Editor's note
The AI wave has reached a new stage of development that transcends the singularity and aligns the future.
On June 6, this year's 21st Century Excellence Board of Directors Artificial Intelligence closed-door meeting gathered more than 20 industry experts, scholars and executives of listed companies to discuss core topics in the field of AI and conduct in-depth discussions on innovation, technical problems and development paths in the field of AI.
In recent years, with the continuous evolution of large model technology represented by GPT, "versatility" has become the latest footnote in the development of artificial intelligence. In the industry's view, general artificial intelligence technology can give new impetus to industrial development.
Recently, at the "21st Century Excellence Board Artificial Intelligence Closed-door Meeting" hosted by 21st Century Business Herald, Beijing Artificial Intelligence Industry Alliance Metaverse Professional Committee, and China Cultural Industry Association Cultural Metaverse Professional Committee, a number of experts, scholars and business executives discussed "the emergence of innovation, technical problems and development paths of listed companies in the field of AI".
For the development of general artificial intelligence technology, Lu Feng, deputy director of the Electronic Information Research Institute of the CCID Research Institute of the Ministry of Industry and Information Technology, said, "Artificial intelligence has developed from special artificial intelligence to the critical point of general artificial intelligence, and we must attach great importance to the future era of artificial intelligence, and artificial intelligence applications will be everywhere." ”
At the same time, a large number of science and technology enterprises in the mainland are successively laying out large-model technology and industrialization, and the application industry is also accelerating its expansion from office, life, entertainment and other directions to medical, industrial, education and other fields. At the meeting, a number of participants discussed the topic of application landing exploration and challenges.
Image source: Xinhua News Agency
General artificial intelligence is on the rise
According to the "Research Report on the Map of Chinese Engineering Intelligence" compiled by the China Institute of Science and Technology Information and the New Generation Artificial Intelligence Development Research Center of the Ministry of Science and Technology and relevant research institutions, China has entered a period of rapid development of large models since 2020, and has now established systematic research and development capabilities covering theoretical methods and software and hardware technologies in large models, forming a large model technology group that closely follows the world's frontier.
From a policy perspective, the Politburo meeting on April 28 clearly pointed out that it is necessary to attach importance to the development of general artificial intelligence and create an innovative ecology.
The first meeting of the 20th Central Finance and Economic Commission on May 5 once again emphasized the need to grasp the wave of new scientific and technological revolutions such as artificial intelligence.
In order to promote the innovation and development of artificial intelligence, Beijing, Shanghai, Shenzhen and other places have issued relevant support measures to gather artificial intelligence innovation resources, seize the opportunities of large models, carry out research on large model innovation algorithms and key technologies, and create an artificial intelligence innovation highland.
According to the China Academy of Information and Communications Technology, the scale of the core artificial intelligence industry in mainland China will reach 508 billion yuan in 2022, a year-on-year increase of 18%. Compared with vertical AI, general AI is not limited to processing a single task, and realizes cross-domain, interdisciplinary, cross-task and cross-modal applications.
In Lu Feng's view, artificial intelligence has experienced several highs and lows, and is currently experiencing a very critical moment of the third high. With the continuous breakthrough of artificial intelligence technology, computing power and algorithms, some basic conditions have been laid for the leapfrog development from digitalization, networking to intelligence.
Therefore, he believes that full attention should be paid to the urgency of the application needs. "In the future, due to changes in the demographic structure in many fields, due to the unadaptation of many new generations to the traditional working environment, machines must be used to replace people." Lu Feng said.
According to IDC's forecast, enterprise investment in artificial intelligence technology will reach $154 billion in 2023, an increase of 27%, which may mean that artificial intelligence will release huge market demand in the future.
Zhou Chengxiong, a researcher at the Institute of Science and Technology Strategy Consulting of the Chinese Academy of Sciences, believes that there are two logics at work in the development of artificial intelligence, one is technical logic, and the other is policy logic, and it can be said that the current policy logic plays a great role in it.
He pointed out that in terms of Chinese intelligent policies, including "Made in China 2025" and "New Generation Artificial Intelligence Development Plan" have been introduced one after another. Recently, the Cyberspace Administration of China (CAC) issued the Measures for the Administration of Generative Artificial Intelligence Services (Draft for Comments), which will regulate generative AI, including social ethics, personal information, and data.
Zhou Chengxiong believes that the value of social subjects lies in accelerating the liberation of human physical strength, brain power and innovation ability, and now artificial intelligence is the liberation of accelerating human innovation ability. He further pointed out that in recent surveys, it has become clear that artificial intelligence is gradually playing more roles.
"Now some experiments in university laboratories are completed by robots, some artificial intelligence technologies can read tens of thousands of documents and write a literature abstract, and some can write academic papers through keyword input." Zhou Chengxiong said.
Technology and industrialization align with the future
"Whether it is the metaverse or AIGC, it is improving efficiency. In the entire application process of AIGC, there is a lot of room for the application of pendants. In fact, after the emergence of GPT, some previous artificial intelligence models can still be optimized. Yan Yang, director of the Metaverse Special Committee of the Beijing Artificial Intelligence Industry Alliance and deputy director of the Network Information Intelligence Center of the National Engineering Laboratory for Big Data Analysis and Application Technology, pointed out.
Yan Yang believes that artificial intelligence must not only have technological breakthroughs, but also breakthroughs in business models. IN THE PROCESS OF BREAKING THROUGH THE TECHNICAL BOTTLENECK, WE MUST FIND OUR OWN BUSINESS MODEL, MANY INSTITUTIONS THAT ARE NOT LARGE-TRAFFIC PLATFORMS WILL DEFINITELY NOT BE ABLE TO DO THE METCALFE EFFECT OF THE OPENAI PLATFORM ON THE C-END MARKET, BUT WE MUST PAY GREAT ATTENTION TO THE TWO-SIDED MARKET EFFECT OF THE WHOLE PROCESS IN THE B-END MARKET.
He said that from a technical point of view, GPT, the Transformer model it relies on most fundamentally, has a lot of underlying point product calculations involving vector matrices, so there is still a lot of room for optimization when processing these operations through hardware. Therefore, considering the overall impact of storage, transmission and computing, general-purpose GPUs are likely to begin to shift to AI acceleration that specifically supports Transformer.
In addition, Yan Yang believes that according to the basic principle of Transformer, many models can be optimized, such as not necessarily need to add encoders and decoders at the same time, but need to match according to the scene. In the future, in the whole ecological process, it may be a herd of heroes chasing deer.
But he also pointed out that the entire GPT model is not the best to do it all at once, because according to relevant institutional tests, the big data model of the large model will also have a large loss during training, but if it breaks through from both ends of the U-shape (that is, small data large model or big data small model), it may be an opportunity for China's "sardines" to come out of the circle.
In his opinion, the financial industry may have a lot of room for application. Now the financial industry is less used, partly because some large models are not very friendly to accurate data calculation, but now the open source model is very effective for structured data after a slight modification. In previous years, logistic regression could achieve this level, and now with the blessing of large models, the threshold for use has been greatly reduced, and there have been laboratory cases overseas for predicting small-cap stock indexes.
In addition, in the field of scientific research, he believes that large models still have a lot of room in structural biology, although Alphafold2 has made great achievements in protein structure prediction before, but there are still many aspects that need to be improved involving insufficient computing power and data, which has brought a very large imagination space for the mainland to be used in this field. In the IT field, DBA (database administrator) work is also relatively complex, and now through Text to SQL, DBA work can also be improved.
"We are also doing localization, the first step is to do open source, for example, after we transfer the relevant code, we can form our own open source ecology, for example, now the NPCs in the virtual space can be formed with AIGC, support commercial partners to use directly, if they are very convenient to use, but also very cheap, then you can start cyclic iteration." Yan Yang said.
"It used to be said that the singularity has come, and now we expect: cross the singularity and align the future." Yan Yang said.
Application landing exploration and challenge
In the process of generalization of artificial intelligence in large models, enterprises are more based on their own background and industrial advantages. At the meeting, a number of leaders of technology companies mentioned the challenges faced in the process of R&D and application landing.
Kunlun Wanwei (300418. SZ) CEO Fang Han said that in terms of commercialization, the company adopts a strategy of paying equal attention to To B and To C. It is reported that in December 2022, Kunlun Wanwei officially released "Kunlun Tiangong", and its AI generation capabilities have covered content modes such as images, music, text, and programming. On April 17 this year, Kunlun Wanwei officially released the Tiangong University Language Model.
On the B-side, Fang Han concluded that how to generate industry data into data usable by large models is the most difficult thing. He said that many domestic enterprises lack better data, and the country now wants to apply To B, to help all industries to supplement the existing knowledge derivation process, not only the topic and answer, but to list the derivation process, so that various industries can use it.
When it comes to overseas market expansion, he believes that Chinese companies are most suitable for the C-end market, and they can do end-to-end content generation tools in the overseas C-end market. "This sounds simple, but in fact, all AIGC tools today are not end-to-end content production tools, but material production tools." Fang Han said.
In addition, in Fang Han's view, there is still a technical difficulty in end-to-end content generation tools that has not been solved, once the consistent content generation is solved, the entire film and television industry and short video industry will be subverted, and consistent video generation is expected to make a breakthrough within 1-3 years. In 3 years at the latest, humans will be able to use AI to generate very consistent long videos.
Tolls (300229. SZ) is expected to be launched by the end of June. At the meeting, Lin Songtao, vice president of Tolls, said that the challenges in the landing of large model technology scenarios are mainly quality, controllability, timeliness and cost.
In terms of quality, he believes that personal service with ChatGPT, the prompt word is not allowed to change one, if the picture generation is not good, you can also change another one, it can be said that the personal tolerance for AIGC is very high. But companies are different, to write a consultant report to the government, the data source must be accurate.
At the controllable level, one is content security, data has values, and models have no values. Second, private domain data security, China's large model to be data-based, how to ensure the security of users' private data while making better use of big data is also a problem.
In terms of timeliness, in big data training, catastrophic forgetting has always been a training problem, so large models are needed to solve this problem. In addition, how to enter real-time data, similar to Chat GPT, the latest data is only up to September 2021, so it is not good to use it when serving To B.
In terms of cost, the privatization of the 100 billion model needs to be well trained, and the landing on the enterprise side also needs to make the enterprise affordable. Tolls does To B service, vertical scene landing has become the core point of "100-model war".
In addition, on June 3, Wenge released the Yayi big model, and launched the large model application in the fields of media, finance, and publicity. At the meeting, Wang Xiaodong, vice president and chief product officer of Wenge, also said that the current big model has four challenges: ecology, scene, evaluation mechanism, and algorithm.
"In terms of ecology, the domestic open source ecology is still relatively weak. In terms of scenes, there are still relatively few landing scenes, and there is a lack of scenes. How the user side contributes data is also crucial. A more scientific model evaluation mechanism in China is still lacking. The level of each model needs to be scientifically open and fairly evaluated, so as to standardize the market and promote healthy development. At the same time, in view of problems such as large model illusions, continuous innovation and breakthroughs in algorithms are also major bottlenecks. Wang Xiaodong said.
For more information, please download 21 Finance APP