"Mockups" are nothing new.
Models of all sizes existed in the industry before this wave of big model booms. However, since most of the industrial fields involve key decisions, it is more cautious in practical application.
With the blessing of large models, the application of artificial intelligence in the industry has ushered in a new opportunity. But in this era of "models" dancing wildly, how far is the large-scale model from the large-scale application in the industry? What are the real core technologies that need to be broken through?
With these questions in mind, we hope to take the opportunity of the WAIC 2023 Global Artificial Intelligence Conference to find a front-line industrial practitioner and explore the answers. Immediately, consumer finance came into our sight.
According to Jiang Ning, deputy general manager and chief information officer (CIO) of Ma Ma Consumer Finance, Ma Ma Consumer Finance was established in 2015 and currently has about 3,000 employees, of which 2,000+ are in research and development, and most of them are artificial intelligence and big data Direction, adhere to independent research and development and not outsourcing, as of the end of June 2023, 1315 invention patent applications, 983 have been published, accounting for 52.9% of the consumer finance industry, ranking first in the industry, a sound technology-driven financial institution. The company currently has 100,000+ risk characteristic variables, 100+ subdivision iterative models, and 2,000+ risk strategies, decision-making and data model algorithms, achieving a 92.4% smart customer service diversion rate, 100% smart quality inspection coverage, 98.6% overall customer satisfaction.
Jiang Ning, Deputy General Manager and Chief Information Officer of Mashang Consumer Finance
Four major challenges , torture the pain points of large-scale model application in the financial industry
As a pioneer of technology-enabled business, the financial industry has always been the vane of new technology application practice. How new technologies are used in the financial industry has always attracted attention. Large models are of course no exception.
In Jiang Ning's view, at present, the application of large-scale models in the financial industry faces four challenges.
One is mission-critical and dynamic adaptability . Critical tasks are often related to life or major assets, such as autonomous driving and bank deposits. It is not easy for artificial intelligence to be 100% accurate. As for the generative model, the chemical reaction combined with the discriminant model in the industry has no obvious effect at present. At the same time, unlike the closed system of AlphaGo, if the big model wants to become more and more intelligent, it must find a way to let the power of the group contribute to the model in an open system, so that as many users as possible can participate in feedback and form an ecology.
The second is personalized requirements and privacy protection . It is also a problem that needs to be solved to abandon the same dogmatic service and achieve thousands of people. The large model can provide users with a personalized experience while ensuring user privacy.
The third is swarm intelligence and safety and controllability . On the one hand, China's mobile Internet and PC Internet are separated, resulting in limited available data; on the other hand, data in industry fields is often closed and difficult to share. How to design a safe and controllable mechanism for cross-industry and organizational data sharing and equity sharing is also a huge challenge.
The fourth is infrastructure capacity . The basic environment such as GPU, network, and computer room must be modified to meet the needs of large-scale model training and reasoning in order to maximize its effectiveness. This challenge will hopefully be the first to be solved.
Three major technologies , decoding the direction of large-scale technology in the financial industry
In response to the first three challenges above, Jiang Ning gave the three real core technologies of the large model that he understands.
One is continuous learning . This is the biggest change brought about by the big model, and it is the key to the success of the big model. Regardless of whether it is thousands of models or ten thousand models, the most important thing is not the model parameters, but whether the large model has the ecological capabilities, so that as many people as possible can use it, become smarter the more it is used, have continuous learning capabilities, and provide positive feedback.
The second is robust decision-making . In the industrial world, including autonomous driving, medical care, finance, etc., although 99% of artificial intelligence is correct, as long as 1% is wrong, it cannot be used in the industrial world because it involves major decisions involving human life. Therefore, it is necessary to help the large model eliminate noise and interference, and maintain decision-making stability, safety and compliance in unexpected and unpredictable situations.
The third is combined AI . There were many models in the industry before. After the emergence of the large model, the generalization ability of the large model was used to disassemble the work, so that the generative model can be effectively combined with the original discriminative model in the industry, so that the advantages of the two models can be better utilized. Good value, automates GPT.
"Forming a new AI system that can implement continuous learning, robustness, and combination in the real industry is the direction for us to build a large model that can be effectively used in the industry in the future." Jiang Ning said.
Three Verticals and Three Horizontals , Immediate AI Practice in Consumer Finance
In Jiang Ning’s view, Immediate Consumer Finance, as an AI-driven financial technology company, must help solve three common problems in the financial industry.
One is personalized service and ultimate user experience . Financial services are characterized by high value and low frequency, making it difficult to provide personalized experience. The traditional approach is to label in layers to provide different products, but the labels are dynamic, so how to make automated decisions to provide customers with an automated experience requires that the products be different.
The second is efficient value transfer efficiency . Although finance is a technology-intensive and data-intensive industry, there are still a large number of offline outlets at present, because artificial intelligence cannot yet be error-free, and many tasks are highly dependent on manual operations.
The third is decision intelligence for compliance and safety , which is related to robustness.
Based on these three considerations, as well as tens of thousands of servers, nearly a thousand GPU cards, and 40PB of data in the form of text, sound, pictures, and videos, Immediately Consumer Finance has realized a closed loop of computing power, algorithms, data, and scenarios, and As a result, an AI strategy of "three verticals and three horizontals" has been created.
"Three Verticals and Three Horizontals" Strategy of Immediate Consumer Finance
three vertical
- Real-time human-machine decision-making : Solve the robustness problems in the industry, and let people take over 1% of the problems that artificial intelligence cannot solve.
- Multi-modal large model : It has accumulated a large amount of language, text, pictures and other materials, and conducted a lot of training around them, forming a complete set of multi-modal assets.
- Data intelligence : 2000+ models have been built to provide automated marketing, risk control and other services for more than 200 million users. This is a scarce scenario and resource in itself.
three horizontal lines
- Continuous learning : Make sure that what you do is not one-time, and the more you use it, the smarter you become.
- Model control : Solve robustness, suddenness, and unexpected problems, make the model have a stable output, automatically find noise, and effectively sort out harmful data that violates social ethics.
- Combined AI : Multiple models are combined and applied to solve problems.
At present, the AI of Immediate Consumer Finance is mainly applied in three scenarios: one is financial intelligent dialogue, which realizes real-time human-machine collaboration, continuous learning, trustworthy security and compliance; the other is financial digital human, through large model + combined AI multi-mode The third is the AI heart engine of financial services, through the organic combination of large model brain and psychology, to realize the human-computer experience with emotion.
"The combination of multiple models such as discriminative models and generative models in the vertical field, to build an open continuous learning, robust, compliant and safe system, is the real implementation of large models, rather than a model relying on thousands of Billion parameters to implement. We will work hard for it." Jiang Ning concluded.