Commercial banks disclosed at the Shareholders ' Conference on 25 June that they were continuing to promote the application of large models in banking operations and had established a specialized cost-benefit measurement system. Management indicated that the key to the promotion of AI in institutions with more than 100,000 staff is to control the efficiency of inputs.
Operating average daily consumption 33 billion Token
The recruiters concerned indicated that, as at the end of May, the average daily consumption of Token, a large model produced mainly by the business sector, had reached $33 billion. Some of the scenes in retailing are relatively more visible.
At the same time, he gave an internal efficiency indicator: the ratio of hours spent by the recruitment AI processing business counterpart to hours actually worked by employees has risen from 1:13 at the end of last year to close to 1:9 at the end of May, indicating that the share of the larger model in part of the process is increasing.
Cost-benefit ratio of approximately 20 per cent
Zhou Tianxiang stated that prior to the introduction of the “AI First” strategy, the call had established a more complete system of cost-benefit measurement along the technology line. The cost end consists mainly of R & D inputs and Token costs, while the benefit end is measured at six dimensions.
Based on the current calibre of the line, the cost-benefits in the direction of the larger model are maintained at or above 20 per cent, i.e., "Inputs $20, which can generate $100." However, the banking chain is long, the precise quantification of the proceeds remains complex and the relevant algorithms, data sources and data quality continue to be adjusted.
Code generation input ratio is only 5%
In terms of resource allocation, recruitment has exercised relative restraint in the programming of large models. Zhou Tianxian indicates that over-budget Token consumption is more likely to occur at the code writing stage than at the business management application.
He mentioned that, at the current stage, the large model was still shortboarded when dealing with large software structures and that the readability, performance and safety of the generation codes needed to be improved. Based on this judgement, a “active follow-up and careful application” strategy for internal software development was introduced. Currently, only 5 per cent of total power input is used for large model programming.
The call indicates that the total annual investment in science and technology is about 13 billion yuan, that the current share of capital purchases in overall technology inputs is not high, and that there is room for additional energy in AI. Management indicators and supporting systems continue to be refined around the goal of “building a smart bank”.