In a market composed of stock funds, the squat jumping and bungee jumping of individual stocks are initiated by small compositions, and the style and track depend on marginal incremental funds. When the share of public funds issued fell back to the level of 2018 and 2014, those small towns that were favored by the era β became the problematic ones. With the framework of "good company and good price", the C managers who once became gods in the star-making movement of fund companies have already Began to lose the pricing power of AH shares. I used to believe in light, but now I dare not AI.
Fund managers who have come out of the ivory tower of the finance department do not know how to translate the mathematical reasoning of the AI big model into Mandarin; The general large model can understand all fragmented markets. The result is that AI institutions will not buy, retail investors will not follow, hot money will not withdraw, and the sector index will not fall. Fear is planted, greed is blooming, and in the end, good luck is bad luck.
After AMD began to challenge Nvidia, the PK between Su Ma and Lao Huang also accelerated the duel of domestic AI large models. Large-scale model manufacturers that rely on acquisitions are regarded as integrated companies, and friends who are overwhelmed by divorce can no longer afford the traffic of model update live broadcasts. The only remaining option has already told the story of AI for N years, and now it is ugly. "Education" attracts new investors, how many people can take heavy positions in Internet giants? So, is it possible for an AI large model listed company to have one more investment option?
In the previous article, we talked about the arrival of the era of AI large models triggered by the double decline phenomenon. ("Challenge: What is the strength of Bairong Cloud's AI large-scale model?") On this basis, we will then explain other elements in the development process of the large-scale model. The reason why the large model can be continuously upgraded is that when the training volume of the model is less than 10 to the 22nd power, the accuracy rate on several natural language processing tasks is around 0, and when the training volume exceeds 10 to the 24th power, The accuracy of the model began to increase significantly, which is the "emergence" effect of the model in the scientific literature, and it has become the reason why the large model is imperative.
The pre-training we often talk about is actually training a large model with unlabeled data so that it can learn general characteristics and structures. Supervised fine-tuning with standard data on the pre-trained model is Finetune that can complete tasks in specific application fields.
Therefore, according to the above principles, the AI large model manufacturers that can be selected must have sufficient training volume, completed pre-training models, and unique labeling data. Only by being able to open up the entire business line process in the vertical field can AI help enterprises complete the online closed loop, and then complete the privatization deployment. In the field of financial AI, if AI can only help the financial industry to complete the preliminary screening of loan applicants, but reject users with abnormally high number of loan applications caused by inter-industry diversion, and complete without a complete data feedback mechanism Automate approvals and model revisions. The large model that Bairongyun will make is to solve these problems.
Banks should expand loan scale and increase performance on the basis of guaranteeing loan recoverability. Taking the case of Bairong Cloud’s banking industry as an example, applicants’ qualifications need to be screened before lending. If a user handles credit business on multiple loan platforms, multiple credits will be formed. Therefore, if the frequency of past loan applications is abnormally high, hit The Bank of China will review such situations as medium risk, frequent application for loans during non-working days, and abnormal number of repeated applications in a short period of time. Bairong Cloud's pre-loan long-term products cover more than 90% of the credit card customer base, and the KS value used to measure the difference between good and bad samples has reached 0.5, increasing the number of loans; the loan process needs to screen applicants Due to the potential risk of multi-account transfer, after the loan, it is necessary to formulate different reminder messages and speeches according to the repayer's monthly payment progress and repayment period.
How about the empowerment and effect of Bairong Cloud's AI to the banking industry? For example, Internet loan regulation has entered the second half, and some banks are faced with the need to switch from offline to online to break their dependence on a single channel. However, the credit strategy of the bank's traditional offline loan model relies heavily on strong rules and expert experience. To put it bluntly, it is no longer applicable to online automatic approval. The coverage of the original rule strategy is less than 50% of the online channel. At this time, hundreds of Assisted by Rongyun's AI technology, through AI algorithms such as multi-fork tree, logistic regression, and LightGBM, in the process of joint modeling with banks, Bairong Cloud helps banks screen or derive highly differentiated credit index variables, and builds a model based on This regular model is applied to all aspects of pre-loan lending. In the process of some bank project cooperation, Bairong Cloud has developed a total of more than 2,000 derivative fields, and finally helped banks increase the approval rate of loan automation from 40% to 80%, which is equivalent to doubling the efficiency. More importantly, the customer's feedback on the application process of the AI large model will be the source of information for Bairong Cloud to continue training the large model.
Bairong Cloud, which has been established for 9 years, uses the behavior information of loan applicants to provide banks with loan approval business. It is the largest business "intelligent analysis and operation". Just need to catch up.
According to the current performance, Bairong Cloud achieved revenue of 566 million yuan in the first quarter of 2023, a year-on-year increase of 25%; the revenue of "intelligent analysis and operation", the largest business based on the self-developed AI platform, increased by 41% year-on-year to 281 million yuan Bairong Cloud's own AI product line "intelligent operation service", revenue increased by 68% compared with the same period last year; the 22-year annual report shows that the company's cash and equivalents totaled nearly 4 billion yuan. Manufacturers who make large models for money.
How do you hear that in the market in 2023, there are still people who want to enter the market on the right side of the institution? This year's paradigm is to make arrangements before institutions adjust their positions. Compared with buying from institutions, it is better to copy the bottom of institutions. There is nothing wrong with this, right?