With the development of ChatGPT set off a boom in the development of large models, the artificial intelligence industry has become more popular, and the industry, academia and the public are highly concerned.
On July 2, the Global Digital Economy Conference, co-hosted by the Beijing Municipal People's Government, the Ministry of Industry and Information Technology, the Cyberspace Administration of China, the Ministry of Commerce and the China Association for Science and Technology, was held in Beijing, including the Artificial Intelligence Summit Forum, focusing on hot issues such as general artificial intelligence and the development of large models.
At present, more than 80 large models have been publicly released, but the commercialization of large models is the most critical. The participants of the forum put forward their ideas on the commercial application of large models: large models must be universal, empower hundreds of industries to lead the industrial revolution, and must contact specific application scenarios to fully tap commercial value.
Large models must be universal
Previously, some voices pointed out that the big model is a vent and a bubble. "The big model is true intelligence, which marks the arrival of general artificial intelligence." Zhou Hongyi, founder of 360, said, "The big model directly improves the productivity of everyone and every organization, while achieving empowerment. When the computer was first invented, it did not lead to the industrial revolution, and it was far from ordinary people and ordinary enterprises, and it was only after computers entered homes and businesses that they changed the world. ”
Zhou Bowen, founder of ChatGPT Technology, pointed out that ChatGPT has brought artificial intelligence to a new inflection point, and the essence of AI should be to interact with people, learn from interaction, and then cooperate with humans to solve problems.
He pointed out that the evolution of artificial intelligence is divided into three stages:
Phase one is ANI, narrow artificial intelligence, which has been realized. This stage of the machine is highly supervised and requires a lot of manually labeled data; The scope of the task is narrow and new patterns need to be trained for new tasks.
The second stage is ABI, generalized artificial intelligence, which is the current stage. Self-supervision can be achieved without too much explicit teaching; End-to-end, i.e. one model can accomplish multiple tasks, as well as skills such as emergence capabilities, zero-shot learning, and more.
Phase three is AGI, general artificial intelligence, which has not yet arrived. At this time, artificial intelligence is smarter than humans, and always learns to become smarter, independent and autonomous, and may get out of control.
"ABI-stage AI is already powerful enough to create value without worrying too much about negative impacts. We should try our best to deepen the research of artificial intelligence and prepare for the future AGI. Zhou Bowen pointed out.
According to McKinsey's 2022 global artificial intelligence research report, the utilization rate of Chinese intelligence temporarily lags behind the global average, and there is still much room for improvement compared with the world's leading countries. Only 9% of Chinese companies can achieve more than 10% revenue growth with AI, compared to 19% of companies surveyed in leading countries.
At the forum, the consensus of the participants was to seize the opportunity of this round of artificial intelligence development and use large models to empower the industry.
Zhou Hongyi believes that the real opportunity for the development of large models is in China, and the government, cities, and enterprises have a broad application market, and China's big model should seize the opportunity of industrial Internet development. He also pointed out that for the industry, only when large models enter thousands of households and empower hundreds of industries can we truly promote the revolution brought by artificial intelligence.
Some people compare the big model to an operating system that most companies don't have the opportunity to use. Zhou Hongyi believes that in the future, there will not be only one or two large models, but like databases, it will become the standard configuration of every digital system, from mobile phone deployment to car deployment, but also internal deployment in enterprises and governments.
Multiple vertical model combinations
How to use big models to empower industrial development? The guests gave their own ideas for solving the problem.
From the hardware level, the answer given by Zhou Hongyi is to make a large model that is industrialized, enterprise, vertical, miniaturized, and proprietary.
He bluntly said that there are "four problems" faced by the implementation of large models in enterprise-level scenarios: First, the public large model lacks industry depth. The public large model is not closely integrated with the internal business of the organization, and the knowledge is not shared, which cannot meet the vertical and professional requirements of enterprise-level application scenarios.
Second, there are data security risks in the public large model. The organization's internal knowledge base is not suitable for training into a public large model; The public large model is easy to cause internal data leakage.
Third, the public model cannot guarantee the credibility of the content. Large models have "illusions", which cannot ensure that the content is authentic and documented; The internal data update iteration speed is fast, and the public large model cannot realize real-time knowledge update.
Fourth, the public large model cannot achieve controllable costs. The cost of directly training and deploying large models with 100 billion parameters is too high, and enterprise-level applications should use 10 billion basic models to train different vertical models according to different requirements, while enterprises only need to bear the cost of vertical training.
"For many enterprises, the cost of training enterprise-level large models has been drastically reduced, and our goal is to 'pull the big models off the altar' and turn them into something that every enterprise and government department can directly use." Zhou Hongyi said.
What kind of big model is needed? How to solve the above problem?
Zhou Hongyi believes that large models should do the following: first, industrialization, data with in-depth training in the industry is valuable; Second, enterpriseization, which needs to cooperate with the internal knowledge base of the enterprise to achieve real-time iterative updates, so as to ensure that the large model understands the enterprise better; The third is verticalization, "do not try to solve all problems with a large model, the large model must be a combination of multiple vertical models in the future landing form of the enterprise, and the vertical model has a stronger ability to solve professional problems." Zhou Hongyi said.
The fourth is miniaturization, making small-scale large models, large models with tens of billions of parameters have lower costs, and deployment and upgrading are more flexible; Fifth, proprietary, proprietary deployment can ensure security and controllability.
Fang Han, CEO of Kunlun Wanwei Technology Co., Ltd., pointed out that high-quality data is crucial to the development of large models. At present, the development prospect of domestic large models is to pay equal attention to To B and To C strategies, but large models must be matched with high-quality industry data on the B side, but there are currently problems of fragmentation and segmentation of industry data.
"Frankly speaking, the last three years of large model training have accumulated the ability to deeply process rich pre-training data. All OpenAI public papers and lectures are public about the training process and training algorithms, but it never discloses the model structure and data processing. Fang Han pointed out that at present, the world's large model pre-training teams are trying to reproduce OpenAI's actions in the model architecture and pre-training data, and the pre-training data processing ability of any enterprise is crucial.
He said that pre-trained large models have extremely high requirements for industry data quality, and only companies with base large models and pre-training data processing can quickly customize industry models.
Look for scenes
In terms of business scenarios, the participants believed that large models should fully touch application scenarios.
Zhou Bowen said that whether the big model can be fully integrated with the business to truly solve business problems is a key factor determining whether AI can achieve economic value. Only the AI strategy design, perfect supporting architecture, sufficient AI talents and sound internal training mechanism that closely follow the business can AI be fully integrated with business development needs and maximize economic benefits.
He put forward his own insights from the aspect of consumption scenarios, "B2C direction, B-end enterprises lead production and sales, it is difficult to meet the individual needs of C-end users, the cost of trial and error is high, and the backlog and waste caused by slow sales of goods; If it is a C2B perspective, consumer-centric, C-end users dominate, and enterprises produce on demand according to consumer demand, which can not only increase sales, but also reduce backlog waste. ”
Zhou Bowen said that in specific application scenarios, the binary management of consumers and products can be reconstructed with a large model, and the data of real-world consumer and product interaction can be used to understand consumers' thoughts and emotions, and gain insight into the state and experience of consumers, focusing on 5D (opportunity insight Discover, explosive product definition Define, product design design, specific research and development Develop, reach conversion Distribute), Compress all consumer and product information into the model.
This round of AI transformation has brought about a fundamental change in the way humans and computers interact. Tu Weiwei, vice president and chief scientist of Fourth Paradigm, believes that from the B-side, this is a very big opportunity. "For a long time, the market user experience for enterprise software was terrible, in most cases, 80% of users only used 20% of the features, and the value of the software was far from being discovered. If there is such a more natural way of human-computer interaction, the threshold for software use and the efficiency of development and use will be much lower. Large models need to be more optimized for such vertical scenarios. He pointed out.
From the perspective of investment, Ji Haiquan, executive director of Legend Capital, pointed out that the primary investment in the large model field is the foundation technology, but investors do not expect the application technology of the base to mature rapidly, and the division of labor in the large model field will not occur in the short term; Secondly, the investment focuses on the vertically integrated large model, from the base to the upper application is completed by one company, and a series of vertically integrated large model companies will appear in the future finance, education, automotive and other industries. As the large model base gradually matures, similar to the public cloud, some cloud-native AGI native companies may be produced in the future, and Killer Apps like ChatGPT may also appear.
For more information, please download 21 Finance APP