the chat robot Bard developed by it in early February, and plans to promote it on a large scale.

　　Domestic companies are also stepping up their deployment. Baidu announced that it will launch the Chinese version of ChatGPT "Wen Xin Yi Yan" in March.

In addition, Internet technology companies such as Ali, Tencent, Xiaomi, Bytedance, and Kuaishou have also joined related fields to carry out research and development and deployment.

Some local governments also "followed up" and announced their support for leading companies to create large models that benchmark against ChatGPT.

The high cost of technology giants will reshuffle the industry

　　The industry believes that the technical potential demonstrated by ChatGPT will quickly penetrate into more service scenarios and become an important tool in multiple industries and fields, giving birth to disruptive and clearer applications.

But at the same time, the current technology is still in the early stage, and there are many problems in itself, which cannot be avoided.

　　First of all, the popularity of ChatGPT is mainly due to the change of technological paradigm, which brings new users, and there is no business scenario for mature application.

In the U.S., the vast majority of current users are students "writing paper assignments," while others are more early adopters.

At the same time, due to limitations in computing power and data scale, the research and development capabilities of academia are stretched, and the industry is difficult to invest large-scale funds because of the huge investment that may affect corporate profits.

　　The data shows that in order to make the language synthesis results of ChatGPT more natural and smooth, OpenAI shared 45TB of data and nearly 1 trillion words to train the model.

That's roughly equivalent to 13.51 million Oxford dictionaries.

The cost of training is as high as tens of millions of dollars.

　　In addition to the training costs, the running costs are equally staggering.

An insider familiar with the development of ChatGPT revealed that the consumption of computing power during ChatGPT’s online service, the company has to bear the cost of tens of dollars for each user every year, and the training to achieve better interaction quality requires a lot of Funding and hardware support.

　　Wu Hequan, an academician of the Chinese Academy of Engineering, believes that the improvement of computing power and large-scale application promote the generation of large pre-training models.

The emergence of ChatGPT has made the pre-trained large model enter the level of commercialization, which is a great progress.

But its application is still limited.

"It is still a kind of company that relies on a large amount of computing power and can afford such a large-scale computing power to support it. It is still a minority in the world." He said.

　　Whether using ChatGPT for customer service or as a character (NPC) in the game, the cost may be unbearable for the enterprise.

This has a precedent in the game industry. According to industry insiders, there was a game that introduced AI-driven NPCs before, but it was not popularized afterwards.

The reason is mainly because of the high cost. High-quality dialogue between NPCs and game players may generate operating costs of 2,000 yuan per user per year, which game companies cannot afford.

　　Secondly, how to form a new business model is still uncertain.

Among them, the problem of low accuracy of ChatGPT cannot be ignored.

　　Thirdly, the social risks of artificial intelligence technology in terms of false information, academic plagiarism, privacy disclosure, and public opinion guidance will be further magnified due to the huge scale of data used.

OpenAI also acknowledged that ChatGPT "occasionally writes plausible but incorrect or nonsensical answers".

　　In the face of challenges, why does ChatGPT still make major companies around the corner and invest in it one after another?

People in the industry believe that this is a major opportunity to subvert the existing business sector.

　　"For a challenger in the industry, he has the opportunity to take advantage of a technological paradigm revolution to acquire new users within a certain time window, and what he pays is the cost of acquiring customers." Li Di.

　　For large companies that firmly occupy the market, this is a "defense battle" that they have to participate in.

Faced with the dimensionality reduction blow brought by new technologies, if you do not follow up, you may only be eliminated.

　　According to Yang Qiang, former chairman of the Council of the International Federation of Artificial Intelligence and a professor at the Hong Kong University of Science and Technology, the development of "big models" is much faster than the industry imagined.

He believes that in the future, the business landscape of artificial intelligence will move towards two categories. One is large companies and institutions that have resources invested and can do pre-training. They do not have a specific task to make a model in advance and become a "winner takes all" in the market. " leader.

The other type is the vertical industry that trains small models for different application content, and many "small and beautiful" application scenarios will be born.

Suppressing bubbles, breaking the "time barrier" is more important than "technical barriers"

　　With the algorithm breakthroughs brought by ChatGPT, many companies have announced to accelerate research and development in the fields of robots, intelligent voice/vision, AIGC intelligent writing, etc. A new round of artificial intelligence applications based on content production is expected to continue to land, expanding pan-AI technology in Coverage in real life.

　　Wang Qiang, director of Tencent Frontier Technology Research Center, believes that the new technology represented by ChatGPT is expected to quickly land in multiple scenarios: in the field of search engines, through technology integration to optimize existing search models and provide more accurate information; in education, medical care, advertising marketing, In the field of professional services such as e-commerce, it provides content services and even replaces some primary professional work; it combines with new smart carriers such as intelligent networked cars, smart speakers, and smart TVs to meet the content needs of the audience.

　　Many experts believe that the emergence of ChatGPT may trigger a new round of artificial intelligence technology competition, and the global competition in the field of large language models has become increasingly fierce.

　　"OpenAI's success is due to its patience and loneliness to complete the debugging work." Li Di believes that what is most needed now is experience, talent, and patience.

　　Some industry giants and local governments began to show signs of "doing big and fast", and some investors began to dig talents.

Wang Huiwen, the co-founder of Meituan, posted on social media that he invested 50 million US dollars to build "China's OpenAI", hoping to recruit top R&D talents in the industry.

　　"Some people are now engaged in ChatGPT. Looking at half a year ago, they were still engaged in metaverse and blockchain. Many capitals lack the spirit of sitting on the bench and the spirit of craftsmanship. What needs to be broken now is the 'time barrier' rather than the 'time barrier' technical barriers'."

　　An industry insider said, "Some R&D personnel's wages are expected to double several times, but the products are still the same. The blind expansion of the competition will inevitably cause unnecessary waste of computing power resources."

　　At the beginning of February this year, a Chinese start-up company released a smart program claiming to declare war on ChatGPT and launched related products, but soon after, its WeChat applet could not be opened.

　　Yang Qiang believes that it is necessary to avoid industry bubbles, promote the stable development of the industry, and reduce problems such as over-development, centralized launching, and waste of resources.

At the same time, in view of possible academic ethics and social order issues that may arise in ChatGPT, the industry should be called on to conduct deduction, and methods such as filing and corpus screening should be adopted in technology research and development to avoid them.

(Reporters Guo Yujing, Ji Ning, Zhang Manzi)

　　"Xinhua Daily Telegraph" February 21, 2023 Edition 06