From the current point of view, large models have been implemented first in the financial and energy industries. Generative applications such as intelligent customer service are greater than decision-making applications, and the implementation value is not as expected. However, with the development of technologies such as industry customization, model lightweighting, and data security, the application of large AI models will usher in a broader space for development.
Author | Doudou
Editor|Piye
Produced | Industrialist
Eight months have passed since the day ChatGPT broke out.
In 8 months, many large models have sprung up in China, penetrating into various industry scenarios at a rapid speed. But so far, there have been no scenes or industries that have been truly disrupted by big models.
Statistical data shows that in the implementation of large models,45% of companies are in the wait-and-see stage, 39% of companies are in the exploration and feasibility research stage, and 16% of companies are in pilot applications. stage, and there are zero companies that have fully applied it.
A question worth seeing is, what is the progress of the implementation of large models in China today?
1. Large-scale model implementation, finance and energy take the lead
"Please help me check my electricity consumption in August this year, and which day I used the most electricity?" "Help me detect any defects in this picture"... On the Southern Artificial Intelligence Innovation Platform, through language interaction, a user can Each item of data is clearly displayed in front of you.
On this platform, workers in the electric power industry can issue instructions to large power models, allowing them to automatically generate data processing results, accurately identify image details of defective scenes, and help workers retrieve and process data during electric power inspections.
Currently, in the field of customer service of China Southern Power Grid, 60% of high-frequency problems can be solved through large-scale power models. In terms of identifying customer mood swings, large-scale power models are even better than manual work.
In addition, in the field of power transmission and distribution,the large-scale power model has the ability to process 100 problem pictures per minute, and can also identify 20 types of defects at the same time. The identification efficiency is higher than that of traditional AI. 10 times the algorithm.
In the field of power dispatching, large power models can assist the dispatching department to quickly and automatically generate treatment plans for abnormal power grid situations, respond to power market regulation requirements in a timely manner, and make the plans safer, more efficient, and less costly.
This is a microcosm of the implementation of large models in the energy field.
It is understood that some leading energy manufacturers have started cooperation with technology companies in large-scale model applications, especially in the fields of power grids and mining, forming some preliminary pilot demonstrations, such as power grid dispatching, defect/fault inquiry, coal mine operation monitoring and other scenarios. .
In addition to the energy field,the financial field is also one of the largest implementation scenarios of large models.
A report from iAnalysis also listed energy and banking as the two industries with the fastest progress in the implementation of large models.
’s wide application in the financial field can also be seen from the number of large models and corporate dynamics. A set of data shows that as of August, the number of domestic large models with parameters above 1 billion is as high as 116, including about 18 large models in the financial industry.
In addition, in their semi-annual reports, nine banks including Industrial and Commercial Bank of China, Agricultural Bank of China, Bank of China, Bank of Communications, China Merchants Bank, China CITIC Bank, Industrial Bank, Hua Xia Bank, and China Zheshang Bank clearly stated that they are exploring the application of large models.
On the side of large model manufacturers, the intensive release of some industry models also reflects the popularity of financial scenarios.
For example, in late May, Du Xiaoman released the 100-billion-level Chinese large model "Xuanyuan"; in June, Tencent Cloud joined hands with China Information to cooperate on financial large models. The Agricultural Bank of China launched ChatABC, a large model application similar to ChatGPT, and the Industrial and Commercial Bank of China released Developed a general model for the financial industry based on Shengteng AI.
From July to August, with the official implementation of the "Interim Measures for Generative Artificial Intelligence Service Management", many companies including Tencent, Baidu, iFlytek, Huawei, ByteDance, etc. have successively released the latest progress in large models; 9 In August, Ant Group also officially released a large financial model and open sourced the generative AI programming platform CodeFuse.
The financial field is undoubtedly one of the most common scenarios for the implementation of large models.
Whether it is the energy field or the financial field,the reason why it can achieve the leading implementation of large models is due to some commonalities between these two industries.
First of all,The energy and banking industries are both highly data-based industries with good data foundation and digital environment, which provides favorable conditions for the training and application of large models. .
Secondly,both major industries have large data processing and decision-making needs,and large model machine learning and deep learning technology can help The industry solves these problems and improves decision-making efficiency and accuracy.
Furthermore,the business models of the energy and banking industries are relatively mature and have high commercial value.Therefore, these industries are very sensitive to large models. The demand for technology is also relatively large, which promotes the implementation of large models.
It can be seen that the two major industries of energy and banking are progressing relatively quickly in the implementation of large models, mainly due to various factors such as theirgood data foundation, large technical needs, and high commercial value. comprehensive effect.
It is worth noting that even in the two implementation scenarios of finance and energy, large models still have some problems that are currently difficult to overcome.
2. Failed to meet expected scene value
In the financial industry, the three directions of marketing, risk control, and operations are the application directions of large models that many banks pay more attention to.
Among them, intelligent question and answer assistants, intelligent customer service, automatic generation of marketing pictures, and post-loan report writing are the segmented scenarios that banks and other financial institutions are actively deploying. But for now, the value of generated scenarios such as intelligent question and answer assistants, intelligent customer service, and automatic generation of marketing pictures is almost the same as expected. However, in decision-making and native application scenarios such as waking up sleeping customers and digital business halls, large models There is still a gap between the expected implementation and the actual effect.
For example, in the smart customer service scenario, there were few smart sparring question banks in the past and lacked pertinence. Nowadays, the production of personalized question banks based on large models can shorten the training cycle; in the automatic generation of marketing pictures, designers used to select and design from the material library. Now they can be automatically generated using Midjourney, which can reduce copyright costs and labor costs.
In terms of the awakening of sleeping customers and the value expectation of digital business hall scenarios, the former uses large models to automatically generate strategies, which can achieve end-to-end strategies to improve the awakening effect; the latter uses digital humans supported by large models to help customers handle business and recommend products to complete transactions. New channels independent of APP can be realized.
However, at present,the practical application value of these two scenarios is not yet known.
The implementation of large-scale models in the energy industry is also similar.
In the energy industry, the value of generated applications such as equipment operation and inspection knowledge assistants and intelligent customer service is almost the same as expected. However, in scenarios such as maintenance documentation, equipment failure repair, and power load forecasting, the actual value of the scenario is still unknown.
Specifically, in the equipment operation and inspection knowledge assistant scenario, in the past, a structured knowledge base was built based on NLP technology. With the support of large models, large models can be used to build operation and inspection assistants to improve efficiency; in the intelligent customer service scenario, in the past, it was based on Intelligent customer service based on the Bert model now uses large models to improve the user experience of intelligent customer service, which can achieve more accurate understanding of intentions, more anthropomorphic language, and improved user experience.
In the scenarios of maintenance document generation, equipment failure repair, and power load forecasting, the values that can be brought by the implementation of large models are: rapid and automatic document generation to improve efficiency; large models can quickly locate the cause of faults and provide maintenance suggestions and solutions; and incorporate more influences Factors can be used to predict loads in real time and improve prediction accuracy.
However, at present, the value brought by large models in these scenarios is not yet known, and it still takes time to explore. It can be found that whether it is the financial industry or the energy industry, generation scenarios are implemented quickly and have many applications, while decision-making scenarios are slower and more difficult to implement and have fewer applications.
3. "Generating Scenarios > Decision Scenarios": Difficult to Transform Productivity
At present,the implementation of large models is still in the pilot application stage and is not fully launched.
As mentioned above, AI large model applications such as intelligent question and answer, intelligent customer service, digital business hall, post-loan report generation, sleeping customer awakening, and financial product recommendation in the financial industry have gradually been implemented; intelligent customer service and equipment operation and inspection knowledge in the energy industry Assistants, maintenance document generation, power system simulation platform, power load forecasting, etc. have been applied on a pilot basis.
However, consumer goods retail, securities, and media are still in the exploratory stage, and manufacturing and pharmaceutical companies are still in the wait-and-see stage.
It can be seen that although the implementation of large models is optimistic in terms of breadth, it is more difficult in terms of depth.
The depth of implementation of a large model depends on its capabilities, scale, computing resources, data quality, domain knowledge, etc. However, for the current large domestic models, they are still in the early stages of development, and many facilities and capabilities are still being gradually improved.
Limited by factors such as model capabilities and application effects, The current application is mainly based on generating scenarios.
Different from large decision-making models, generative large models are mainly used in fields such as text generation, dialogue systems, and language translation. By analyzing a large amount of text data, they learn the generation rules and intrinsic semantic relationships of text, so that they can generate high-quality text output. Representative models of generative large models include OpenAI’s GPT series and Baidu Wenxin Yiyan.
Large decision-making models are mainly used in recommendation systems, reinforcement learning and other fields. The data that needs to be processed usually contains continuous numerical variables, and decisions need to be made or future behaviors need to be predicted. Representative models of large decision-making models include DeepMind’s AlphaZero series and OpenAI’s Dota2 AI, etc.
Compared with large decision-making models, generative large models are first used in text generation and dialogue systems, where data can be collected and organized through a large number of text corpora, while in recommendation systems and reinforcement learning, data usually require manual design and construction, which is relatively More complex.
Secondly, research in fields such as text generation and dialogue systems is relatively mature, and there are many ready-made algorithms and frameworks available, while areas such as recommendation systems and reinforcement learning require more exploration and research.
In addition, text generation and dialogue systems have a wide range of application scenarios, such as search engines, chat robots, automatic writing, etc., while recommendation systems and reinforcement learning are mainly used in e-commerce, advertising, games and other fields.
One fact is that,Although generative scenarios are widely used, predictive decision-making scenarios are high-value scenarios in the future. Whether it is a large model supplier or an enterprise, if you want to improve business value based on large model capabilities, the latter is the direction of efforts.
4. In industry scenarios, look at the large AI model again
To implement a large model, you first need to choose the appropriate field and scene. Scenarios in this field have strong digital capabilities and digital foundation.
For example, in the field of intelligent customer service, you can consider applying large models to scenarios such as FAQ question and answer systems and chat robots; in the field of advertising recommendation, you can apply it to scenarios such as personalized recommendations on e-commerce platforms; in the field of public opinion monitoring, you can apply it to scenarios such as personalized recommendations on e-commerce platforms. It is used in scenarios such as content classification and sentiment analysis of news media.
Secondly, it is necessary to have high model capabilities and application effects. Judging from the current main paths for enterprise users to implement large models, group companies focus on building large model capabilities, while general enterprises/departments focus on application scenario exploration. Large model capacity building is divided into three levels: infrastructure construction, large model training and large model application. Currently, infrastructure construction and large model training are the main tasks, and large model applications are rare.
It is worth noting that the current application directions of large models are mainly divided into two types.One is mainly small models, and large models improve the development efficiency of small models; the second is large models Cascade with small models, connect applications with small models, and enhance the capabilities of small models with large models.
This landing path limits the model’s capabilities.
If you want to promote the depth of implementation of large models,large model suppliers and enterprises need to continuously explore capabilities and cooperation models.
Some paths to deepening the implementation of large models are gradually becoming clear.
In the future, with the continuous development and popularization of large model technology, the application of model cascade will become more and more widespread.
For example, multiple large models can be combined and cascaded to achieve more complex and accurate application scenarios such as speech recognition, image recognition, and natural language processing. At the same time, large models and small models can also be cascaded to give full play to their respective advantages and improve the performance and generalization capabilities of the model.
Based on this, the application depth of large model implementation is expanded and the application implementation of decision-making scenarios in various fields is accelerated.
Secondly, different industries have different specific needs, and large models need to develop in a more customized direction in the future. Through training on industry-specific corpora, large models can better adapt to actual application scenarios in different industries.
Secondly, in order to better meet the efficiency and resource requirements in practical applications, large models need to develop in a more lightweight direction. Through technologies such as model compression and pruning, model size and computing resource consumption can be reduced while ensuring model performance.
In addition, as data privacy protection issues become increasingly prominent, large models need to pay more attention to data security and privacy protection.
Model customization, lightweighting, and data security have become important factors for its implementation.
The implementation of large-scale AI models in China has achieved a series of results in areas such as intelligent customer service, advertising recommendations, and public opinion monitoring. However, also faced many difficulties during the implementation process. In the future, with the development of technologies such as industry customization, model lightweighting, and data security, the application of large AI models will usher in a broader development space.