After ChatGPT, GPT-4 and other large models came out, everyone was pleasantly surprised to find that these models were fundamentally different from the previous versions. The new model has basic language ability, some general knowledge, and simple logical reasoning ability (Chain Of Thought). What impact will such technological breakthroughs have on the field of supply chain finance in the future?
1. A breakthrough in the human language system
There is an amazing example in the recent release of GPT-4. The questioner asked AI to tell the story of Cinderella with words starting with 26 letters of AZ. Such language ability is almost impossible for humans to complete in a short time.
Output：
A beautiful Cinderella, dwelling eagerly, finally gains happiness; inspiring jealous kin, love magically nurtures opulent prince; quietly rescues, slipper triumphs, uniting very wondrously, xenial youth zealously.
Although the technical details and model parameters behind GPT have not been announced yet, it can be guessed that after a large amount of data, an increase in the magnitude of model parameters, pre-training, and reinforcement learning in human feedback, the GPT model has changed from the original Transformer algorithm to simple The "cloze" function has achieved a breakthrough in the human language system after evolution. It can be understood that the GPT large model has found the hidden variables behind the language system, so that it can truly understand the relationship between languages and words, while the previous GPT model is more like a parrot in terms of language processing.
Compared with stock and bond financial products, supply chain finance is a business scenario that requires more communication, and each scenario may involve different processes and different materials. GPT-4 can achieve more accurate and faster information and document processing, including the communication content of each participant in supply chain finance, processing various documents and contracts, etc., which helps to improve the efficiency of each participant in supply chain finance. Collaborative efficiency.
2. GPT model: Let machines understand unstructured data for the first time
Before the GPT large model came out, machines could not really understand unstructured data, such as pictures, text, videos, etc. These different types of data need to use various professional tools or models to convert them into structured data and then process them.
Bill Gates, co-founder of Microsoft Corporation, said OpenAI's GPT artificial intelligence model is the most revolutionary technological advance since he first saw a modern graphical desktop environment (GUI) in 1980. The emergence of graphical interfaces enables machines to generate pictures, interfaces and other interactive content that are more in line with human habits according to human needs. However, no matter which software needs to change and customize the graphical interface, it requires high costs. After the GPT model comes out, AI can not only understand unstructured data, but also automatically label pictures and automatically generate pictures through text descriptions. As a result, the cost of interaction between humans and machines is further reduced.
In the scenario of supply chain finance, most of the existing data has been highly abstracted, such as ERP data, customer credit, tax data, contracts and bills, etc. The GPT model can directly understand and break through various original unstructured data such as videos, such as the matching of site construction monitoring videos with business flows and capital flows, warehouse monitoring conditions, and product lifecycle tracking. This enables supply chain finance to better handle the original information when the business occurs, and to better match financial resources with the business. For example, the operating cost of the leasing financial business may be greatly reduced.
3. Logical reasoning ability (Chain Of Thought)
The logical reasoning ability (Chain Of Thought) of the GPT model may come from the reading and training of a large number of logical codes and logical reasoning content. In a conversation between Huang Renxun and Ilya Sutske, Huang Renxun asked how the GPT model evolved from the Transformer model's "cloze" to inference logic, and Ilya Sutske cited a reasoning novel. example to explain. In mystery novels, there are various characters and plots, and readers need to reason to solve puzzles and finally get the answer. In the GPT model, this is similar to the "cloze" function. After a lot of training, the logical reasoning ability of the GPT model can be improved.
The logical reasoning ability of the GPT model makes it have a certain degree of intelligence. Under the guidance of humans, it can decompose complex problems into multiple simple ones, and then gradually get the final answer. For example, the problem of crows drinking water can be broken down into feasible simple solutions through observation, perception, cognition, learning, reasoning and execution. The current intelligence level of GPT has begun to "emerge" similar to crow intelligence.
Using the logical reasoning ability of the GPT model, we can disassemble the complex needs of customers into individual links, which are satisfied by different products, and finally form a complete solution. In addition, logical reasoning ability can also help financial institutions customize suitable financial products for specific customer groups under given constraints, and create more supply chain financial products suitable for various industries. In terms of risk control, false trade or mismatch between financing needs and business often occur in the supply chain. The logical reasoning ability of the GPT model can help organizations reduce the cost of dealing with these problems.
4. The longer the link in the business scenario, the greater the room for AI improvement
Supply chain finance is a typical long-link scenario, especially in the case of multi-level circulation, the link may have as many as a dozen layers; in the international supply chain system, the full life cycle of a product may involve multiple countries and Integration of resources from multiple suppliers. Supply chain finance involves core enterprises, multiple departments within core enterprises, suppliers of core enterprises, multi-level suppliers, multiple capital parties and regulatory authorities. Due to the limitation of professional division of labor, each department and position can only achieve efficient communication in the field they are familiar with, and the efficiency of cross-field and cross-team communication may be reduced, such as 95% communication efficiency. In cross-enterprise communication, due to the lack of trust relationship, the efficiency is further reduced, for example, further down to 75%. Efficiency loss in each link may make the overall efficiency lower than 10%. The longer the link, the higher the probability of service being stuck.
The GPT model can assist each participant in translating their own content into a familiar and accustomed style of the other party, thereby maintaining efficient communication. Even through multiple links, there is not much loss in the overall information transmission, thereby improving the efficiency of each participant in the long-link business scenario.
In addition to the business scenarios of supply chain finance, software development is also a typical long-link scenario. One of the biggest pain points in the software industry is that the speed of software development is not as fast as that of customer needs. From the customer's distributor to initiate demand, to the customer's business department, through the translation of the customer's IT department, and finally to the product manufacturer's business, product, project manager and IT development, every link may lose efficiency, resulting in the failure of the final delivery satisfy customer's request. It may only take a few seconds for a customer to discover a new question or have a new idea, but it may take days or even months to finally solve even a short-answer question. The GPT model can help customers quickly realize the prototype of ideas, and transmit accurate demand information to the development team without loss, thereby greatly improving the efficiency of demand delivery.
5. Quantitative changes in efficiency improvement lead to qualitative changes, promoting more supply chain financial transactions
The GPT model itself is a practical process in which quantitative changes lead to qualitative changes. When GPT-1 came out, it had only 117 million parameters, GPT-2 had 1.5 billion parameters, and GPT-3 had 175 billion parameters. Larger parameters, more data training sets, more manual proofreading, and more powerful AI hardware capabilities make the GPT model finally break through the ceiling of traditional AI models.
For individual consumers, it may be the case that they often go shopping without buying things. But for enterprises, the ultimate goal of information acquisition and analysis is still for transactions, whether it is to provide services to customers, or to allocate internal resources within the enterprise, or to purchase resources needed for production. In the era before the GPT large model, the cost of information acquisition and processing was very high, making many transactions impossible to reach, and many things have been in a state of unresolved discussions. When enterprises are looking for partners and suppliers, they often spend too much time in the preparation of decision-making, and the proportion of final transactions is not necessarily ideal.
A more capable AI model can reduce every link in supply chain finance, lower transaction costs will bring more transaction scale, and resources can be better integrated. The large model's understanding of unstructured data can make supply chain finance expand to richer scenarios, and finally realize the integration of information flow, capital flow and logistics, and the contribution of each resource can be calculated. There is huge room for AI improvement in long-link business scenarios. It can make the entire supply chain financial system more collaborative and efficient by improving communication efficiency across domains, teams, and companies, thereby achieving an improvement from quantitative change to qualitative change. .
The advent of strong artificial intelligence will change the transaction of enterprises, which will be similar to the upgrading of offline shopping scenarios by Internet e-commerce in the past. The industrial Internet, which has not been realized in the past, may usher in new opportunities.
For more reading, please pay attention to the WeChat public account - informatization and digitalization